{"id": "rzQGHXNReU", "venue": "COLM 2024", "paper_decision": "Accept", "title": "RAFT: Adapting Language Model to Domain Specific RAG", "authors": ["Tianjun Zhang", "Shishir G Patil", "Naman Jain", "Sheng Shen", "Matei Zaharia", "Ion Stoica", "Joseph E. Gonzalez"], "abstract": "Pretraining Large Language Models (LLMs) on large corpora of textual data is now a standard paradigm. \nWhen using these LLMs for many downstream applications, it is common to additionally incorporate new information into the pretrained model either through RAG-based-prompting, or finetuning. \nHowever, the best methodology to incorporate information remains an open question. \nIn this paper, we present Retrieval Augmented Fine Tuning (RAFT), a training recipe which improves the model's ability to answer questions in \"open-book\" in-domain settings. In training RAFT, given a question, and a set of retrieved documents, we train the model to ignore those documents that don't help in answering the question, which we call, distractor documents. RAFT accomplishes this by citing verbatim the right sequence from the relevant document to help answer the question. This coupled with RAFT's chain-of-thought-style response helps improve the model's ability to reason. In domain specific RAG, RAFT consistently improves the model's performance across PubMed, HotpotQA, and Gorilla datasets, presenting a post-training recipe to improve pre-trained LLMs to in-domain RAG.", "keywords": "R;e;t;r;i;e;v;e;r;;; ;F;i;n;e;t;u;n;i;n;g", "qa_pairs": [{"question": "Do the authors include general domain datasets like NQ and TQA in their experiments, and if so, why are these datasets used?", "answer": "Yes, authors include the NQ [1] and TQA [2] datasets in our experiments because they are standard benchmarks for RAG. Authors use them to study the effect of the percentage of golden documents in the training dataset, employing the standard retriever from the Replug[3] paper and reporting RAFT's performance.\n[1] Shi, Weijia, et al. \"Replug: Retrieval-augmented black-box language models.\" arXiv preprint arXiv:2301.12652 (2023). \n[2] Kwiatkowski, Tom, et al. \"Natural questions: a benchmark for question answering research.\" Transactions of the Association for Computational Linguistics 7 (2019): 453-466 \n[3] Joshi, Mandar, et al. \"Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension.\" arXiv preprint arXiv:1705.03551 (2017)\n", "category": "Experimental_Analysis"}, {"question": "What are the specific hyperparameters used in the main experiment (Table 2) that are necessary for reproducing the results?", "answer": "Authors use a standard supervised fine-tuning method from the FastChat repository [4], the main experiment uses a learning rate of 1e-5, a warmup ratio of 0.03%, a batch size of 64, a total epoch of 2, and a percentage of golden documents of 80%, following a standard supervised fine-tuning method from the FastChat repository.\n[4] Zheng, Lianmin, et al. \"Judging llm-as-a-judge with mt-bench and chatbot arena.\" Advances in Neural Information Processing Systems 36 (2024)", "category": "Experimental_Setup"}, {"question": "How does the proposed RAFT method contribute novel insights or advancements in the domain-specific 'open-book' exam setting, given that domain-specific fine-tuning and RAG are already known to be effective for domain-specific tasks?", "answer": "The RAFT method contributes novel insights by thoroughly studying the effect of combining domain-specific fine-tuning and RAG, demonstrating that a smaller model (Llama2-7B) can achieve comparable performance to GPT-3.5. It identifies techniques like verbatim chain-of-thought, distractor documents, and percentage of golden documents, improving understanding and reasoning on the given domain and answering styles. Authors' experiments study why RAFT is improving over baselines. There are two potential reasons: (1) improved understanding and reasoning on the given domain and (2) improved answering styles that are easier for the exact matching methods to check. This is shown through comparisons with baselines like Llama2-7B-instruct + RAG, GPT3.5 + RAG, and DSF, Note that this baseline even uses a much stronger base model GPT3.5, compared to RAFT, which uses Llama2. Second, the main reason authors compare RAFT to DSF is to understand the effect of answering style improvement. Using DSF, the model is finetuned to provide the same answering style that is the same as the RAFT model, with the only difference to memorize the knowledge. Thus, this illustrates RAFT does improve the model’s capability of understanding and reasoning over the given context.", "category": "Method_Disambiguation"}, {"question": "Why was RAFT not tested on tasks such as text-to-SQL or reasoning-based QA, given its potential effectiveness in those domains?", "answer": "RAFT is specifically designed to enhance the base LLM's performance in domain-specific RAG settings. While it might improve performance in coding or reasoning domains, studying RAFT's performance on those tasks falls outside the scope of this paper, as it focuses on creating domain expert RAG models rather than general post-training strategies.", "category": "Motivation_Analysis"}, {"question": "What are the main results and ablations for NQ and Trivia QA when applying Chain-of-Thought (CoT), and how do they compare to non-CoT settings?", "answer": "Authors will provide one example here, in one of NQ questions, 'What is the last movie of Jason Bourne' the right answer as per the dataset is 'Jason Bourne'. With CoT, the LLM can answer 'Jason Bourne(2016)' by referencing the documentation within its context. The additional ablation studies on CoT for NQ and TQA showed slight degradation in performance compared to non-CoT settings (but roughly the same number): NQ (CoT: 49.10%, Non-CoT: 50.90%) and TQA (CoT: 62.77%, Non-CoT: 64.77%). Despite this, CoT is still generally helpful in settings where reasoning is heavily involved. The results will be added to the paper along with a discussion on this finding.", "category": "Experimental_Exposition"}, {"question": "Can the fine-tuned RAFT model effectively apply its acquired domain-specific knowledge from seen datasets to unseen datasets within the same domain, as demonstrated in the PubmedQA example?", "answer": "Authors' experiments include two settings: (a) In TorchHub, TensorHub, and HuggingFace, test questions are based on the same documents as the training set, This would mean that authors train question-document-answer triplets on document set A, and then use question-document from the same document set A for testing.and (b) in Hotpot QA, NQ, TQA, and MedPub, test questions don't necessarily refer to the same documents in the training set.  This would mean that authors train question-document-answer triplets on document set A, and then use question-document from document set B for testing. Results in Table 1 and Figure 5 show that RAFT can generalize to similar domain questions with unseen documents. Additional evaluations will be added to further understand RAFT's performance.", "category": "Experimental_Setup"}, {"question": "Was RAFT tested on tasks such as text-to-SQL or reasoning-based QA to validate its effectiveness outside open-domain QA?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "TrloAXEJ2B", "venue": "COLM 2024", "paper_decision": "Accept", "title": "LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition", "authors": ["Chengsong Huang", "Qian Liu", "Bill Yuchen Lin", "Tianyu Pang", "Chao Du", "Min Lin"], "abstract": "Low-rank adaptation (LoRA) is often employed to fine-tune large language models (LLMs) for new tasks. This paper investigates LoRA composability for cross-task generalization and introduces LoraHub, a simple framework devised for the purposive assembly of LoRA modules trained on diverse given tasks, with the objective of achieving adaptable performance on unseen tasks. With just a few examples from a new task, LoraHub can fluidly combine multiple LoRA modules, eliminating the need for human expertise and assumptions. Notably, the composition requires neither additional model parameters nor gradients. Empirical results on the Big-Bench Hard benchmark suggest that LoraHub, while not surpassing the performance of in-context learning, offers a notable performance-efficiency trade-off in few-shot scenarios by employing a significantly reduced number of tokens per example during inference. Notably, LoraHub establishes a better upper bound compared to in-context learning when paired with different demonstration examples, demonstrating its potential for future development. Our vision is to establish a platform for LoRA modules, empowering users to share their trained LoRA modules. This collaborative approach facilitates the seamless application of LoRA modules to novel tasks, contributing to an adaptive ecosystem.", "keywords": "P;a;r;a;m;e;t;e;r; ;E;f;f;i;c;i;e;n;t; ;T;u;n;i;n;g;,; ;C;r;o;s;s;-;T;a;s;k; ;G;e;n;e;r;a;l;i;z;a;t;i;o;n;,; ;M;o;d;e;l; ;M;e;r;g;i;n;g", "qa_pairs": [{"question": "Why was the FLAN-T5 model chosen for the experiments despite its lower performance on the BBH metric, and how does this choice support the validity of the ablation studies?", "answer": "The FLAN-T5 model was chosen because it is highly popular on HuggingFace and known for its proficiency in zero-shot and few-shot learning scenarios, as well as its robust problem-solving capabilities. Its use ensures a fair comparison in zero-shot scenarios and demonstrates the effectiveness of the methods in terms of learning efficiency and adaptability. Additionally, the integration of LoRA modules trained solely on the FLAN collection maintains consistent experimental conditions, preventing bias from additional training data. Results from FLAN-T5-XL, a 3B model in the Appendix further validate the generalizability of LoraHub across model scales.", "category": "Motivation_Analysis"}, {"question": "Why does LoraHub underperform in some scenarios compared to in-context learning and gradient-based methods, and what are its practical advantages?", "answer": "LoraHub's underperformance in some scenarios is acknowledged, but it offers a gray-box approach that requires only model output logits, avoiding the need for access to model weights and extensive GPU memory. This allows efficient operation in inference-only mode, ideal for CPU-only environments, simplifying deployment and reducing computational demands. Additionally, LoraHub achieves a peak performance of 41.2 on the BBH benchmark, exceeding ICL's 38.4 and all other baselines except full fine-tuning, highlighting its potential for superior performance.", "category": "Method_Disambiguation"}, {"question": "How does LoraHub differentiate its novelty from existing works such as MHR, SPoT, and Polytropon?", "answer": "LoraHub employs a unique gray-box method that uses only model output logits, unlike the gradient methods applicable mainly to open-source models in previous work. Additionally, LoraHub functions solely during inference without requiring fine-tuning, facilitating efficient learning on CPU-only machines.", "category": "Method_Comparison"}, {"question": "How does LoraHub address the issue of arbitrary LoRA module selection and its potential negative impact on downstream task performance?", "answer": "LoraHub has introduced $⁣\\rm LoraHub_{filter}$ in Appendix E, which enhances the selection process by identifying the top 20 LoRA module candidates with the lowest loss on few-shot examples, improving selection accuracy and average performance from 34.7 to 35.4.", "category": "Method_Mechanics"}, {"question": "How does the computational resource consumption of LoraHub compare to full fine-tuning and LoRA fine-tuning, particularly in terms of memory usage?", "answer": "Full fine-tuning required about 40GB of memory, LoRA fine-tuning used around 34GB, and LoraHub utilized about 5GB of memory. LoraHub's efficiency is due to its inference-only mode, which eliminates the need for storing gradients and optimization states.", "category": "Method_Comparison"}, {"question": "What factors contribute to the performance gap between LoraHub and ICL, and how does LoraHub demonstrate its potential despite this gap?", "answer": "The performance gap between LoraHub and ICL is primarily due to the current implementation strategy rather than intrinsic limitations of LoraHub. Refining module composition algorithms can significantly narrow this gap. In Appendix B, LoraHub achieves a score of 41.2 on the BBH, surpassing ICL's 38.4, demonstrating its potential. Additionally, LoraHub offers significant advantages in inference latency and efficiency, reducing the number of input tokens needed, which is crucial for latency-sensitive applications. Ongoing enhancements aim to optimize LoraHub's performance across various tasks.", "category": "Experimental_Analysis"}, {"question": "Why did the authors use a randomly selected subset of 20 LoRA modules instead of all available modules for their experiments, and how does increasing the number of LoRA modules impact the computational time required to find optimal configurations?", "answer": "The authors used a subset of 20 LoRA modules primarily for computational efficiency. Increasing the number of LoRA modules increases the dimensionality of the optimization problem, significantly extending the time required to find an optimal configuration. Experiments in Appendix H confirmed that while scaling up the number of modules is feasible, it necessitates more optimization time, highlighting computational trade-offs. Using a fixed number of modules also ensures stability and robustness, allowing systematic evaluation of the LoraHub method without excessive computational overhead.", "category": "Motivation_Analysis"}, {"question": "What is the time efficiency of authoring high-quality few-shot samples using LoraHub, and how does it compare to LoRA modules in terms of processing time and memory usage?", "answer": "In the experiments, authoring five high-quality few-shot examples using LoraHub takes approximately 10 seconds on A100 cards. Additionally, LoraHub uses 5GB of memory with the FLAN-T5-Large model, whereas LoRA fine-tuning requires 34GB, highlighting LoraHub's efficiency in both processing time and memory usage.", "category": "Method_Comparison"}]}
{"id": "FdVXgSJhvz", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "AlpaGasus: Training a Better Alpaca with Fewer Data", "authors": ["Lichang Chen", "Shiyang Li", "Jun Yan", "Hai Wang", "Kalpa Gunaratna", "Vikas Yadav", "Zheng Tang", "Vijay Srinivasan", "Tianyi Zhou", "Heng Huang", "Hongxia Jin"], "abstract": "Large language models~(LLMs) strengthen instruction-following capability through instruction-finetuning (IFT) on supervised instruction/response data. However, widely used IFT datasets (e.g., Alpaca's 52k data) surprisingly contain many low-quality instances with incorrect or irrelevant responses, which are misleading and detrimental to IFT.  In this paper, we propose a simple and effective data selection strategy that automatically identifies and removes low-quality data using a strong LLM (e.g., ChatGPT). To this end, we introduce Alpagasus, which is finetuned on only 9k high-quality data filtered from the 52k Alpaca data. Alpagasus significantly outperforms the original Alpaca as evaluated by GPT-4 on multiple test sets and the controlled human study. Its 13B variant matches $>90\\%$ performance of its teacher LLM (i.e., Text-Davinci-003) on test tasks. It also provides 5.7x faster training, reducing the training time for a 7B variant from 80 minutes (for Alpaca) to 14 minutes \\footnote{We apply IFT for the same number of epochs as Alpaca(7B) but on fewer data, using 4$\\times$NVIDIA A100 (80GB) GPUs and following the original Alpaca setting and hyperparameters.}.  In the experiment, we also demonstrate that our method can work not only for machine-generated datasets but also for human-written datasets. Overall, Alpagasus demonstrates a novel data-centric IFT paradigm that can be generally applied to instruction-tuning data, leading to faster training and better instruction-following models.", "keywords": "data selection; large language model; instruction tuning;", "qa_pairs": [{"question": "What is the frequency of mistakes made by GPT-4 as the filtering model, and how does the performance of Alpagasus vary when using different filtering models such as GPT-3.5, Claude-2, and Llama?", "answer": "The paper does not use GPT-4 as the filtering model; it uses ChatGPT (GPT-3.5-turbo) and Claude2. Comprehensive evaluations across four test sets, including GPT4-based and human studies, demonstrate the effectiveness of the filtering method. The follow-up work, Humpback, shows that the 70B-LLaMA model can self-filter with proper prompts and thresholds, indicating the method is not highly dependent on specific models. Manual examination of 400 examples ( i.e., 200 examples from the high-quality 9k data and another 200 examples from the low-quality 43k) revealed no mistakes made by the LLM filter. The filtering method is reliable across different models, including ChatGPT, Claude, and LLaMA.", "category": "Experimental_Exposition"}, {"question": "What are the costs associated with using ChatGPT for data filtering, and how do they compare to the costs of human annotation?", "answer": "The cost of using ChatGPT for filtering 52,000 data samples is approximately 50 dollars and takes around 1.5 hours. In contrast, human annotation would cost around 10,825 dollars and take at least 433 hours, assuming one person takes 30 seconds to rate each data sample, the total time will be: `T=(52,002 x 30s) / (60 x 60)=433.35 hrs`. Additionally, the investment in data filtering is a one-off expense, and the filtered data supports multiple model iterations, such as LLaMA-1 and LLaMA-2. The methodology also reduces training costs for larger models, such as reducing the cost for 13B models from 225.28 dollars to 40.96 dollars, with potential for even greater savings for ultra-large models like BLOOM-176B.", "category": "Method_Comparison"}, {"question": "Why does AlpaGasus perform partially weaker than Alpaca on the MMLU dataset, and why is the improvement of AlpaGasus over Alpaca considered marginal?", "answer": "AlpaGasus performs partially weaker than Alpaca on the MMLU dataset because MMLU is a knowledge-intensive benchmark not specifically designed to evaluate instruction-following capability, which is the focus of AlpaGasus's instruction-finetuning (IFT). The goal of IFT is to refine the model's ability to comprehend and respond to complex user queries, not to enhance knowledge acquisition. AlpaGasus is designed to maintain performance on MMLU rather than significantly improve it. The improvement of AlpaGasus over Alpaca is considered marginal because instruction-following is a more important evaluation metric for instruction tuning, and AlpaGasus consistently demonstrates gains in instruction-following scores. Additionally, AlpaGasus outperforms models trained with randomly selected samples across all model scales, highlighting the efficacy of the data filtering approach. For example, on Dolly, the 13B(3k) model is better than 13B(15k) on all four benchmark datasets.", "category": "Method_Disambiguation"}, {"question": "How do you address the potential bias introduced by using GPT-4 for both data quality evaluation and final model output evaluation in your study?", "answer": "Authors used GPT-3.5-turbo for data filtering and GPT-4 for automatic evaluation, which are distinct models with different functionalities. Additionally, authors conducted human-based evaluations to validate  findings, and applied Claude2 as an alternative LLM filter to mitigate bias concerns. It shows AlpaGasus' answers are preferred when compared with the answers generated by Alpaca. ", "category": "Method_Mechanics"}, {"question": "How many runs are performed for the random splits, and what are the standard deviations of the results?", "answer": "Authors generate another random split with a different random seed and evaluate the model trained with it. The benchmark results are shown in the provided table, which includes the performance of Alpaca(9k-random)-v1, Alpaca(9k-random)-v2, and AlpaGasus. The instruction-following evaluations are also provided, with the winning score calculated as `1 + (#Win−#Lose) / (#Testset)`, as shown in the caption of Figure 1.\n\n| Dataset | BBH | DROP | Humaneval | MMLU |\n | --- | --- | --- | --- | --- | \n| Alpaca(9k-random)-v1 | 31.89 | 25.88 | 11.59 | 36.93 | \n| Alpaca(9k-random)-v2 | 31.71 | 25.63 | 11.72 | 36.71 | \n| AlpaGasus | 33.76 | 26.03 | 12.20 | 38.78 | \n\n|Win score| Vicuna | Koala | WizardLM | Self-Instruct | \n|--- | --- | --- | --- | --- |\n |Ours-9k vs. 9k-random-v1 | 1.213 | 1.022 | 1.060 | 1.056 | \n|Ours-9k vs. 9k-random-v2 | 1.188 | 1.033 | 1.069 | 1.071 | \n\n", "category": "Experimental_Exposition"}, {"question": "What is the agreement between the scores obtained using Claude2 as the LLM filter and those using ChatGPT, and how does this address potential bias concerns?", "answer": "The results demonstrate that the effectiveness of the method is consistent across different LLM filters, including Claude2 and ChatGPT, as evaluated by GPT-4 as the judge. This consistency mitigates potential bias concerns.", "category": "Method_Mechanics"}, {"question": "What evidence supports the reliability of the LLM filter in selecting high-quality data for instruction tuning?", "answer": "The reliability of the LLM filter is supported by its effectiveness across various rating dimensions (e.g., accuracy and helpfulness), LLM filters (e.g., ChatGPT and Claude-2), base model families (e.g., LLaMA-1 and LLaMA-2), model sizes (e.g., 7B and 13B), dataset types (e.g., machine-generated and human-written), and evaluation methods (e.g., GPT-4-based evaluations and human studies). This demonstrates its capability in selecting high-quality data for instruction tuning.", "category": "Method_Disambiguation"}, {"question": "Does the filtering method used in the study scale effectively to larger models, such as 30B, despite the computational constraints mentioned in the limitations?", "answer": "Yes, the filtering method scales effectively to larger models. Experiments were conducted on a 30B model using LoRA, showing that AlpaGasus-30B with 9k filtered data outperforms Alpaca-30B with 52k data. The models were trained on 8xRTX A6000 for ~18 hours for Alpaca and ~3.5 hours for AlpaGasus, with a batch size of 128, a learning rate of 1e-5, a max length of 1024, and 5 epochs. Alpagasus-30B(9k data) vs. Alpaca-30B(52k data): | Testset | Win | Tie | Lose | |--- | --- | --- | --- | | Vicuna | 20 | 50 | 10 | | Koala | 38 | 113 | 29 | | WizardLM | 41 | 141 | 36 | | Self-Instruct | 43 | 172 | 37 | Alpagasus-30B(9k data) vs. Alpaca-30B(9k random data): | Testset | Win | Tie | Lose | |--- | --- | --- | --- | | Vicuna | 18 | 53 | 9 | | Koala | 35 | 114 | 31 | | WizardLM | 47 | 148 | 23 | | Self-Instruct | 50 | 165 | 37 | From the table, authors can see that when the model size goes up to 30B, the paper's Alpagasus can still perform better than the Alpaca-52k with only 9k filtered data.", "category": "Experimental_Exposition"}, {"question": "What additional experiments were conducted to evaluate the adversarial robustness of the data filtering method, and how do the results demonstrate its effectiveness?", "answer": "Additional experiments were conducted on the llm-rule benchmark, which evaluates how well LLMs follow developer-provided rules under adversarial inputs. The adversarial robustness of AlpaGasus-7B(9k), Alpaca-7B(52k), and Alpaca-7B(9k random) was tested. \n| | Avg. | Indirect | Just-Ask | Legalese | Obfs | Rule-C | Sim |\n |---------------|---------------|---|---------------|---------------|---------------|---------------|---------------| \n| Alpagasus | **38.12** | 37.97 | 36.36 | 35.23 | 38.96 | 38.64 | 41.67 | \n| Alpaca | 38.02 | 38.23 | 40.68 | 32.95 | 38.96 | 38.63 | 41.67 |\n | Alpaca(random) | 35.58 | 34.22 | 36.36 | 36.36 | 38.96 | 35.61 | 35.71 | \n\nThe results, shown in a table, indicate that AlpaGasus performs the best among the three models across six adversarial strategies (`Indirect`, `Just-ask`, `legalese`, `Obfuscation`, `Rule Change`, and `Simulation`), demonstrating its effectiveness on adversarial robustness.", "category": "Experimental_Exposition"}, {"question": "How does the granularity of the score (0.5 for ChatGPT and 1 for Claude2) affect the performance of the models, and is this difference inherent to the models' outputs?", "answer": "The granularity of ChatGPT is 0.5, and the granularity of Claude2 is 1. This difference in granularity is inherent to the models' outputs in response to rating prompts. The score distributions for both models are shown in Figure 5 and Figure 14, respectively. The paper's analysis indicates that both granularity levels effectively contribute to the filtration and refinement of data.", "category": "Method_Mechanics"}, {"question": "What is the effect of system prompts on the final performance of Claude2 compared to ChatGPT?", "answer": "Claude2 operates without utilizing system prompts, unlike ChatGPT. Despite this difference, Claude2 demonstrates considerable proficiency in executing filtering tasks, as shown in Figure 14, suggesting that the absence of system prompts does not significantly impede its performance efficiency.", "category": "Method_Comparison"}, {"question": "Does reducing the quantity of training data negatively impact the adversarial robustness of the tuned models, specifically AlpaGasus-7B(9k), Alpaca-7B(52k), and Alpaca-7B(9k random)?", "answer": "No, the robustness is not affected. Based on the adversarial robustness benchmark results, AlpaGasus-7B(9k) outperforms Alpaca-7B(52k) and Alpaca-7B(9k random) across six adversarial strategies (`Indirect`, `Just-ask`, `legalese`, `Obfuscation`, `Rule Change`, and `Simulation`). Additionally, the filtering method used in AlpaGasus enhances safety by filtering biased responses and backdoor examples in the training data.\n | | Avg. | Indirect | Just-Ask | Legalese | Obfs | Rule-C | Sim |\n |---------------|---------------|---|---------------|---------------|---------------|---------------|---------------| \n| Alpagasus | **38.12** | 37.97 | 36.36 | 35.23 | 38.96 | 38.64 | 41.67 |\n | Alpaca | 38.02 | 38.23 | 40.68 | 32.95 | 38.96 | 38.63 | 41.67 | \n| Alpaca(random) | 35.58 | 34.22 | 36.36 | 36.36 | 38.96 | 35.61 | 35.71 | ", "category": "Experimental_Exposition"}, {"question": "Does the follow-up work Humpback demonstrate that the LLaMA-2-70B model can self-filter its responses using a method similar to the one employed in the study?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was Claude2 used as an alternative LLM filter to address potential bias concerns related to GPT-3.5-turbo?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the results in Figure 14 evaluated using GPT-4?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the cost of using ChatGPT for filtering 52,000 data samples less than the cost of human annotation for the same amount of data?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the reduced quantity of data negatively impact the adversarial robustness of the tuned models according to the experimental results?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the experimental results on the 30B model show that Alpagasus outperforms Alpaca-52k with only 9k filtered data across various testsets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the standard deviations of the results provided in the paper for the random splits?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were the experiments on the 30B model conducted with the same computational resources and training parameters as specified in the rebuttal?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the authors account for the costs associated with using ChatGPT for data filtering in their comparison of finetuning costs?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the filtering method's effectiveness demonstrated on models larger than 13B, specifically on a 30B model?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was more than one random split generated and evaluated to assess the model's performance consistency?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the study use GPT-4 for both data filtering and final model output evaluation, potentially leading to bias towards GPT-4's preferences?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "IEduRUO55F", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Eureka: Human-Level Reward Design via Coding Large Language Models", "authors": ["Yecheng Jason Ma", "William Liang", "Guanzhi Wang", "De-An Huang", "Osbert Bastani", "Dinesh Jayaraman", "Yuke Zhu", "Linxi Fan", "Anima Anandkumar"], "abstract": "Large Language Models (LLMs) have excelled as high-level semantic planners for sequential decision-making tasks. However, harnessing them to learn complex low-level manipulation tasks, such as dexterous pen spinning, remains an open problem. We bridge this fundamental gap and present Eureka, a human-level reward design algorithm powered by LLMs. Eureka exploits the remarkable zero-shot generation, code-writing, and in-context improvement capabilities of state-of-the-art LLMs, such as GPT-4, to perform evolutionary optimization over reward code. The resulting rewards can then be used to acquire complex skills via reinforcement learning. Without any task-specific prompting or pre-defined reward templates, Eureka generates reward functions that outperform expert human-engineered rewards. In a diverse suite of 29 open-source RL environments that include 10 distinct robot morphologies, Eureka outperforms human experts on 83% of the tasks, leading to an average normalized improvement of 52%. The generality of Eureka also enables a new gradient-free in-context learning approach to reinforcement learning from human feedback (RLHF), readily incorporating human inputs to improve the quality and the safety of the generated rewards without model updating. Finally, using Eureka rewards in a curriculum learning setting, we demonstrate for the first time, a simulated Shadow Hand capable of performing pen spinning tricks, adeptly manipulating a pen in circles at rapid speed.", "keywords": "Large Language Models;Reinforcement Learning;Dexterous Manipulation;Reward Learning;Robotics", "qa_pairs": [{"question": "Does Eureka use the same number of environment samples as the baseline for training, and how does its sample efficiency compare to the baselines?", "answer": "For the final training runs, Eureka uses the exact same number of environment samples as the baseline final reward functions (e.g., L2R and Human) for training. As stated in Section 4.2 of the paper, 'for each task, all final reward functions are optimized using the same RL algorithm with the same set of hyperparameters'; this fixed set of hyperparameters includes the predetermined number of environment samples used for RL training. Given that the hyperparameters are chosen by the original benchmark designers to ensure the Human reward function is effective, the paper's evaluation setting in fact disadvantages Eureka’s reward functions as Eureka does not tune the RL algorithm (PPO) to work well with its reward functions. as shown in the aggregate RL training curves on the 20 Dexterity tasks in Figure 10, Appendix F, Eureka reward functions exhibit better sample efficiency compared to the baselines, but given that the RL algorithm is not tuned to Eureka rewards, the learning curves do exhibit slightly more variance throughout training.", "category": "Experimental_Setup"}, {"question": "How does Eureka handle adaptation to complex environments that may exceed the model's context window length?", "answer": "Eureka has been demonstrated on complex environments like the bi-manual Dexterity benchmark, which have state dimensions larger than 400 and observation source code spanning over 200 lines. To ensure compatibility, an automatic script extracts the observation portion of the environment source code, keeping it within the context length. Eureka primarily requires information about environment state and action spaces, which is of reasonable token length even for large environments. Additionally, advancements in LLM context lengths and future improvements in efficient LLM inference methods are expected to mitigate potential bottlenecks.", "category": "Method_Mechanics"}, {"question": "What are the exact inputs and outputs in the evolutionary search iteration described in the subsection on evolutionary search?", "answer": "In the evolutionary search iteration, the inputs are (a) the best reward function from the previous iteration, (b) its reward reflection (Section 3.3), and (c) the prompt instruction for how to mutate the provided reward function given its reward reflection feedback. These inputs prompt the LLM to generate $K$ i.i.d. outputs, corresponding to $K$ reward samples.", "category": "Method_Disambiguation"}, {"question": "Can the success/fitness function be used to initialize the Eureka reward search process?", "answer": "Yes, the success/fitness function can be used to initialize the Eureka reward search process. The authors have conducted an experiment where a human-supplied reward function was used to initialize Eureka, as described in Section 4.4. Using the fitness function, which is the optimization objective for Eureka reward search, can lead to reward functions that better optimize the objective, such as creating a success bonus in reward generations.", "category": "Experimental_Setup"}, {"question": "Why do the paper claims that EUREKA generates reward functions that outperform expert human-engineered rewards despite the lack of statistically significant differences in performance?", "answer": "First and foremost, the authors disagree with the notion that significance tests should be considered in the first place. [1] in fact argues against the significance test driven way of method comparison, stating that evaluation should embrace ‘statistical thinking but avoid statistical significance tests (e.g., p-value < 0.05) because of their dichotomous nature (significant vs. not significant) and common misinterpretations such as 1) lack of statistically significant results does not demonstrate the absence of effect.’ The authors agree with this statement in its entirety, and have already included an extensive amount of evaluation, including many aspects suggested by the reader and by [1], that together construct a holistic performance profile that suggests the improvement is very likely to exist. Second, even if statistical significance were considered, Eureka in fact produces reward functions that are statistically significantly better than human-designed reward functions. In particular, as suggested by [1], the probability of improvement is considered a robust comparison metric. These results are presented in Figure 12; the probability that the Eureka reward outperforms the human reward on a randomly selected task is above 50% with statistical significance as determined by a Mann-Whitney U-Test (also suggested by [1]) at p-value threshold of 0.05 (p<<0.05); the lower end of the 95% CI is well above 55%. In conclusion, in light of the added metrics that provide detailed performance profiles of Eureka and the baselines, the paper’s original claims are only strengthened.\n\n[1] Agarwal, Rishabh, et al. \"Deep reinforcement learning at the edge of the statistical precipice.\" Advances in neural information processing systems 34 (2021): 29304-29320.", "category": "Method_Disambiguation"}, {"question": "Does Eureka use significantly more environment samples than the baseline during training, and are the RL training curves included in the paper?", "answer": "No, Eureka uses the exact same number of environment samples as the baseline final reward functions (e.g., L2R and Human) for training, as stated in Section 4.2 of the paper. This fixed set of hyperparameters includes the predetermined number of environment samples used for RL training. Given that the hyperparameters are chosen by the original benchmark designers to ensure the Human reward function is effective, the paper's evaluation setting in fact disadvantages Eureka’s reward functions as Eureka does not tune the RL algorithm (PPO) to work well with its reward functions. The RL training curves for the 20 Dexterity tasks are included in Figure 10, Appendix F, showing that Eureka reward functions exhibit better sample efficiency compared to the baselines, albeit with slightly more variance due to the RL algorithm not being tuned for Eureka rewards.", "category": "Experimental_Setup"}, {"question": "How does the generality of Eureka's approach hold when tested on different simulators and RL algorithms, given that it has only been evaluated on Isaac Gym and a fixed RL algorithm?", "answer": "Eureka has been tested on the Mujoco Humanoid environment, where it outperformed the official human-written reward function. This demonstrates its effectiveness across different simulators and code syntax. The authors argue that Eureka's reward generation pipeline abstracts away information about the simulator and RL algorithm, making it likely to work on new simulators or algorithms. Additionally, Eureka's generality is claimed over the diversity and complexity of robots and tasks, supported by its performance on 29 tasks spanning 10 distinct robots with varying observation code. Isaac Gym's status as a widely used benchmark [3,4]further supports the promise of Eureka's results for real-world policy deployment.\n|                 | Eureka   | Human |\n|-----------------|----------|-------|\n| Mujoco Humanoid | **7.68** | 5.92  |\n\n[3] Rudin, Nikita, et al. \"Learning to walk in minutes using massively parallel deep reinforcement learning.\" Conference on Robot Learning. PMLR, 2022.\n\n[4] Handa, Ankur, et al. \"Dextreme: Transfer of agile in-hand manipulation from simulation to reality.\" 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023.\n\n", "category": "Experimental_Analysis"}, {"question": "How was the fairness of the comparison between Eureka and the tuned human rewards ensured, and what were the results of this comparison?", "answer": "The fairness of the comparison was ensured by implementing a baseline for Humanoid in the Isaac suite, where 4 components from the human reward functions were selected and their weights were grid-searched. Specifically, 3 possible values (0.1x, 1x, 10x the original value) were tested for each weight, resulting in 81 tuned human reward functions (80 new ones). The results showed that while Human (Tuned) improved, the relative ordering between Eureka and Human did not change. Eureka's best reward generated from scratch outperformed the best tuned human reward function, even though the baseline was unfairly advantaged as it tuned an existing reward function rather than designing one from scratch.\n|          | Eureka   | Human | Human (Tuned) |\n|----------|----------|-------|---------------|\n| Humanoid | **8.24** | 6.28  | 7.47          |\n", "category": "Experimental_Exposition"}, {"question": "What is the impact on the pen spinning results when considering more detailed material characteristics and physics, such as friction, inertia, and actuator latencies?", "answer": "The pen spinning environment is based on the original Isaac Gym ShadowHand environment, which mirrors the OpenAI Shadow Hand Rubik’s cube environment that has enabled real-world transfer. No changes were made to the physics parameters to simplify the task, ensuring realism. Therefore, the pen spinning environment effectively demonstrates Eureka’s capability in scaling up to very difficult tasks without needing additional modifications to material characteristics or physics.", "category": "Experimental_Analysis"}, {"question": "What is the error bar for the Human (Tuned) reward function in the Humanoid task, and is the improvement of Eureka's reward function over Human (Tuned) statistically significant?", "answer": "The error bar for the Human (Tuned) reward function in the Humanoid task is 7.47 ± 0.12. The improvement of Eureka's reward function over Human (Tuned) is statistically significant, as evidenced by an independent sample t-test with p=0.0002 and a Mann-Whitney U-test with p=0.007.\n|          | Eureka               | Human (Tuned)   |\n|----------|----------------------|-----------------|\n| Humanoid | **8.24** $\\pm$ 0.05  | 7.47 $\\pm$ 0.12 |", "category": "Experimental_Exposition"}, {"question": "What is the relationship between reward reflection and the choices of RL method, and how are these concepts clarified in the manuscript?", "answer": "Reward reflection is a textual evaluation of the reward function, such as summarizing the training dynamics of its components. The discussion about the choices of RL method provides intuitive justification for why reward reflection is helpful. The manuscript has been updated to more clearly separate the procedure of reward reflection from its motivations, with concrete examples provided in Appendix G.1.", "category": "Motivation_Analysis"}, {"question": "What level of detail is required to describe the environment for 'environment as context,' and what specific information is needed for simulators different from Isaac to ensure compatibility?", "answer": "For 'environment as context,' only the state and action variables in the environment need to be exposed in the source code. An automatic script extracts the observation function from the raw environment source code. The prompts and observation source code do not reveal the simulator's identity, ensuring generality. For simulators different from Isaac, the same level of information—specifically, the names of state and action variables—is required. This has been empirically validated with the Mujoco Humanoid environment, where Eureka successfully generated effective reward functions despite differences in how variables are exposed (comment block vs. class attributes).", "category": "Method_Mechanics"}, {"question": "How does the Eureka approach address the limitations of not having full access to the environment code or operating in a real-world setting?", "answer": "The Eureka approach is expected to be useful even without full access to the environment code, as it only requires knowledge of the state and action variables, which can be extracted automatically or accessed via an API. For real-world environments, Eureka can integrate with Sim2Real techniques for policy transfer from simulation, or use state-estimation techniques to construct a symbolic representation of the environment, enabling the definition of reward functions. State-estimation based solutions have been attempted by many other works targeted at learning and deploying in the real world [1,2].\n[1] Smith, Laura, Ilya Kostrikov, and Sergey Levine. \"A walk in the park: Learning to walk in 20 minutes with model-free reinforcement learning.\" arXiv preprint arXiv:2208.07860 (2022).\n\n[2] Handa, Ankur, et al. \"Dextreme: Transfer of agile in-hand manipulation from simulation to reality.\" 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023.", "category": "Method_Mechanics"}, {"question": "Is the classification of the generation as 'zero-shot' accurate, and how is 'zero-shot' defined in the context of this study?", "answer": "In this study, 'zero-shot' refers to generation performed without any examples in the prompt, distinguishing it from 'few-shot prompting' where input-output examples are provided. The classification is based on the lack of examples in the prompt, not the composition of the training dataset. The manuscript has been updated to clarify this definition. While the possibility that GPT-4 has seen similar environments in its training data cannot be ruled out, benchmarks were selected after the cut-off date to ensure identical environments were unseen. Additionally, most bi-manual Dexterity tasks are novel, and Eureka demonstrates the ability to transfer general reward design knowledge across domains, as supported by reward correlation analysis in Section 4.3, which finds Eureka to discover novel reward functions compared to the original human-written ones.", "category": "Concept_Understanding"}, {"question": "What is the justification for assuming access to parts of the environment source code in Eureka, and how does this differ from traditional reinforcement learning setups?", "answer": "The problem setting of Eureka is reward design, not reinforcement learning. Reward design traditionally involves humans writing and refining reward function codes through trial and error, and Eureka automates this process. The outer loop of Eureka is a code generation problem, where it is standard to assume access to parts of the source code to generate or complete other parts[1]. Additionally, Eureka’s 'environment as context' procedure only requires the environment’s state and action variables to be exposed, which can be provided via an API without requiring the full environment code. The inner loop of Eureka uses reinforcement learning to evaluate reward candidates, and this loop accesses the environment as a black-box simulation.\n[1] Roziere, Baptiste, et al. \"Code llama: Open foundation models for code.\" arXiv preprint arXiv:2308.12950 (2023).", "category": "Experimental_Setup"}, {"question": "What is the justification for the assumption that the method is applicable only to tasks with known ground-truth reward functions, and how does Eureka handle tasks where such functions are not analytically expressible?", "answer": "The ground-truth reward functions are sparse task success criteria that are easy to define in practice, as demonstrated by their one-line implementation across 29 benchmark tasks (see Appendix B). Eureka is applicable to tasks with programmatic success conditions, which constitute a large space of tasks. Additionally, for tasks where desired behaviors are difficult to express mathematically (e.g., running in a stable gait), Eureka from Human Feedback allows humans to provide fitness feedback in textual form, bypassing the need for an analytical fitness function. This approach has been shown to align reward functions with human intention effectively. Furthermore, the updated manuscript discusses the potential use of vision-language models (VLMs) to automate textual feedback, offering a scalable alternative to human-based fitness feedback.", "category": "Method_Mechanics"}, {"question": "How was the pre-trained policy obtained in the study?", "answer": "The pre-trained policy was obtained by using the Eureka framework, which iteratively discovers better reward functions in the Shadow Hand environment of the Isaac Gym benchmark suite. The cube object was replaced with a pen object. The best Eureka reward was used to train a policy to solve the task, and the resulting converged policy from this training is referred to as the 'pre-trained' policy.", "category": "Method_Mechanics"}, {"question": "What are the key differences between the 'pre-training' and 'fine-tuning' stages in terms of the distribution and sequence of target pen poses?", "answer": "The key difference lies in the distribution of pen target poses. In pre-training, random poses from the SO(3) 3D orientation group are sampled uniformly. When the current target pose is deemed achieved (i.e., pose difference lower than certain threshold), then a new target pose is provided. In the pre-training stage, a random pose from SO(3) 3D orientation group is sampled uniformly. In fine-tuning, target poses are a predetermined sequence of waypoints specifying pen spinning patterns. The policy switches to the next waypoint upon reaching the current one, completing a cycle when all waypoints are reached consecutively, allowing continuous spinning until the episode length is reached.", "category": "Experimental_Analysis"}, {"question": "Is the success of the pen spinning task primarily due to the dual-stage policy optimization rather than the reward design, and does this cast doubt on the central argument of the paper?", "answer": "The success of the pen spinning task is not solely due to either the dual-stage policy optimization or the reward design alone, but rather their combination. An ablation study was conducted where the dual-stage curriculum was fixed, and the Eureka  reward with a simple reward function that rewards the policy proportional to the negative distance between the pen’s current pose to the target pose. The resulting policy cannot solve the pre-training stage competently, often dropping the pen with unnatural and jerky hand motion. . The resulting policy failed to solve the pre-training stage competently, demonstrating that both components are necessary. Additionally, the central argument of the paper is not that a good reward function alone is sufficient, but rather that automated reward design is effective at scale, as supported by other experiments in the paper that do not use a curriculum.", "category": "Experimental_Analysis"}, {"question": "What are the computational resources and runtime required for a single run of Eureka with default hyperparameters?", "answer": "A single run of Eureka with default hyperparameters can be executed on 4 32GB memory GPUs (e.g., NVIDIA V100) and takes less than one day of wall clock time, even for the most complex environments such as bimanual manipulation.", "category": "Experimental_Exposition"}, {"question": "For each sampled reward, how many environments are created for training, and how is this number determined?", "answer": "The number of environments created for training is determined by the task-specific hyperparameters set in the original benchmarks. For example, the number of environments is 2048 for Humanoid. The default number works well for all rewards, demonstrating the robustness of the Eureka approach.", "category": "Experimental_Setup"}, {"question": "How is the termination of each running experiment determined in the RL runs?", "answer": "Early termination is not performed for RL runs. Each RL run executes for a fixed number of episodes according to the default task-specific hyperparameters provided in the benchmarks. ", "category": "Experimental_Setup"}, {"question": "Is it possible that RL training is not long enough to miss an effective reward sample, and how is this addressed?", "answer": "It is possible that RL training might miss an effective reward sample if the training duration is insufficient. However, in practice, the benchmark default training hyperparameters are sufficient for discovering good reward functions across tasks. With additional compute budget, the training time can be increased to reduce false negatives.", "category": "Method_Mechanics"}, {"question": "What is the total time required to find a good reward function, and what type of device is used?", "answer": "All experiments took less than one day of wall clock time, and each individual experiment can be executed on 4 V100 GPUs.", "category": "Experimental_Exposition"}, {"question": "Is the Eureka method applicable to tasks involving a variety of object types?", "answer": "Yes, Eureka is a general reward design method applicable to diverse object types. It has been demonstrated to design effective reward functions for tasks like pen re-orientation and spinning, as well as re-orienting a cube on different robot hands. Prior works also support the use of similar reward functions across diverse objects[1,2], suggesting Eureka's applicability.\n\n[1] Chen, Tao, Jie Xu, and Pulkit Agrawal. \"A system for general in-hand object re-orientation.\" Conference on Robot Learning. PMLR, 2022.\n\n[2] Qi, Haozhi, et al. \"In-hand object rotation via rapid motor adaptation.\" Conference on Robot Learning. PMLR, 2023.", "category": "Method_Disambiguation"}, {"question": "How does the Eureka method handle tasks that require image inputs, and what are the potential challenges and strategies involved?", "answer": "The Eureka method can infer state variables relevant to reward definition from vision, leveraging state estimation techniques that have been effective in vision-based robotic learning[3,4]. While accuracy may be challenging in complex scenes, this is a common issue for real-world policy learning methods. Additionally, if only the final policy needs visual inputs, Eureka can design the reward function in simulation and then distill the state-based policy to a vision-based policy, a strategy used in Sim2Real approaches[5].\n[3] Gu, Shixiang, et al. \"Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates.\" 2017 IEEE international conference on robotics and automation (ICRA). IEEE, 2017.\n\n[4] Büchler, Dieter, et al. \"Learning to play table tennis from scratch using muscular robots.\" IEEE Transactions on Robotics 38.6 (2022): 3850-3860.\n\n[5] Lee, Joonho, et al. \"Learning quadrupedal locomotion over challenging terrain.\" Science robotics 5.47 (2020)\n\n", "category": "Method_Mechanics"}, {"question": "Is Eureka capable of writing reward functions for generalized grasping, and how does it address the challenges associated with this task?", "answer": "Yes, Eureka is capable of writing reward functions for generalized grasping. Recent literature has shown that RL-based approaches, combined with an object curriculum and manually-engineered reward components (e.g., object reaching, touching, lifting, and moving to target pose), can effectively learn generalized grasping policies. Eureka has successfully generated these reward components in the Dexterity benchmark suite, demonstrating its capability to incorporate such components and more due to its open-ended search nature. Additionally, Eureka has been shown to work well with curriculum-based learning, as demonstrated in the pen spinning task (Section 4.3), suggesting that a curriculum-based approach with Eureka can also address the challenges of generalized dexterous grasping.", "category": "Method_Mechanics"}, {"question": "Does image-based input reduce environment parallelism due to GPU memory constraints and negatively impact training performance?", "answer": "Yes, training with visual inputs reduces the number of parallel simulations from 2048 to 256 on a single RTX 3090 GPU and negatively impacts RL training performance, as shown in Figure 8 of the original Dexterity paper. The main challenges for vision-based training in simulation include slower simulation speed and instability in vision-based reinforcement learning, which are generic issues in simulation-based visual RL. However, recent advancements, such as a ten-fold increase in vision-based simulation speed without sacrificing parallel environments, suggest potential improvements. Additionally, Eureka can support state-of-the-art dexterous manipulation approaches by designing reward functions for state-based teacher policies, which can then be distilled into vision-based student policies.", "category": "Experimental_Exposition"}, {"question": "What is the detailed evaluation protocol used to measure the performance of reward functions in Section 4.2, and how does it ensure fairness and mitigate biases in comparisons?", "answer": "The evaluation protocol involves obtaining a reward function using each approach (authors and the baselines). For each reward function, 5 PPO runs are performed. For each PPO run, the best checkpoint out of 10 checkpoints taken at fixed training intervals is selected, and its performance is evaluated by averaging over 1000+ test rollouts that randomize initial environment states. The performance is then averaged over the best checkpoints from all PPO runs. This procedure is applied uniformly to all reward functions, including baselines, ensuring no bias towards method. The protocol is fair, as the maximum is taken over the same number of checkpoints for each approach, and outliers are mitigated by averaging over multiple runs. The manuscript has been updated to clarify these details, and additional evaluation metrics have been added to provide a comprehensive assessment of Eureka’s efficacy in reward design.", "category": "Method_Mechanics"}, {"question": "Does the paper claim that a good reward function alone is sufficient to solve all difficult tasks?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the RL algorithm (PPO) tuned specifically for Eureka’s reward functions in the evaluation setting?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the EUREKA generate reward functions statistically significantly better than human-designed rewards, based on a Mann-Whitney U-Test with a p-value threshold of 0.05?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the ablation study show that the success of the pen spinning task is due to the dual-stage policy optimization alone?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do all experiments in the paper use a curriculum-based approach?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was Eureka evaluated on multiple base simulators and with different RL algorithms to support its claimed generality?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Eureka use a different number of environment samples compared to the baseline for final training runs?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Eureka's approach to generalized grasping rely solely on end-to-end RL without additional strategies like an object curriculum?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "Rc7dAwVL3v", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers", "authors": ["Kai Shen", "Zeqian Ju", "Xu Tan", "Eric Liu", "Yichong Leng", "Lei He", "Tao Qin", "sheng zhao", "Jiang Bian"], "abstract": "Scaling text-to-speech (TTS) to large-scale, multi-speaker, and in-the-wild datasets is important to capture the diversity in human speech such as speaker identities, prosodies, and styles (e.g., singing). Current large TTS systems usually quantize speech into discrete tokens and use language models to generate these tokens one by one, which suffer from unstable prosody, word skipping/repeating issue, and poor voice quality. In this paper, we develop NaturalSpeech 2, a TTS system that leverages a neural audio codec with residual vector quantizers to get the quantized latent vectors and uses a diffusion model to generate these latent vectors conditioned on text input. To enhance the zero-shot capability that is important to achieve diverse speech synthesis, we design a speech prompting mechanism to facilitate in-context learning in the diffusion model and the duration/pitch predictor. We scale NaturalSpeech 2 to large-scale datasets with 44K hours of speech and singing data and evaluate its voice quality on unseen speakers. NaturalSpeech 2 outperforms previous TTS systems by a large margin in terms of prosody/timbre similarity, robustness, and voice quality in a zero-shot setting, and performs novel zero-shot singing synthesis with only a speech prompt. Audio samples are available at https://naturalspeech2.github.io/.", "keywords": "text-to-speech;large-scale corpus;non-autoregressive;diffusion", "qa_pairs": [{"question": "What experimental results and analyses demonstrate the effectiveness of the modules in applications beyond Text-to-Speech (TTS), specifically in singing voice synthesis?", "answer": "The paper has conducted ablation experiments in singing voice synthesis: 1) Removing CE loss significantly worsens singing generation quality, similar to TTS results, demonstrating CE loss effectiveness. 2) Without explicit pitch modeling, the model loses the ability to generate singing audio following specific music scores, resulting in poor quality. Pitch information is crucial for guiding the singing voice synthesis process.", "category": "Experimental_Exposition"}, {"question": "What are the roles of SoundStream and WaveNet in the proposed model, and how do they interact with other components such as the prior model, diffusion model, and audio codec?", "answer": "SoundStream is used as the neural audio codec architecture because it provides good reconstruction quality, low bitrate, and converts waveforms into continuous vectors while retaining fine-grained details. Based on the quantized token IDs, the paper can design the regularization loss, which further enhances the prediction precision. The architecture is adapted by using the post-quantized latent $z$ as the representation for waveform $x$ and as the training target for the diffusion model, along with a regularization loss $L_{ce-rvq}$. WaveNet is chosen as the architecture for the diffusion model due to its superior performance in audio quality (CMOS) compared to other architectures like the vanilla transformer, conformer, and UNet. It is enhanced with a Q-K-V attention and a FiLM layer for in-context learning. The system consists of three components: a prior model, a diffusion model, and an audio codec. Similar to [1], the prior model, which includes a phoneme encoder and a duration/pitch predictor, encodes text $y$ into a prior $c$. The diffusion model predicts the speech representation $z$ using $c$ as a condition, and the audio codec generates the speech signal $x$ from $z$.\"\n\n[1] Ren, Yi, et al. Fastspeech 2: Fast and high-quality end-to-end text to speech.", "category": "Method_Mechanics"}, {"question": "Does the paper claim that discrete token sequences must necessarily be longer than continuous sequences, and how does NaturalSpeech 2 address the issue of increased sequence length?", "answer": "The paper does not claim that discrete token sequences must necessarily be longer than continuous sequences. It points out that previous works using multiple RVQ discrete token sequences for speech representation result in a multiple-fold increase in length when flattened. It's true that some attempts have been made to address this issue, but they still have their own problems. AudioLM[1] predicts the first two RVQ token sequences in the first stage, and predicts the remaining RVQ token sequences in the second stage. However, this approach has two drawbacks: 1) it requires two separate AM models and a two-stage generation process; 2) it still needs the flattening of the code sequences, which also extends the sequence length. VALL-E [2] first predicts the first RVQ token sequence in an AR manner, and then predicts the other RVQ token sequences in a NAR manner, which makes the modeling much more complex compared with one-stage generation. Moreover, MusicGen[3] claims that such parallel modeling may impact system performance. In contrast, NaturalSpeech 2 uses a single continuous latent instead of multiple discrete tokens for each speech frame. It can prevent the increase in sequence length and support one-stage generation, which helps enhance overall performance. \n\n[1] Borsos, Zalán, et al. Audiolm: a language modeling approach to audio generation. \n[2] Wang Chengyi, et al. Neural codec language models are zero-shot text to speech synthesizers. \n[3] Copet, Jade, et al. Simple and Controllable Music Generation.", "category": "Method_Disambiguation"}, {"question": "What are the advantages of using a diffusion model over other non-autoregressive (NAR) models, and how does the performance comparison with FastSpeech2 demonstrate these advantages?", "answer": "The diffusion model incorporates multiple iterations during inference, enabling progressive refinement over these iterations, which enhances modeling capability and makes it particularly effective for high-quality generation. To demonstrate this, FastSpeech2[1] was adapted as a NAR baseline by adding cross-attention on speech prompts for zero-shot synthesis and changing the prediction target from mel-spectrogram to latent representation. Scaled to 400M parameters and trained on the same MLS dataset, the significant performance gap shown in Table 2 highlights the limitations of FastSpeech2's modeling capability compared to the diffusion model.\n\n[1] Ren, Yi, et al. Fastspeech 2: Fast and high-quality end-to-end text to speech.", "category": "Method_Comparison"}, {"question": "What does the term 'prior model' in Section 3.2 refer to, and how is it trained?", "answer": "The 'prior model' in Section 3.2 collectively refers to the phoneme encoder and the duration/pitch predictor, which are trained jointly as described in equation (6).", "category": "Concept_Understanding"}, {"question": "What are the variance values for the prosodic measures, specifically for pitch and duration, along with their 95% confidence intervals?", "answer": "For pitch, the measures (Mean, Std, Skew, Kurt) are $10.11_{\\pm 0.08}, 6.18_{\\pm 0.06}, 0.50_{\\pm0.01},1.01_{\\pm 0.03}$, respectively. For duration, the measures are $0.65_{\\pm 0.01}, 0.70_{\\pm 0.01}, 0.60_{\\pm 0.01}, 2.99_{\\pm 0.07}$, respectively. These results demonstrate that the improvements in Table 3 are statistically significant.", "category": "Experimental_Exposition"}, {"question": "Why is NaturalSpeech2 a good fit for ICLR despite being related to speech representation learning, which is typically addressed in speech-specific venues like InterSpeech or ICASSP?", "answer": "The NaturalSpeech 2 paper is fitting for ICLR because it addresses representation issues in large-scale TTS systems by using a single continuous sequence instead of multiple discrete token sequences, thereby resolving the dilemma between the codec and the acoustic model. Additionally, it introduces a new learning paradigm involving latent diffusion and continuous representation, which effectively addresses such issues in the field of speech representation learning in large-scale TTS. This research is aligned with the core interests of ICLR, particularly in the primary area of ‘representation learning for computer vision, audio, language, and other modalities’.", "category": "Method_Disambiguation"}, {"question": "What is the shortest prompt length in seconds that can be used for zero-speech synthesis while maintaining high-quality speech generation, as measured by SMOS on the LibriSpeech test set?", "answer": "The shortest prompt length tested is 1.0 seconds, which results in an SMOS score of 3.79 on the LibriSpeech test set. This demonstrates that NaturalSpeech 2 can generate high-quality speech even with a reduced prompt length, highlighting the system's robustness and stability.\n\n| Prompt Length| SMOS|\n| ------- | ---- | \n| 3.0 s  |  4.06  |\n| 2.0 s  |  3.98  |\n| 1.0 s  |  3.79  |", "category": "Experimental_Exposition"}, {"question": "What are the results of experiments conducted by varying model sizes and data sizes, and what do these results indicate about the importance of data size and model size in generating high-quality speech?", "answer": "The results of experiments conducted by varying model sizes and data sizes are as follows:\n| Model Size | Data Size | CMOS |\n| -------------- | ----- | ----- |\n| 400M | 44,000 h | 0.00 |\n| 400M | 10,000 h | -0.06 |\n| 400M | 1,000 h | -0.41 |\n| 60M | 1,000 h | -0.52 |\n\nThese results indicate that:\n1. Decreasing the data size from 44k hours to 1k hours while keeping the model size unchanged leads to a consistent drop in CMOS, showing that data size is important for generating high-quality speech.\n2. Decreasing the model size from 400M to 60M while keeping the data size unchanged results in a CMOS drop of 0.11, indicating that model size is also important.", "category": "Experimental_Exposition"}, {"question": "What are the additional metrics, such as WER and SMOS results, for NaturalSpeech 2 under different diffusion steps?", "answer": "The SMOS and WER results are reported in the table below:\n\n| Diffusion Step | RTF   | CMOS  | SMOS  | WER  |\n| -------------- | ----- | ----- | --------------- | ---- |\n| 150            | 0.366 | 0.00  | 4.06 | 2.01 |\n| 100            | 0.244 | -0.02 | 4.04 | 2.04 |\n| 50             | 0.124 | -0.08 | 3.98 | 2.09 |\n| 20             | 0.050 | -0.21 | 3.88 | 2.15 |\n\nThese results are in line with CMOS, further illustrating the trade-off between quality and inference speed.", "category": "Experimental_Exposition"}, {"question": "Did the paper experiment with training your model using pre-quantized versus post-quantized latent representations, and what were the outcomes in terms of performance and storage efficiency?", "answer": "In the preliminary experiment, the paper did not observe performance improvements when replacing the post-quantized representation with the pre-quantized representation.\n\nMoreover, using post-quantized representation is driven by engineering considerations. For example, rather than directly storing 256-dimensional FP32 (256 x 32 bits) latent representations, the paper only needed to store 16 INT16 quantized IDs (16 x 16 bits), leading to a 32x boost in storage efficiency.", "category": "Experimental_Analysis"}, {"question": "How is the duration of phonemes extracted in this work, and what is the source of the singing voice data?", "answer": "The singing voice data comes in the form of recorded audio, not musical scores. The duration of phonemes is extracted using the same alignment tool as described in the paper (see Appendix E for additional details)", "category": "Experimental_Setup"}, {"question": "What is the method used to extract pitch information, including the algorithm and parameters?", "answer": "The pitch of each frame is extracted using the pyworld[1] library with the dio + stonemask algorithm, and the parameters are set to default.\n\n[1] https://github.com/JeremyCCHsu/Python-Wrapper-for-World-Vocoder", "category": "Experimental_Setup"}, {"question": "How does the performance of the paper's audio codec implementation compare to the official SoundStream codec in terms of VISQOL score at 12.8kbps?", "answer": "The paper's implementation achieves a VISQOL score of 4.38, which is competitive with the official SoundStream paper's score of 4.26 at 12.8kbps.", "category": "Experimental_Exposition"}, {"question": "What is the definition of singing voice synthesis as used in the paper, and how does it differ from singing voice conversion?", "answer": "Singing voice synthesis in the paper's work involves generating singing from lyrics, ground truth duration, and ground truth pitch, rather than from recorded singing voice. This process is not singing voice conversion. The ground truth pitch and duration can be determined by a separately-trained neural network conditioned on musical scores, and the paper uses these during inference for simplification, following previous works [1,2].\n\n[1] Liu, Jinglin, et al. Diffsinger: Singing voice synthesis via shallow diffusion mechanism.\n\n[2] Huang, Rongjie, et al. Make-A-Voice: Unified Voice Synthesis With Discrete Representation.", "category": "Concept_Understanding"}, {"question": "Can the proposed method be adapted or refined to reduce its dependency on additional modules such as the prompt encoder, second attention block, CE-RVQ loss, and pitch predictor?", "answer": "1) For the prompt encoder, it is used to encode the prompt speech for high-quality zero-shot text-to-speech synthesis, a technique that is also widely used in previous works, such as [1,2,3]. 2) For the second attention block, it is crucial for NaturalSpeech 2 for high-quality synthesis. Intuitively, the prompt needs to provide the model with aspects like timbre, prosody, energy, background, etc. Thus, the paper adopts mm query tokens in the second attention block to model these important aspects. With experiments in Sec. 4.3 (without query attention), it can be observed that this component is beneficial to the audio generation quality. 3) For the CE-RVQ loss, it can significantly enhance the performance while incurring low computational cost. Intuitively, it is a regularization constraint inspired by the hierarchical structure of the audio codec. To ensure high-quality generation, the paper constrains the latent distribution and proposes the CE Loss by matching the predicted $z_0$ with each quantizer. With experiments in Sec. 4.3 (without CE loss), it can be observed that both the CMOS and WER will degrade by removing it, which demonstrates the effectiveness of CE-RVQ loss. 4) For the pitch predictor, it is a flexible module in NaturalSpeech 2. The paper has conducted the ablation experiment by removing the pitch predictor and observed no noticeable decline in performance. However, it significantly enhances the controllability of pitch, which is necessary for singing voice synthesis (refer to Sec. 4.4 for more information).\n\n[1] Arik, Sercan, et al. Neural voice cloning with a few samples. \n[2] Jia, Ye, et al. Transfer learning from speaker verification to multispeaker text-to-speech synthesis. \n[3] Jiang, Ziyue, et al. Mega-TTS: Zero-Shot Text-to-Speech at Scale with Intrinsic Inductive Bias.", "category": "Experimental_Exposition"}]}
{"id": "efFmBWioSc", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Multimodal Web Navigation with Instruction-Finetuned Foundation Models", "authors": ["Hiroki Furuta", "Kuang-Huei Lee", "Ofir Nachum", "Yutaka Matsuo", "Aleksandra Faust", "Shixiang Shane Gu", "Izzeddin Gur"], "abstract": "The progress of autonomous web navigation has been hindered by the dependence on billions of exploratory interactions via online reinforcement learning, and domain-specific model designs that make it difficult to leverage generalization from rich out-of-domain data.\nIn this work, we study data-driven offline training for web agents with vision-language foundation models.\nWe propose an instruction-following multimodal agent, WebGUM, that observes both webpage screenshots and HTML pages and outputs web navigation actions, such as click and type.\nWebGUM is trained by jointly finetuning an instruction-finetuned language model and a vision encoder with temporal and local perception on a large corpus of demonstrations.\nWe empirically demonstrate this recipe improves the agent's ability of grounded multimodal perception, HTML comprehension, and multi-step reasoning, outperforming prior works by a significant margin. \nOn the MiniWoB, we improve over the previous best offline methods by more than 45.8%, even outperforming online-finetuned SoTA, humans, and GPT-4-based agent. \nOn the WebShop benchmark, our 3-billion-parameter model achieves superior performance to the existing SoTA, PaLM-540B.\nFurthermore, WebGUM exhibits strong positive transfer to the real-world planning tasks on the Mind2Web.\nWe also collect 347K high-quality demonstrations using our trained models, 38 times larger than prior work, and make them available to promote future research in this direction.", "keywords": "Web Navigation;Foundation Models;Large Language Models;Instruction Finetuning;Decision Making;Multimodal Document Understanding", "qa_pairs": [{"question": "What measures were taken to ensure the high quality of the generated dataset, and is there evidence to support its quality?", "answer": "The dataset quality was maintained by running LLM agents with 10,000 episodes per task and retaining only successful trajectories, a practice known as reward filtering. This is a common practice in web navigation, as used by Humphreys et al. (2022) [1]. This ensures the dataset does not include noisy data from failure episodes. Additionally, qualitative analysis in Table 15 (Appendix L) compares episodes from the 347K dataset with a previous 12K dataset, showing that the paper's dataset exhibits 'shortest' behaviors compared to 'hesitant' behaviors (e.g. clicking the same checkbox several times) in the previous dataset, thus demonstrating controlled quality.\n\n[1]Humphreys et al., (2022) A data-driven approach for learning to control computers (https://arxiv.org/abs/2202.08137) ", "category": "Experimental_Setup"}, {"question": "What does 'rollout a LLM policy with 100 episodes per task' mean in the context of the data generation process, and what are the inputs and outputs of this process?", "answer": "'Rollout' refers to deploying WebN-T5 (the LLM policy) on MiniWoB environments, where WebN-T5 was tasked with solving 100 episodes per task across all 56 tasks. The inputs include HTML, instruction, and action history, and the output is a dictionary-format action, as described in Figure 2.", "category": "Concept_Understanding"}, {"question": "What are the reasons that a small model like WebGUM can outperform humans in general web navigation, and is this performance due to overfitting to a specific dataset(Table 1)?", "answer": "In web navigation literature, the number of parameters does not always correlate to the performance. For instance, CC-Net [1], trained through supervised learning and reinforcement learning, only has small-scale architectures compared to LLMs (transformer with 8 layers, 8 heads, and a 512-dimensional embedding, a dual-layer LSTM with 512 hidden units per layer, and 4 blocks ResNet). However, CC-Net can exhibit human-level web navigation performance (93.5%). Among the human-level web navigation methods (CC-Net, WebGUM, RCI/Synapse (GPT-3.5/4 with prompting)), there are trade-offs between model size, training cost/risk, and inference cost. The paper's method provides promising solutions among those by (1) mitigating training costs and risks through offline supervised finetuning and (2) achieving on-premise agents with low inference costs.\n\n| | CC-Net | WebGUM | RCI / Synapse |\n|--|--|--|--|\n| **Model Size** | small | up to 3B parameters | API-based private LLM |\n| **Training Cost/Risk** | High (RL) | Low (SL) | -- (in-context)|\n| **Inference Cost** | Low | Low | High |\n\nIn addition, the paper has conducted an out-of-distribution evaluation in Section 5.3 (Figure 5, and 6), and a transfer learning evaluation in Section 5.5 (Table 3). WebGUM performs the best among competitive baselines. In particular, WebGUM achieved 78.5% success rate for unseen combinational tasks, outperforming Synapse (73.8%), a SoTA web navigation agent in Table 1. In real-world action prediction tasks from the Mind2Web dataset, WebGUM exhibits strong transfer achieving the best performance in all the metrics. These results support that WebGUM is generalizable enough to other web navigation tasks beyond overfitting.\n\n[1] Humphreys et al., (2022) A data-driven approach for learning to control computers (https://arxiv.org/abs/2202.08137) ", "category": "Experimental_Analysis"}, {"question": "How does the proposed method, WebGUM, generalize to web navigation tasks involving HTML that exceeds the context length of 4096 tokens, and what strategies are suggested to handle longer HTML?", "answer": "WebGUM accepts 4096 tokens as input, which is the soft upper limit for dense attention models due to computational costs. For HTML exceeding this length, preprocessing is applied to shorten it, as demonstrated in the WebSRC evaluation. To handle much longer HTML, the authors suggest incorporating (1) ranking language models[1] to prioritize HTML elements, (2) extractive summarization of HTML[2], or (3) long-context pre-trained language models with sparse attention[3]. The applicability of such a pipelined approach was demonstrated by achieving the best performance among baselines on Mind2Web using cached ranking results.\n\n[1] Deng et al., (2023) Mind2Web: Towards a Generalist Agent for the Web (https://arxiv.org/abs/2306.06070)\n\n[2] Gur et al., (2023) A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (https://arxiv.org/abs/2307.12856)\n\n[3] Guo et al., (2021) LongT5: Efficient Text-To-Text Transformer for Long Sequences (https://arxiv.org/abs/2112.07916)\n", "category": "Method_Mechanics"}, {"question": "What failure cases have been included for the MiniWoB dataset, and what are the specific challenges or issues identified in these cases?", "answer": "The paper has included a failure example of WebGUM in MiniWoB in Appendix L. Specific challenges include (1) ambiguous instructions, such as small items ('count-shape'), (2) confusion with repetitive structures on the page ('social-media-all'), and (3) long-horizon tasks that may require memory ('guess-number'). Please refer to Figure 11 in Appendix L for details.", "category": "Experimental_Exposition"}, {"question": "What distinguishes WebGUM from recent multimodal LLMs like LLaVA[1] and InstructBLIP[2]?\n\n[1] Liu, Haotian, et al. \"Visual instruction tuning.\" arXiv preprint arXiv:2304.08485 (2023).\n\n[2] Dai, Wenliang, et al. “InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning”, arXiv preprint arXiv:2305.06500", "answer": "WebGUM differs from recent multimodal LLMs like LLaVA and InstructBLIP in its focus on web navigation, which is an interactive decision-making problem requiring temporal understanding and planning over multiple timesteps. In contrast, LLaVA and InstructBLIP focus on static inference tasks like visual reasoning or Science QA. Decision making for autonomy and Static VQA are orthogonal research fields that have their own specific challenges.", "category": "Method_Comparison"}, {"question": "Why did the paper choose Flan-T5 as the base language model for WebGUM instead of LLaMA or LLaMA2, especially when some prior MLLMs [1-3] are already using LLaMA as their base LM? \n\n[1] Liu, Haotian, et al. \"Visual instruction tuning.\" arXiv preprint arXiv:2304.08485 (2023).\n\n[2] Li, Bo, et al. \"Otter: A multi-modal model with in-context instruction tuning.\" arXiv preprint arXiv:2305.03726 (2023).\n\n[3] Dai, Wenliang, et al. “InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning”, arXiv preprint arXiv:2305.06500", "answer": "The paper has investigated the architectural choices for base language models, comparing Flan-T5 (220M, 770M, 3B, 11B) and Flan-PaLM-8B. Flan-T5 is the encoder-decoder model and Flan-PaLM is the decoder-only model. In Figure 4 (right), while Flan-T5 shows a clear scaling trend, Flan-PaLM-8B performs sub-optimally (72.8%), compared to similar scale Flan-T5 models (770M: 72.4%, 3B: 75.5%, 11B: 79.0%). The paper's results imply that in web navigation, encoder-decoder architecture (T5) is better at understanding HTML than the decoder-only model (PaLM, LLaMA, etc), since encoder-decoder architecture can capture the hierarchical and nested structure of HTML well. The paper's results also suggest that advanced language models are not always the best for downstream finetuning tasks. Because both Flan-T5 and Flan-PaLM are instruction-tuned on the same Flan dataset, the paper's comparison should be more fair than the case with LLaMA. Moreover, while recent LLMs employ larger model sizes (LLaMA-7B + ViT-L14 for LLaVA, LLaMA-7B/13B + ViT-G14 for InstructBLIP), WebGUM only has 3 billion parameters (Flan-T5-XL + ViT-B16). Such parameter efficiency is a desirable property for autonomous agent research because it ensures on-premise deployment and low latency. ", "category": "Motivation_Analysis"}]}
{"id": "Th6NyL07na", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models", "authors": ["Yung-Sung Chuang", "Yujia Xie", "Hongyin Luo", "Yoon Kim", "James R. Glass", "Pengcheng He"], "abstract": "Despite their impressive capabilities, large language models (LLMs) are prone to hallucinations, i.e., generating content that deviates from facts seen during pretraining. We propose a simple decoding strategy for reducing hallucinations with pretrained LLMs that does not require conditioning on retrieved external knowledge nor additional fine-tuning. Our approach obtains the next-token distribution by contrasting the differences in logits obtained from projecting the later layers versus earlier layers to the vocabulary space, exploiting the fact that factual knowledge in an LLMs has generally been shown to be localized to particular transformer layers. We find that this **D**ecoding by C**o**ntrasting **La**yers (DoLa) approach is able to better surface factual knowledge and reduce the generation of incorrect facts. DoLa consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving the performance of LLaMA family models on TruthfulQA by 12-17% absolute points, demonstrating its potential in making LLMs reliably generate truthful facts.", "keywords": "Large Language Models;Hallucination;Factuality;Decoding;Text Generation", "qa_pairs": [{"question": "Why do the authors restrict the candidate layers considered to only a subset of the non-final layers instead of using all preceding layers for computing the JS divergence?", "answer": "Using all layers can still improve the scores compared to the Vanilla baseline, but the improvement is not as significant as DoLa. The essential information to be contrasted is located in a certain part of the layers, and the bucket-based selection helps to find a suitable bucket within 2~4 validation tests. While using all layers can improve performance in most cases, it is not as significant as the improvement achieved with DoLa. The performance of TruthfulQA of 'DoLa - all layers' is shown below:\n\n|                   | 7b    |       |       | 13b   |       |       | 30b   |       |       | 65b   |       |       |\n|-------------------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|\n|                   | MC1   | MC2   | MC3   | MC1   | MC2   | MC3   | MC1   | MC2   | MC3   | MC1   | MC2   | MC3   |\n| Vanilla           | 25.58 | 40.55 | 19.20 | 28.27 | 43.32 | 20.85 | 31.70 | 49.52 | 24.23 | 30.84 | 46.88 | 22.74 |\n| DoLa - all layers | 31.95 | 63.86 | 31.17 | 30.48 | 62.25 | 30.96 | 29.13 | 61.50 | 30.69 | 30.48 | 62.00 | 31.67 |\n| DoLa              | 31.82 | 64.35 | 32.23 | 29.74 | 65.17 | 34.98 | 29.99 | 62.32 | 33.79 | 30.97 | 64.61 | 34.16 |", "category": "Motivation_Analysis"}, {"question": "In what sense is the search space considered large for DoLa-static, especially when there are only 10s of layers to consider?", "answer": "For smaller models like llama-7b/13b, the search space of DoLa-static is not very large. However, compared to DoLa, DoLa-static does have a larger search space. If users have resources for 10s of validation runs for a specific validation set, DoLa-static remains a good option. Without a specific validation set, DoLa is better at transferring to different data distributions, while DoLa-static is more sensitive to selected layers, as observed in Section 4.1.", "category": "Method_Disambiguation"}, {"question": "How do the computational costs of hyperparameter search and JS-divergence computation in DoLa and DoLa-static compare, and which method is more efficient depending on the use case?", "answer": "DoLa-static reduces inference latency by computing JS-divergence (JSD) less frequently but requires more validation runs. The advantages of DoLa and DoLa-static depend on the use case.\n\n- If it is a popular downstream application with millions of users and a fixed downstream task with a reliable validation set available, then the optimal inference time is the priority consideration and it is worthy to have tens of times more validation runs. In this case, DoLa-static is a good choice.\n\n- If it is a customized application that has a moderate amount of users (e.g. users in a specific small region, which is the usual case for even popular apps), then the cost of adapting the model to multiple user groups is significant, and an inference speed of 1.08x slower is relatively neglectable. In this case, DoLa will be preferred over DoLa-static.\n\nFrom the perspective of efficiency, DoLa and DoLa-static have their own advantages depending on different use cases. DoLa provides another trade-off option for better balancing the cost of validation and inference.", "category": "Method_Comparison"}, {"question": "How does the performance of the method change when all previous layers are considered at each sampling step, and why are the candidate layers restricted to a subset of the non-final layers?", "answer": "When all layers are considered, the method still improves the scores compared to the Vanilla baseline, but the improvement is not as significant as with DoLa. This suggests that the essential information to be contrasted is located in a certain part of the layers. The bucket-based selection helps to find a suitable bucket within 2~4 validation tests, and while using all layers can improve performance in most cases, it is not as significant as using DoLa. The performance of TruthfulQA of 'DoLa - all layers' is shown below.\n\n|                   | 7b    |       |       | 13b   |       |       | 30b   |       |       | 65b   |       |       |\n|-------------------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|\n|                   | MC1   | MC2   | MC3   | MC1   | MC2   | MC3   | MC1   | MC2   | MC3   | MC1   | MC2   | MC3   |\n| Vanilla           | 25.58 | 40.55 | 19.20 | 28.27 | 43.32 | 20.85 | 31.70 | 49.52 | 24.23 | 30.84 | 46.88 | 22.74 |\n| DoLa - all layers | 31.95 | 63.86 | 31.17 | 30.48 | 62.25 | 30.96 | 29.13 | 61.50 | 30.69 | 30.48 | 62.00 | 31.67 |\n| DoLa              | 31.82 | 64.35 | 32.23 | 29.74 | 65.17 | 34.98 | 29.99 | 62.32 | 33.79 | 30.97 | 64.61 | 34.16 |", "category": "Experimental_Analysis"}, {"question": "Why does the contrastive decoding performance in the proposed method (DoLa) not match the higher performance reported in ‘CONTRASTIVE DECODING IMPROVES REASONING IN LARGE LANGUAGE MODELS’, which used a smaller amateur model (1.5B parameters)[1]?\n\n[1] O'Brien, Sean, and Mike Lewis. \"Contrastive decoding improves reasoning in large language models.\" arXiv preprint arXiv:2309.09117 (2023).", "answer": "The Contrastive Decoding (CD) baseline in the reference paper [1] has two significant differences compared to our CD baseline. 1. **They train a small LLaMA-1.5B model as the amateur LM:** This paper [1] pretrains a Llama-1.5B model that is **not publicly released**. As we do not have that 1.5B model, we use the 7B model in our experiment. the authors try to use the 1.3B and 2.7B Sheared-LLaMA and 3B/7B [OpenLLaMA] as the amateur LMs. These are all of the publicly released small LLaMA models that share the same vocabulary as LLaMA (so that it can contrast them on the same vocab set). The results on GSM8K are shown below. \n\n| Method / Score (%)                     | LLaMA-7B | LLaMA-13B | LLaMA-33B | LLaMA-65B |\n|-----------------------------------------|----------|-----------|-----------|-----------|\n| Vanilla                                 | 10.77    | 16.68     | 33.81     | 51.18     |\n| + CD w/ LLaMA-7B                        | ---      | 9.10      | 28.43     | 44.05     |\n| + CD w/ OpenLLaMA-7B                    | 6.44     | 13.50     | 30.48     | 38.82     |\n| + CD w/ OpenLLaMA-7B_v2                 | 6.90     | 14.33     | 27.14     | 39.50     |\n| + CD w/ OpenLLaMA-3B                    | 6.60     | 11.07     | 27.60     | 41.77     |\n| + CD w/ OpenLLaMA-3B_v2                 | 8.11     | 11.52     | 29.34     | 40.33     |\n| + CD w/ Sheared-LLaMA-2.7B              | 5.00     | 14.10     | 32.30     | 47.08     |\n| + CD w/ Sheared-LLaMA-1.3B              | 9.02     | 16.38     | 34.87     | 46.40     |\n| + DoLa                                  | 10.46    | 18.04     | 35.41     | 53.60     |\n\nIt can be seen that using a small amateur LM, especially the 1.3B one, can improve the scores for contrastive decoding compared to using the 7B one as the amateur LM. However, the scores are still not better than DoLa.  The paper [1] may have put some effort into finding a suitable LLaMA-1.5B model that hits the sweet spot. The paper's experiments on these open-sourced small LLaMA models still cannot match the performance in [1], showing the difficulty in choosing amateur LMs for CD. \n\n2. **They introduce an extra hyperparameter β that does not exist in the original CD paper:** While the paper only follows the original version of contrastive decoding [2], in Section 2.1 of this paper [1], the authors propose to add an extra hyperparameter β, the strength of the amateur penalty, for better balancing between expert logits and amateur logits. Thus, the contrastive decoding formula will not be the same as the one in the original CD paper, unless β=∞.  The 65B CD result of GSM8K is 56.8 in [1] which is regarded much better than the paper's 65B CD baseline (44.0) and the paper's 65B DoLa (54.0). However, the authors argue that 44.0 is already a reasonable number that consistently with the results in [1] by the following evidence. In Table 1 from [1], the high GSM8K score of 65B (56.8) is under the setting of β = 0.5, while it drops significantly to 44.6 when setting β = 1.0, which is very close to the paper's 65B CD baseline (44.0) which essentially equals the situation of setting β=∞. Based on the above evidence, it is argued that the high score in [1] is probably achieved by carefully tuning β at a sweet spot. If following the original version of CD [2], the paper’s score (44.0) already matches the score shown in [1] with β=1.\n\n[2] Li, Xiang Lisa, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. \"Contrastive decoding: Open-ended text generation as optimization.\" arXiv preprint arXiv:2210.15097 (2022).", "category": "Experimental_Analysis"}, {"question": "Why does the proposed method (DoLa) show worse performance than the baselines (basic decoding and contrastive decoding) in Table 1 for TruthfulQA (MC1) with LLaMa-33B, and how does the sensitivity of the MC1 metric contribute to this inconsistency?", "answer": "The MC1 metric in TruthfulQA is highly sensitive because it is a 'winner takes all' metric. If any false answer has a higher score than the best true answer, the score becomes 0.0; otherwise, it is 1.0. Small fluctuations in model outputs on a single example can drastically change the score. In contrast, MC2 and MC3 are more stable as they consider all correct answers together. This sensitivity of MC1 explains the inconsistent results, and the authors conducted extensive experiments to demonstrate the overall trend of improvements across datasets and tasks.", "category": "Experimental_Analysis"}, {"question": "What are the results of using brute force for static and dynamic layer selection in the DoLa approach, and why is dynamic layer selection with brute force not feasible?", "answer": "1.  **Static layer selection with brute force, namely DoLa-static:** Using the same premature layer for the entire decoding process. The layer was selected by the overall scores on a validation set. The results are shown in Table 10 and 11. It can be seen that dynamic DoLa is slightly better than DoLa-static on CoT reasoning tasks, while DoLa-static can be slightly better than dynamic DoLa on FACTOR.\n2.  **Dynamic layer selection with brute force:** Applying the brute force idea at every time step of dynamic DoLa constitutes a very complex search problem. Given a validation example (e.g., a question in GSM8K), one must finish decoding the full-sentence answer (e.g., the CoT reasoning path in GSM8K) before calculating the score (e.g., exact match scores in GSM8K). For example, if there are 20 candidate layers to select from and 50 new tokens need to be generated, the possible combinations of dynamic layer selection strategies become $20^{50}$, making brute force infeasible. Furthermore, since the instant reward for each decoding step cannot be obtained until the full sentence is generated, it is difficult to perform operations like beam search to eliminate low-score search paths during the dynamic layer selection process. Even if the optimal path for a specific validation example is successfully obtained, transferring it to unseen examples in the testing set would be difficult. Applying brute force to dynamic layer selection remains an open question.\n", "category": "Experimental_Analysis"}, {"question": "Why is the selected layer for TruthfulQA closer to the final layer, while for FACTOR-News, the 0-th layer is a good premature layer, as shown in Figure 6?", "answer": "The examples in TruthfulQA are mainly short answers with difficult crucial facts. For example:\nQ: Where did fortune cookies originate?\nCorrect Answer: Fortune cookies originated in San Francisco.\nIncorrect Answer: Fortune cookies originated in China.\n\nIt is reasonable to surmise that the crucial facts in TruthfulQA require focusing more on directly contrasting with the layers right before the topmost layers where factual knowledge is located. In contrast, examples in FACTOR consist of long paragraph prefixes with 4 possible long sentence completions as multiple-choice options. These completions contain many ‘non-fact’ tokens that rely on general knowledge (e.g., linguistic knowledge), which is not necessarily located in the topmost layers. In this case, contrasting with the lower part of the layers can better handle all the tokens in the whole sentence.\n\nAlthough this is just a hypothesis, the best bucket of candidate layers for FACTOR/GSM8K/StrategyQA are all the same: the lower part of the layers, which also transfer well to VicunaQA. All these datasets involve long-sentence answers.", "category": "Experimental_Analysis"}, {"question": "What evidence supports the hypothesis that factual knowledge is injected at the top layers in DoLA, and how does this differ for easy-to-predict tokens?", "answer": "In Figure 2, when predicting important tokens related to factual information, the higher layers are more likely to be selected. For easy-to-predict tokens like function words, the word embedding layer is more likely to be selected. The claim that 'almost all layers selected will be the word embedding layer' applies only to easy-to-predict tokens, not all tokens.", "category": "Method_Disambiguation"}, {"question": "Are there quantitative results to support the observations presented in Figure 2?", "answer": "Yes, the authors conducted a quantitative study using the CoNLL-2003 name entity recognition dataset[1] with 3.25K examples. The authors calculated the layer with the largest JS-divergence with the final layer (referred to as the 'critical layer') when LLaMA-7B predicts the next token using teacher forcing. The results show that 75% of the time, the critical layer is layer 0 for non-entity tokens, while for entity tokens, only 35% of the time the critical layer is layer 0, and more than 50% of the time it is at a higher layer. This supports the observations in Figure 2.\n\n| Layer | Entity Tokens | Non-Entity Tokens |\n|-------|---------------|-------------------|\n|   0   |    35.56%     |      75.55%       |\n|   2   |    0.05%      |      0.08%        |\n|   4   |    0.94%      |      0.36%        |\n|   6   |    0.94%      |      0.14%        |\n|   8   |    1.05%      |      0.27%        |\n|  10   |    0.05%      |      0.33%        |\n|  12   |    2.10%      |      0.65%        |\n|  14   |    0.00%      |      0.33%        |\n|  16   |    0.00%      |      0.16%        |\n|  18   |    0.00%      |      0.05%        |\n|  20   |    1.69%      |      0.47%        |\n|  22   |    9.69%      |      1.76%        |\n|  24   |    10.38%     |      2.62%        |\n|  26   |    2.08%      |      2.17%        |\n|  28   |    10.06%     |      2.11%        |\n|  30   |    25.40%     |      12.98%       |\n\nNote that the authors use teacher forcing to send the ground truth into LLaMA to predict the next word for each token in the sentence. And the ground truth sentences are not generated by LLaMA. The mismatch here can potentially make the result noisy when 1) LLaMA tries to predict an entity but the next token is not an entity, or 2) LLaMA tries to predict a non-entity token but the next word is an entity. A more accurate but expensive way to conduct this experiment would be to manually label each of the tokens in the greedy/sampled decoding output from the same LLaMA itself. However, such a trend can be seen in the NER dataset from the current experiments.\n\n[1] Sang, Erik Tjong Kim, and Fien De Meulder. \"Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition.\" In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pp. 142-147. 2003. https://huggingface.co/datasets/conll2003", "category": "Experimental_Exposition"}, {"question": "What are the specific experimental settings and computational resources used in the latency analysis of DoLa, as described in Section 4.2?", "answer": "In the latency analysis in Section 4.2, the paper used 817 examples from TruthfulQA with a default 6-shot in-context demonstration prompt, resulting in an average input length of 250.3 tokens. The model was forced to decode 50 new tokens without stopping criteria. Efficiency was achieved through batched tensor operations for computing JS-divergence across all candidate layers simultaneously. Experiments were conducted on a machine with NVIDIA 32G V100 GPUs and 40-core Intel(R) Xeon(R) Platinum 8168 CPUs @ 2.70GHz. Models were run with 16-bit floating point precision and a batch size of 1. For 7b, 13b, 33b, and 65b models, 1, 2, 4, and 8 V100 GPUs were used, respectively. Cross-GPU inference with model weight sharding was managed using the Huggingface accelerate package.", "category": "Experimental_Setup"}, {"question": "What is the impact of DoLA on the language model’s generation quality, specifically in terms of grammaticality and cohesiveness, across different model sizes (7B, 13B, 33B, and 65B)?", "answer": "An additional study was conducted using GPT-4 to evaluate the quality of generated text, focusing on grammaticality and cohesiveness without considering factual correctness. The results indicate that for 7B, 13B, and 33B models, DoLA demonstrates better grammaticality and cohesiveness compared to the vanilla decoding baseline. For the 65B model, DoLA achieves a score nearly identical to vanilla decoding. Thus, when evaluating text generation quality without factuality, DoLA performs on par with (65B) or better than (7B/13B/33B) vanilla decoding.", "category": "Experimental_Exposition"}, {"question": "How was the efficiency of decoding from every layer measured in Section 4.2, including details on the tasks, hardware, and generation lengths used?", "answer": "The efficiency of decoding from every layer was measured using 817 examples from TruthfulQA with a default 6-shot in-context demonstration prompt, resulting in an average input length of 250.3 tokens. The model was forced to decode 50 new tokens without stopping criteria. Experiments were conducted on machines with NVIDIA 32G V100 GPUs and Intel(R) Xeon(R) Platinum 8168 CPUs @ 2.70GHz. Models were run with 16-bit floating point precision and a batch size of 1. For 7b, 13b, 33b, and 65b models, 1, 2, 4, and 8 V100 GPUs were used, respectively, with cross-GPU inference managed by the Huggingface accelerate package.", "category": "Experimental_Setup"}, {"question": "How does DoLA impact throughput and memory overhead compared to the vanilla decoding baseline?", "answer": "Additional experiments show that the memory overhead difference between DoLA and the vanilla decoding baseline is negligible, calculated as the difference between peak GPU memory during forward passes and the occupied GPU memory before the first forward pass. \n\n| Model       | LLaMA-7b | LLaMA-7b | LLaMA-13b | LLaMA-13b | LLaMA-30b | LLaMA-30b | LLaMA-65b | LLaMA-65b |\n|-------------|----------|----------|-----------|-----------|-----------|-----------|-----------|-----------|\n|             | Vanilla  | DoLa     | Vanilla   | DoLa      | Vanilla   | DoLa      | Vanilla   | DoLa      |\n| (a) GPU Memory Before Forward (MB) | 12916.5  | 12916.5  | 25025.8   | 25025.8   | 55715.7   | 55715.7   | 124682.6  | 124682.6  |\n| (b) Peak GPU Memory During Forward (MB) | 13233.9  | 13385.7  | 25510.7   | 25674.8   | 57057.5   | 57390.2   | 126950.0  | 127606.8  |\n| (b) - (a) GPU Memory Overhead (MB) | 317.4    | 469.2    | 484.9     | 681.6     | 1341.9    | 1674.5    | 2267.4    | 2924.3    |\n| [(b) - (a)] / (a) GPU Memory Overhead (%) | 2.5%     | 3.6%     | 1.9%      | 2.7%      | 2.4%      | 3.0%      | 1.8%      | 2.4%      |\n\n\nFor throughput, measured in tokens/sec for various models (7B/13B/33B/65B), the differences are generally less than 7%, similar to the increased latency extent in Section 4.2.\n\n| Model | Baseline (tokens/sec) | DoLa (tokens/sec) | Ratio of DoLa/Baseline |\n|---|---|---|---|\n| 7B | 22.03 | 20.83 | ×0.95 |\n| 13B | 12.94 | 12.03 | ×0.93 |\n| 33B | 6.82 | 6.38 | ×0.94 |\n| 65B | 3.11 | 3.08 | ×0.99 |\n", "category": "Experimental_Exposition"}, {"question": "Why does DoLA not discourage the model from generating the same token as the previous one?", "answer": "Because LLaMA’s word embedding layer is not tied with the LM head, feeding the word embedding of the previous token directly into the LM head results in an output distribution ($q_0(x_t)$) that does not put a substantial amount of mass on the previous token ($x_t$). Instead, the output distribution is closer to a non-contextual prediction of the next token ($x_{t+1}$) conditioned only on the previous token ($x_t$), similar to a bi-gram language model ($P(x_t+1|(x_t))$). Therefore, DoLA does not discourage the model from generating the same token as the previous one.", "category": "Method_Disambiguation"}, {"question": "How does the performance of a baseline that always selects the 0-th layer compare to DoLA, and how does this comparison demonstrate the effectiveness of DoLA in improving performance?", "answer": "The authors compared their method to a baseline that always selects the 0-th layer. Contrasting with the 0-th layer improves performance because the 0-th layer predictions are similar to a bi-gram LM, which contains non-contextual language statistics bias learned from pretraining data. By contrasting the final predictions with the 0-th layer, the non-contextual bias is removed, emphasizing the contextual information injected in the middle layers, leading to unbiased and better final predictions. Using only the 0-th layer already improves performance compared to the vanilla decoding baseline, and DoLa, which incorporates information from multiple layers, further increases the scores.", "category": "Method_Comparison"}, {"question": "What is the impact of the proposed method (DoLa) on inference speed, specifically in terms of latency and throughput, compared to the baseline decoding method?", "answer": "For the latency analysis, the paper uses 817 examples from TruthfulQA with the default 6-shot in-context demonstration prompt, which has an average input length of 250.3 after concatenating the prompt with the questions. The paper forces the model to decode 50 new tokens without any stopping criteria. All experiments were run on a machine equipped with NVIDIA 32G V100s and 40-core Intel® Xeon® Platinum 8168 CPUs @ 2.70GHz. The models were run with 16-bit floating point precision and a batch size of 1. For the 7B/13B/33B/65B models, 1/2/4/8 V100 GPUs were used, respectively. The cross-GPU inference with model weight sharding was handled by the Huggingface Accelerate package. [1]\n\nHere is the latency (ms/token) and throughput (tokens/sec) for 4 models.\n\nLatency:\n| Model | Baseline (ms/token) | DoLa (ms/token) | Ratio of DoLa/Baseline |\n|-------|---------------------|-----------------|-----------------|\n| 7B    | 45.4                | 48.0            | ×1.06           |\n| 13B   | 77.3                | 83.1            | ×1.08           |\n| 33B   | 146.7               | 156.7           | ×1.07           |\n| 65B   | 321.6               | 324.9           | ×1.01           |\n\nThroughput:\n| Model | Baseline (tokens/sec) | DoLa (tokens/sec) | Ratio of DoLa/Baseline |\n|-------|-----------------------|-------------------|------------------------|\n| 7B    | 22.03                 | 20.83             | ×0.95                  |\n| 13B   | 12.94                 | 12.03             | ×0.93                  |\n| 33B   | 6.82                  | 6.38              | ×0.94                  |\n| 65B   | 3.11                  | 3.08              | ×0.99                  |\n\nThe results show that the speed is slightly slower than the vanilla decoding baseline by 1%~8%, which is in an acceptable range. This result suggests that DoLa is a practical decoding strategy that has the potential to be applied to real-world applications.\n\n[1] https://huggingface.co/docs/accelerate/concept_guides/big_model_inference", "category": "Method_Comparison"}, {"question": "What happens if the residual connection between the premature layer and the final layer is removed, and would this approach be beneficial without significantly reducing inference speed?", "answer": "Removing the residual connections of the premature layer could potentially downplay the early layers and emphasize the higher layers without causing latency. Initial experiments showed that removing the residual connection from one of the attention layers or MLP layers resulted in the decoding output becoming a sequence of repeated tokens.For example, when removing the MLP residual connection of layer 12, the result was:\n\n`Question: Eliza's rate per hour for the first 40 hours she works each week is $10. She also receives an overtime pay of 1.2 times her regular hourly rate. If Eliza worked for 45 hours this week, how much are her earnings for this week?`\n\n`Model Output: s that that that that s that s s s that that s that that s that that s s s that that that s that that that that s that that that that that that …`. \n\nThis suggests that removing residual connections causes a significant distribution shift for the hidden states, leading to out-of-distribution inputs for the next layer and unusual model behavior. Manipulating continuous hidden states is more challenging than manipulating output logits, as the latter are more explainable. However, this approach could be worth further exploration with additional fine-tuning to address the distribution shift issue.", "category": "Experimental_Analysis"}, {"question": "How does the inference speed of the DoLa method compare to the baseline across different model sizes, as measured in latency (ms/token) and throughput (tokens/sec)?", "answer": "For the latency analysis, the paper uses 817 examples from TruthfulQA with the default 6-shot in-context demonstration prompt, which has an average input length of 250.3 after concatenating the prompt with the questions. The model is forced to decode 50 new tokens without any stopping criteria. All experiments were run on a machine equipped with NVIDIA 32G V100s and 40-core Intel® Xeon® Platinum 8168 CPUs @ 2.70GHz. The models were run with 16-bit floating point precision and a batch size of 1. For the 7b, 13b, 33b, and 65b models, 1, 2, 4, and 8 V100 GPUs were used, respectively. Cross-GPU inference with model weight sharding was handled using the Hugging Face Accelerate package.[1]\n\nHere is the latency (ms/token) and throughput (tokens/sec) for 4 models.\n\nLatency:\n| Model | Baseline (ms/token) | DoLa (ms/token) | Ratio of DoLa/Baseline |\n|-------|---------------------|-----------------|-----------------|\n| 7B    | 45.4                | 48.0            | ×1.06           |\n| 13B   | 77.3                | 83.1            | ×1.08           |\n| 33B   | 146.7               | 156.7           | ×1.07           |\n| 65B   | 321.6               | 324.9           | ×1.01           |\n\nThroughput:\n| Model | Baseline (tokens/sec) | DoLa (tokens/sec) | Ratio of DoLa/Baseline |\n|-------|-----------------------|-------------------|------------------------|\n| 7B    | 22.03                 | 20.83             | ×0.95                  |\n| 13B   | 12.94                 | 12.03             | ×0.93                  |\n| 33B   | 6.82                  | 6.38              | ×0.94                  |\n| 65B   | 3.11                  | 3.08              | ×0.99                  |\n\nIt can be seen that the speed is slightly slower than the vanilla decoding baseline by 1%–8%, which falls within an acceptable range. This result suggests that DoLa is a practical decoding strategy with the potential to be applied in real-world applications.\n\n[1] https://huggingface.co/docs/accelerate/concept_guides/big_model_inference", "category": "Method_Comparison"}]}
{"id": "TyFrPOKYXw", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "Safe RLHF: Safe Reinforcement Learning from Human Feedback", "authors": ["Josef Dai", "Xuehai Pan", "Ruiyang Sun", "Jiaming Ji", "Xinbo Xu", "Mickel Liu", "Yizhou Wang", "Yaodong Yang"], "abstract": "With the development of large language models (LLMs), striking a balance between the performance and safety of AI systems has never been more critical. However, the inherent tension between the objectives of helpfulness and harmlessness presents a significant challenge during LLM training. To address this issue, we propose Safe Reinforcement Learning from Human Feedback (Safe RLHF), a novel algorithm for human value alignment. Safe RLHF explicitly decouples human preferences regarding helpfulness and harmlessness, effectively avoiding the crowd workers' confusion about the tension and allowing us to train separate reward and cost models. We formalize the safety concern of LLMs as an optimization task of maximizing the reward function while satisfying specified cost constraints. Leveraging the Lagrangian method to solve this constrained problem, Safe RLHF dynamically adjusts the balance between the two objectives during fine-tuning. Through a three-round fine-tuning using Safe RLHF, we demonstrate a superior ability to mitigate harmful responses while enhancing model performance compared to existing value-aligned algorithms. Experimentally, we fine-tuned the Alpaca-7B using Safe RLHF and aligned it with collected human preferences, significantly improving its helpfulness and harmlessness according to human evaluations.\n\nCode is available at https://github.com/PKU-Alignment/safe-rlhf.\n\n\nWarning: This paper contains example data that may be offensive or harmful.", "keywords": "Safe Reinforcement Learning;Reinforcement Learning from Human Feedback;Large Language Model;AI Safety", "qa_pairs": [{"question": "What were the experimental results comparing the paper's method to additional safety RLHF baselines such as Constitutional AI and Safety SFT?", "answer": "Win rates for three different methods compared to Aplaca-7B rated by GPT-4:\n| | Safe RLHF (the paper's; Beaver-v1) | Constitutional AI | Safety SFT |\n| :---: | :---: | :---: | :---: |\n| Helpfulness Win Rate | 71.8% | 40.2% | 53.6% |\n| Harmlessness Win Rate | 73.3% | 69.7% | 53.6% |\n\nthe experiments reveal interesting insights:\n\n**Helpfulness and Harmlessness Balance:** Neither Constitutional AI (40.2% helpfulness, 69.7% harmlessness) nor Safety SFT (53.6% for both) matched the performance of the paper's Safe RLHF (Beaver-v1) approach, which achieved a helpfulness win rate of 71.8% and harmlessness win rate of 73.3%. This underscores **the challenge in balancing helpfulness and harmlessness** in large language models and highlights the efficacy of the paper's approach.\n\nImpact of SFT on RLHF: the findings suggest that while the SFT stage contributes to model safety, the gains from the RLHF stage, especially with advanced reward models like Beaver-v1, are more pronounced. **This distinction is crucial in understanding the relative impacts of SFT and RLHF stages on model performance.**\n\nThe experiments demonstrate that while incorporating safety in the SFT stage is beneficial, **the RLHF stage offers a more substantial improvement in balancing helpfulness and harmlessness**. The inclusion of Constitutional AI and Safety SFT provides a comprehensive view of the current landscape of the paper's Safe RLHF method and validates the effectiveness of the paper's approach.\n\n**Implementation details of Constitutional AI and Safety SFT:**\n- **Constitutional AI:** The authors followed the Constitutional AI paper to revise the raw responses based on constitutional critiques. The Constitutional AI method requires a large model with the capability to follow complex instructions to generate critiques for the responses and then revise the responses. The authors use the GPT-3.5 model to revise the responses generated by Alpaca-7B using the same constitutional prompts from the Constitutional AI paper. The authors let the reward model tend to prefer the revised response over the original response. And apply the standard RLHF pipeline over the Alpaca-7B model.\n\n- **Safety SFT:** For the Safety SFT method, the authors first filter out preference pairs in the preference dataset used in this study if both responses are labeled unsafe. Then the authors finetune the Alpaca-7B model to fit the safer response in the preference dataset.", "category": "Method_Comparison"}, {"question": "What are the relative impacts of the SFT and RLHF stages on model performance, particularly in terms of balancing helpfulness and harmlessness?", "answer": "The experiments show that while the SFT stage contributes to model safety, the RLHF stage provides a more substantial improvement in balancing helpfulness and harmlessness. The inclusion of new baselines validates the effectiveness of the paper's Safe RLHF method, with advanced reward models like Beaver-v1 demonstrating more pronounced gains in the RLHF stage. This distinction helps in understanding the relative impacts of the SFT and RLHF stages on model performance.", "category": "Experimental_Analysis"}, {"question": "What other safe reinforcement learning baselines were considered apart from Sparrow and reward shaping, and how do they compare to the proposed Safe RLHF (Beaver-v1) approach?", "answer": "Additional experiments were conducted with Constitutional AI and Safety SFT baselines. Constitutional AI achieved 40.2% helpfulness and 69.7% harmlessness, while Safety SFT achieved 53.6% for both metrics. Neither baseline matched the performance of the proposed Safe RLHF (Beaver-v1) approach, which achieved a helpfulness win rate of 71.8% and harmlessness win rate of 73.3%. This highlights the challenge in balancing helpfulness and harmlessness and underscores the efficacy of the proposed approach.", "category": "Method_Comparison"}, {"question": "What are the reasons for not training stronger supervised fine-tuning (SFT) models, given that some safety-related data was gathered before performing RLHF?", "answer": "The reasons for not training stronger SFT models are: (1) The Alpaca team has made their data, training code, and final model open source, which facilitates replication by the academic community. (2) There is insufficient expert data available to train a stronger SFT model. (3) Using Alpaca-7B as the starting model and applying three rounds of Safe RLHF fine-tuning allows observation of the effects of Safe RLHF under different levels of helpfulness and harmlessness.", "category": "Motivation_Analysis"}, {"question": "How is model selection conducted during the reward model (RM)/cost model training stage and the reinforcement learning with human feedback (RLHF) training stage?", "answer": "Model selection is common in RLHF to ensure **the correctness** at every step. To ensure **the fairness**, the paper applied this engineering trick to all types of RLHF algorithms used in the paper's experiments, including Safe RLHF, RLHF (PPO), Reward Shaping, and ablation experiments. In fact, the paper's approach of model selection enhances the fairness of comparisons as it mitigates randomness and ensures each algorithm operates correctly.\n**For both the reward model and cost model, the selection of models primarily aims to achieve higher prediction accuracy on the test set.** For different parameter training outcomes, the paper evaluates their predictive accuracy on the reserved test set and selects the one with the highest accuracy for the next step. Typically, an accuracy above 75% is considered acceptable. With a fixed dataset, the impact of different hyper-parameters on the reward model and cost model is not significant. Therefore, the paper does not perform model selection repeatedly many times at this stage. For the best hyperparameters, please refer to the Appendix G.1.\n**For the RL phase, model selection primarily aims to prevent the over-optimization [1].** Since the reward model and cost model are learned from human preference data, their ability to correctly predict has a range. Continuously training with the same reward and cost models can easily lead to the phenomenon of reward hacking. Therefore, the model selection during the RL phase mainly involves evaluating models at different training steps within the same training cycle to identify the point where the model is sufficiently trained but not over-optimized. Existing evaluations for alignment rely on GPT-4 and human evaluators, and due to their high financial and time cost, the paper opts for a rapid model selection process that involves human evaluation on a small test set combined with a trained unified score model (as mentioned in the Section 5.2.1 Model-based Evaluations). Only models deemed sufficiently good are then tested with the entire test set using GPT-4 and human evaluations.\nAlso, due to the issue of over-optimization, it is necessary to continually collect data from new models, train new preference models, and perform multiple RLHF iterations.\n\n\n**References:**\n\n[1] Gao, Leo, John Schulman, and Jacob Hilton. \"Scaling laws for reward model overoptimization.\" International Conference on Machine Learning. PMLR, 2023.", "category": "Experimental_Setup"}, {"question": "What is the significance of the red dashed line in Figure 2(a), and how does the proposed cost model fit the ambiguous boundary between safety and unsafety in human values?", "answer": "The red dashed line in Figure 2(a) represents the ambiguous boundary between safety and unsafety in human values. The proposed cost model fits this boundary by introducing a comparison loss between responses and a virtual response $y_0$ near the safety threshold, setting the cost of $y_0$ to zero (Equation 4). This approach maintains the original properties of the Bradley-Terry model and is fundamental for integrating RLHF with Safe RL algorithms.", "category": "Method_Mechanics"}, {"question": "What are the specific contributions of the Safe RLHF framework, and how does it differ from standard CMDP formulations and existing Lagrangian methods in Safe RL?", "answer": "The Safe RLHF framework is the first work to combine Safe RL with the RLHF framework. It includes decoupled data collection, training of two different preference models (reward model and cost model), and fine-tuning through the integration of Safe RL. It formalizes safety as a constraint and dynamically adjusts it during training to navigate the tension between helpfulness and harmlessness. Logically, the human value towards LLMs is first decoupled into two parts: helpfulness and harmlessness. After modeling harmlessness as a constraint that needs to be prioritized, the entire problem is formalized under the CMDP framework. Reversing this logical order is a disheartening accusation. It overshadows the paper’s efforts in guiding the crowdworkers to generate decoupling data reflecting their preferences for helpfulness and harmlessness; utilizing the reward model and the paper’s proposed novel cost model to represent objective and constraint signals; integrating Safe RL with the training of LLMs. While standard CMDPs encompass both reward and cost signals, the paper's innovation lies in how these signals are modeled, acquired, and utilized for LLM alignment. The Lagrangian method is acknowledged as a mature technology in traditional Safe RL, but it is only one component of the Safe RLHF framework.  The paper does not claim the introduction of the Lagrangian method to solve Safe RL problem as the paper's contribution. The contributions include: Fine-tuning with Safe RLHF framework effectively enhances the helpfulness and harmlessness of LLMs under human values (Section 5.2.1). Decoupling helpfulness from harmlessness effectively increases agreement among crowd workers when annotating preference data, reducing bias due to the contradiction (Section 5.2.2). Compared to a fixed ratio of optimizing for helpfulness and harmlessness, dynamically adjusting the trade-off between them during training more effectively guides the inherent tension between the two (Section 5.2.3). Providing preference-based scores, rather than just a binary safe/unsafe classifier, improves LLM safety more efficiently (Section 5.2.4).", "category": "Method_Disambiguation"}, {"question": "How does the paper define the trade-off between safety and preference performance in LLM alignment, and why is the cost-reward formulation of CMDP necessary instead of a simpler reward function, for example, one where A is preferred over B only if (preference(A > B) and (A passes a safety check threshold))?", "answer": "Firstly, many prior works [1,2,3,4] have noted that the trade-off between helpfulness and harmlessness during training remains a critical challenge in the field of LLM alignment. This is due to the inherent tension between these two optimization goals and the complex, high-dimensional nature of human values that are hard to represent with rule-based approaches. Therefore, this is not a task that can be overlooked or simplified. Secondly, it is also believed that harmlessness should be prioritized over helpfulness. As stated in the paper, the trade-off that needs to be achieved during training is: to maximize helpfulness as much as possible, while ensuring that the output aligns with human values of safety. Therefore, the paper models safety as a constraint that must be met first. Naturally, compared to the traditional RLHF's MDP framework, it becomes necessary to formalize this problem within the CMDP framework. Thirdly, using a simple reward format to express both helpfulness and harmlessness is the approach taken by traditional RLHF. As many previous works [1,2,3,4] have mentioned, this approach poses significant challenges during data annotation and training due to the tension between helpfulness and harmlessness. Meanwhile, in section 5.2.2, the paper empirically demonstrated the shortcomings of the traditional single-reward RLHF method. Compared to Safe RLHF, it results in lower data agreement, as well as reduced efficiency in optimizing both helpfulness and harmlessness, as illustrated in Figure 6(a). \n\nReferences: \n[1] Kenton, Zachary, et al. 'Alignment of language agents.' \n[2] Bai, Yuntao, et al. 'Training a helpful and harmless assistant with reinforcement learning from human feedback.' \n[3] Glaese, Amelia, et al. 'Improving alignment of dialogue agents via targeted human judgements.' \n[4] Zhao, Wayne Xin, et al. 'A survey of large language models.'", "category": "Method_Mechanics"}, {"question": "Why is the conclusion that Lagrangian methods are superior to reward shaping for LLMs fine-tuning valid?", "answer": "The authors tested seven different reward shaping coefficients (ν = 0.01, 0.5, 1, 2, 5, 10, 100) to cover the coefficient space. Extremely high and low values led to over-optimization of one objective at the expense of the other, while moderate values (ν = 1, 2) were ineffective in resolving the conflict between helpfulness and harmlessness, with their enhancements remaining subpar compared to Safe RLHF. These results demonstrate that Safe RLHF outperforms reward shaping across the tested coefficient range, supporting the conclusion that Lagrangian methods are superior.", "category": "Experimental_Analysis"}, {"question": "What are the limitations and potential risks associated with Safe RLHF, particularly in scenarios where the algorithm might fail to balance helpfulness and harmlessness effectively?", "answer": "The limitations of Safe RLHF include a lack of high-quality SFT data for SFT and PTX loss, no specific optimization for multi-turn dialogue scenarios, and the need for additional mechanisms (such as pre- and post-check strategies) in real-world applications to achieve a safer AI system. These limitations are discussed in detail in Appendix A of the paper.", "category": "Method_Disambiguation"}, {"question": "How does the proposed RLHF framework contribute to the field of Safe RL and LLM alignment, and what are the key contributions and experimental findings of the Safe RLHF methodology?", "answer": "The paper’s primary goal is to propose a new RLHF framework that addresses the challenging tension between helpfulness and harmlessness goals in the field of LLM alignment, rather than introducing a new Safe RL algorithm. Instead, the paper aims to broaden the application scope of Safe RL algorithms, bringing new perspectives to both fields.\n\nIts contributions:\n\nSafe RLHF is the first work to combine Safe RL with the RLHF framework. It is a comprehensive methodology that includes decoupled data collection, training of two different preference models, and fine-tuning through the integration of Safe RL. It formalizes safety as a constraint and dynamically adjusts it during training, aiming to navigate the inherent tension between helpfulness and harmlessness.\n\nThe paper’s extensive experiments demonstrate the following conclusions: Fine-tuning with the Safe RLHF framework effectively enhances the helpfulness and harmlessness of LLMs under human values (Section 5.2.1). Decoupling helpfulness from harmlessness effectively increases agreement among crowd workers when annotating preference data, reducing bias due to the contradiction (Section 5.2.2). Compared to using a fixed ratio for optimizing helpfulness and harmlessness, dynamically adjusting the trade-off between them during training more effectively manages the inherent tension between the two (Section 5.2.3). Providing preference-based scores, rather than just a binary safe/unsafe classifier, improves LLM safety more efficiently (Section 5.2.4).", "category": "Method_Disambiguation"}, {"question": "How does the reward and cost model in Section 4.2 of Safe RLHF differ from the classic reward and cost signals used in traditional Safe Reinforcement Learning (Safe RL)?", "answer": "Safe RLHF differs from traditional Safe RL in the following ways: (1) Human values are high-dimensional and complex, making it impossible to design reward and cost functions in a rule-based manner. Instead, human preferences must be modeled from collected datasets. This involves guiding crowd workers to provide high-quality annotations (Section 4.1), constructing prompt datasets for large language models (Section 5.1 Prompts and Red-teaming), controlling the distribution of prompt-response pairs in the dataset (Section 5.1 Preference Datasets), and training better preference models (Section 4.2). (2) The threshold for what is considered safe in human values cannot be formulaically defined. Therefore, a new training method for a Cost Model is proposed to model the safety threshold. This method, based on the Bradley-Terry model, introduces a comparison loss between responses and a virtual response $y_0$ near the safety threshold, setting the cost of $y_0$ to zero. This design separates safe from unsafe responses at a cost boundary of zero (as shown in Figure 2(a)), making zero the safety threshold for an individual response.", "category": "Method_Comparison"}, {"question": "Given the high-dimensional complexity of LLMs and the randomness in training preference models, how can the stability and convergence of learning in Safe RLHF be ensured, especially in the absence of theoretical guarantees?", "answer": "Theoretical guarantees for the convergence of RLHF are challenging due to the high-dimensional complexity of LLMs and the randomness in training preference models. This is a common scenario in this research field. Second, in the RLHF framework, training often needs to be early stopped before convergence due to the issue of over-optimization [1]. The reward model and cost model are learned from human preference data. They are proxies of human preference (or: preferences). Their ability to accurately predict scores is limited to a finite range. Continuously training on the same reward model and cost model can easily lead to a phenomenon known as ‘reward hacking’ [2]. Thus, an early stop before over-optimization is necessary. In the experiments, the authors also observed that when LLMs converge with the current reward model and cost model, it usually indicates the occurrence of reward hacking. Third, empirically, through three iterations of Safe RLHF, the authors have validated that this framework is stable and efficient for enhancing both the helpfulness and harmlessness of large language models simultaneously (Section 5.2.1). In these experiments, the primary measure for stability is conducting periodic evaluations with human evaluators and early stopping training in cases of over-optimization.\n\nReferences: [1] Gao, Leo, John Schulman, and Jacob Hilton. 'Scaling laws for reward model overoptimization.' International Conference on Machine Learning. PMLR, 2023. \n\n[2] Di Langosco, L. L., Koch, J., Sharkey, L. D., Pfau, J., & Krueger, D. Goal misgeneralization in deep reinforcement learning. In International Conference on Machine Learning (pp. 12004-12019). PMLR.", "category": "Method_Mechanics"}, {"question": "Can Lagrangian safe RL methods, such as TRPO-Lag and PPO-Lag, be applied within the Safe RLHF framework?", "answer": "First-order algorithms like PPO-Lag, FOCOPS, and P3O are promising for application within the Safe RLHF framework. However, second-order algorithms such as TRPO-Lag and CPO, which require solving the inverse of the Fisher matrix, may demand excessive computational resources for large language models (LLMs).", "category": "Method_Mechanics"}, {"question": "What measures have been taken to ensure the stability of the reward and cost signals in the reward model and cost model?", "answer": "The stability of the reward and cost signals is ensured through the following measures: 1) Dataset Quality: Additional optimizations were implemented in the annotation process and data distribution. 2) Randomness in Deep Learning: Predictive accuracy is evaluated on a reserved test set, and the model with the highest accuracy (>75%) is selected for the next step. 3) Over-optimization: Periodic testing is conducted, and training is stopped prematurely to prevent over-optimization leading to preference model failure.", "category": "Method_Mechanics"}, {"question": "How does Equation (29) align with the Lagrange multiplier approach, given that it appears to use a different method of differentiation?", "answer": "It is important to highlight that Equation (29) is consistent with the Lagrange method and its implementation in the code. The authors have applied a common engineering trick. Since $\\lambda \\geq 0$, they set $\\lambda_k \\doteq e^{\\eta_k}$. By using $\\eta$ as the actual update parameter, they ensure $\\lambda \\geq 0$. Therefore, according to Equation (12): $$ \\eta_{k+1} = \\eta_k + \\alpha\\frac{\\partial}{\\partial \\eta} e^\\eta\\mathcal J(\\theta_k)\\Big |_{\\eta_k} = \\eta_k + \\alpha \\cdot \\mathcal J(\\theta_k) e^{\\eta_k} $$ Substituting in $\\eta_k = \\ln \\lambda_k$, the equation becomes: $$ \\ln \\lambda_{k+1} = \\ln \\lambda_k + \\alpha \\cdot \\lambda_k \\cdot \\mathcal J_C(\\theta_k) $$", "category": "Concept_Understanding"}, {"question": "What are the computational resources, time requirements, and costs associated with the Safe RLHF training process, particularly for large-scale language models?", "answer": "The experiments used a 7-billion-parameter large language model. The server had an Intel(R) Xeon(R) Platinum 8378A CPU @ 3.00GHz with 64 cores and 8 NVIDIA A800-SXM4-80GB GPUs with NVLink support. Data collection took approximately two weeks for crowdworker annotation and one week for quality control. Training for a single round of Safe RLHF required 10 to 20 hours, depending on inference length. The total cost for data annotations was around $70,000, and the training equipment cost was about $120,000. All data and code are released to help reduce research costs.", "category": "Experimental_Setup"}, {"question": "How many values of cost threshold were tried in the experiments, and how robust is the Lagrangian method for RLHF to different threshold values?", "answer": "The authors did not try multiple values of the cost threshold. Instead, they introduced a new training methodology for the Cost Model, which computes a reference threshold directly. This threshold is calculated by estimating cost values for the LLM on the training set, adjusting values exceeding zero down to zero, and computing their mean. This approach ensures robustness and avoids the need to search for the threshold in the real number space. The negative thresholds in the second and third rounds of Safe RLHF reflect the fact that most responses on the training set were predicted as safe (negative numbers) by the Cost Model.", "category": "Experimental_Setup"}, {"question": "Was the total cost for data annotations and training equipment for Safe RLHF around 190,000 U.S. dollars?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the detailed computational efficiency and scalability discussion planned to be included in the appendix of the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the training for a single round of Safe RLHF require between 10 to 20 hours, varying with the average length of inference?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper provide detailed information on the computational resources and time required for the Safe RLHF training process?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the average time for a single round of Safe RLHF data collection approximately two weeks for crowdworker annotation and one week for professional quality control?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "AqN23oqraW", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "KoLA: Carefully Benchmarking World Knowledge of Large Language Models", "authors": ["Jifan Yu", "Xiaozhi Wang", "Shangqing Tu", "Shulin Cao", "Daniel Zhang-Li", "Xin Lv", "Hao Peng", "Zijun Yao", "Xiaohan Zhang", "Hanming Li", "Chunyang Li", "Zheyuan Zhang", "Yushi Bai", "Yantao Liu", "Amy Xin", "Kaifeng Yun", "Linlu Gong", "Nianyi Lin", "Jianhui Chen", "Zhili Wu", "Yunjia Qi", "Weikai Li", "Yong Guan", "Kaisheng Zeng", "Ji Qi", "Hailong Jin", "Jinxin Liu", "Yu Gu", "Yuan Yao", "Ning Ding", "Lei Hou", "Zhiyuan Liu", "Bin Xu", "Jie Tang", "Juanzi Li"], "abstract": "The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth of LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, and applicable evaluations. Given the importance of world knowledge to LLMs, we construct a Knowledge-oriented LLM Assessment benchmark (KoLA), in which we carefully design three crucial factors: (1) For ability modeling, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering 19 tasks. (2) For data, to ensure fair comparisons, we use both Wikipedia, a corpus prevalently pre-trained by LLMs, along with continuously collected emerging corpora, aiming to evaluate the capacity to handle unseen data and evolving knowledge. (3) For evaluation criteria, we adopt a contrastive system, including overall standard scores for better numerical comparability across tasks and models, and a unique self-contrast metric for automatically evaluating knowledge-creating ability. We evaluate 21 open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset will be updated every three months to provide timely references for developing LLMs and knowledge-related systems.", "keywords": "Large Language Model;World Knowledge;Evolving Benchmark", "qa_pairs": [{"question": "How does the paper differentiate between knowledge understanding and knowledge application, and why are tasks like reading comprehension classified under knowledge understanding rather than knowledge attention? Additionally, why does knowledge understanding primarily involve extraction tasks?", "answer": "The paper differentiates between knowledge understanding and knowledge application based on Bloom's Cognitive Taxonomy[1]. Understanding involves determining the meaning of instructional messages (e.g., explaining or summarizing), while applying involves carrying out or using a procedure in a given situation (e.g., implementing or executing). Indeed, since the revision of Bloom's theory [2], there has been considerable debate about the demarcation between these two cognitive levels. A key principle for differentiation is whether the task involves information and abilities beyond the given content[3]. Tasks involving information extraction(which ensures that the output information is faithfully extracted from the input text) are categorized as understanding because they ensure that the output information is faithfully extracted from the input text. Knowledge understanding is not limited to extraction tasks; it also includes conceptual abstraction, as seen in the COPEN task. Reading comprehension is classified as a knowledge understanding task if it requires only knowledge from the text. If it involves more complex background knowledge and reasoning, it may lean towards knowledge application.\n\n[1] Krathwohl D R. A revision of Bloom's taxonomy: An overview[J]. Theory into practice, 2002, 41(4): 212-218.\n\n[2] Forehand M. Bloom's taxonomy: Original and revised[J]. Emerging perspectives on learning, teaching, and technology, 2005, 8: 41-44.\n\n[3] Wilson L O. Anderson and Krathwohl Bloom’s taxonomy revised understanding the new version of Bloom’s taxonomy[J]. The Second Principle, 2016, 1(1): 1-8.", "category": "Method_Mechanics"}, {"question": "How is the self-contrast metric applied in the knowledge creation task as illustrated in Figure 2?", "answer": "The self-contrast metric is applied in three steps: (1) The model generates a completion $T$ using only the context $C$(The Battle of Evesham marked the defeat of Simon de Montfort...), allowing it to freely create content. (2) The model generates another completion $T_{k}$ using both the context $C$ and event knowledge $K$ (i.e., the structured content in the middle part of the figure). (3) The rationality of the model's event knowledge creation is calculated by contrasting $T$ and $T_{k}$, $\\partial \\left( T, T_{k} \\right)$ (its ability to freely create subsequent event knowledge, e.g., *Prince Edward’s Escape*), which, along with two other contrast metrics, leads to the scoring of the knowledge creation level.", "category": "Method_Mechanics"}, {"question": "What are the detailed costs involved in conducting the studies for each season, and how are these costs managed to sustain the project?", "answer": "The costs for each season are kept below USD 2,000 and include two main parts: data annotation and model deployment/API calls. Data annotation costs approximately USD 660-700 for triple annotation and USD 400-420 for event annotation. Due to the scale of the KoLA test sets, the costs for model deployment and API calls have been kept within an acceptable range. Model deployment costs are approximately USD 200-240 (GPU usage), and API calls cost USD 150-180. For the two completed seasons, costs have not exceeded USD 1,600 per season. A budget of USD 2,000 per season is anticipated to be sufficient for future seasons.", "category": "Experimental_Setup"}, {"question": "How will the evolving benchmark be maintained in the long term, and how will it interact with the known dataset?", "answer": "The authors are committed to updating the evaluation for at least six seasons, after which they will introduce new versions of KoLA. Each new season's data will be up-to-date to prevent test data leakage and maintain evaluation fairness. Ended seasons will not be added to the known set immediately but will be publicly released after six seasons, followed by a retrospective analysis. The maintenance involves collecting and annotating new data, constructing new model lists, and publishing results. A maintenance team has been established, and costs are controlled through automated pre-annotation and evaluation set size management. The cost per season is expected to be under USD 2,000, and community contributions are welcomed.", "category": "Experimental_Setup"}, {"question": "What are the key differences between the KoLA benchmark and existing LLM benchmarks, and why is KoLA necessary despite the availability of numerous existing benchmarks?", "answer": "(1) As described in Section 1, rather than assessing the breadth of model capabilities like most existing benchmarks (such as MMLU), KoLA focuses more on analyzing the deep interconnections of model knowledge capabilities through carefully-designed multi-level testing. (2) Besides combining some existing tasks, KoLA maintains a unique evolving mechanism, aimed at minimizing data leakage. Recent studies [1,2,3] have widely revealed a serious issue of 'test set contamination' in the evaluation of many LLMs with existing benchmarks, where the results may not genuinely reflect the actual capabilities of the models. These features make KoLA distinct and necessary in the current landscape of LLM evaluation.\n\nRef:\n\n[1] Wei T, Zhao L, Zhang L, et al. Skywork: A More Open Bilingual Foundation Model[J]. arXiv preprint arXiv:2310.19341, 2023.\n\n[2] Yang S, Chiang W L, Zheng L, et al. Rethinking Benchmark and Contamination for Language Models with Rephrased Samples[J]. arXiv preprint arXiv:2311.04850, 2023.\n\n[3] Zhou K, Zhu Y, Chen Z, et al. Don't Make Your LLM an Evaluation Benchmark Cheater[J]. arXiv preprint arXiv:2311.01964, 2023. ", "category": "Method_Comparison"}, {"question": "What are the motivations and benefits of using standardized overall scoring in the evaluation of large language models (LLMs) compared to simple ranking, especially considering the potential variations caused by the addition of new models?", "answer": "Standardized scores have become the mainstream choice in fields like educational assessment and intelligence testing [1], with systems like the SAT, GRE, and Wechsler Intelligence Scales (IQ Test) adopting them. Compared to absolute values, using standard scores in the multi-task evaluation of LLMs offers the following benefits:1. Standardized scores enable fair comparisons across different seasons and datasets. The varying difficulty of tasks and the different sensitivities of metrics mean that the absolute numerical performance of models is incomparable. For example, if there is a dataset where models typically perform at 80, and another where the usual performance is 40, directly comparing a model's performance scores on these two datasets cannot reflect whether the model is better at one task over the other. However, comparing standardized scores derived from relative performance can yield conclusive insights.\n2. For overall rankings, standardized scores can negate the bias caused by specific tasks. For instance, without standardized scoring, ChatGLM (6B) would rank much higher in overall rankings solely due to its high performance in the 2-2 task, surpassing its improved version ChatGLM2-32k.\n\n Although the inclusion of new models causes variations in standardized scores, these benefits make such variations acceptable. KoLA aims to provide a reliable reference for model selection, believing that the relative downscaling of evaluations due to superior new models is beneficial for users.\n\n[1]Haladyna T M, Nolen S B, Haas N S. Raising standardized achievement test scores and the origins of test score pollution[J]. Educational Researcher, 1991, 20(5): 2-7.", "category": "Motivation_Analysis"}, {"question": "How do the conclusions related to knowledge in this paper contribute to its overall contributions, given that they do not seem directly tied to the paper's stated goals?", "answer": "The experimental conclusions contribute in two key ways: (1) the analysis based on Bloom’s cognitive taxonomy provides a novel perspective on the knowledge capabilities of LLMs, and (2) the results offer new experimental support for widely discussed speculations, such as the alignment tax. While some conclusions may not be novel, the evaluation framework and the long-term utility of the KoLA benchmark add significant value.", "category": "Experimental_Analysis"}]}
{"id": "YCWjhGrJFD", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Training Diffusion Models with Reinforcement Learning", "authors": ["Kevin Black", "Michael Janner", "Yilun Du", "Ilya Kostrikov", "Sergey Levine"], "abstract": "Diffusion models are a class of flexible generative models trained with an approximation to the log-likelihood objective. However, most use cases of diffusion models are not concerned with likelihoods, but instead with downstream objectives such as human-perceived image quality or drug effectiveness. In this paper, we investigate reinforcement learning methods for directly optimizing diffusion models for such objectives. We describe how posing denoising as a multi-step decision-making problem enables a class of policy gradient algorithms, which we refer to as denoising diffusion policy optimization ( DDPO), that are more effective than alternative reward-weighted likelihood approaches. Empirically, DDPO can adapt text-to-image diffusion models to objectives that are difficult to express via prompting, such as image compressibility, and those derived from human feedback, such as aesthetic quality. Finally, we show that DDPO can improve prompt-image alignment using feedback from a vision-language model without the need for additional data collection or human annotation. The project’s website can be found at http://rl-diffusion.github.io.", "keywords": "reinforcement learning;RLHF;diffusion models", "qa_pairs": [{"question": "How does this paper's contribution differ from DPOK and 'Optimizing ddpm sampling with shortcut fine-tuning'[1]?\n\n[1]Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning. arXiv preprint arXiv:2301.13362, 2023.", "answer": "This paper was the first to propose a policy gradient objective for training diffusion models, as highlighted in the related work and methods sections. Unlike 'Optimizing ddpm sampling with shortcut fine-tuning,' which focused on improving fast sampling using a GAN-like objective and included small-scale experiments with unconditional diffusion models, this work demonstrates that RL can optimize various objectives, including prompt alignment with VLM-derived rewards. The experiments illustrate the flexibility and usefulness of this approach, addressing a different problem.", "category": "Method_Comparison"}, {"question": "How does the experimental setup address the concern that using the same Aesthetic Quality metric for both fine-tuning and evaluation may not accurately reflect improvements in image generation quality?", "answer": "The experimental evaluation includes out-of-domain assessment by training using ImageReward and evaluating using Aesthetic Score, which shows an increase in both metrics across all prompts. Additionally, an experiment in Appendix B applying universal guidance to the aesthetic predictor was conducted, resulting in a non-negligible increase in aesthetic score, though it is much smaller than DDPO finetuning and significantly slower(2 minutes per image on an A100)  due to the backpropagation of gradients through the large guidance network.", "category": "Experimental_Analysis"}, {"question": "Does the aesthetic predictor bug affect the fairness of the quantitative comparison in the experiments, and are there any other bugs in the paper that might impact the completeness of the results?", "answer": "The aesthetic predictor bug, caused by hardware-specific numerical instability, affected the scale of the rewards and the style of the resulting images but did not compromise the fairness of the quantitative comparison because it affected all methods equally. The primary comparison (Figure 4) was re-run after resolving the bug, and the results remain valid. The anonymized code provided is written in PyTorch, and the results have been successfully replicated on GPUs, indicating that the results are not limited to customized devices like TPU. There is no evidence of other bugs in the paper.", "category": "Method_Disambiguation"}, {"question": "Regarding the capability of DDPO, is it possible to control the image generation to have predicted reward? For example, Classifier-Free guidance provides control over sample fidelity and diversity. It seems like DDPO improves prompt fidelity at the expense of sample diversity", "answer": "The most direct way to control the amount of reward during generation in DDPO is through early stopping. Earlier checkpoints in the training process have lower reward but are closer to the distribution of the pretrained model.", "category": "Method_Mechanics"}, {"question": "How does the difficulty of prompts affect the optimization analysis in the context of the reward function normalization used by DDPO?", "answer": "Some prompts are naturally more difficult for the pretrained model than others. For instance, generating 'a man riding a horse' is easier than generating 'a horse riding a man.' Figure 5 quantitatively demonstrates this, showing that 'riding a bike' prompts achieve the highest average reward, followed by 'playing chess,' and then 'washing dishes.'", "category": "Method_Mechanics"}, {"question": "How does the proposed method address the issue of overoptimization, particularly in scenarios where visual inspection is not feasible? For instance, one might not be able to apply this method to a medical imaging task, where visual inspection by a human supervisor does not allow for determining when to stop the optimization process.", "answer": "It is reasonably straightforward to run The paper's method for a fixed number of iterations (e.g., 15k reward queries) and recover good generations; completely degenerated images were not observed in the VLM or aesthetic quality experiments; they were only observed in compressibility and incompressibility tests when run for long enough. Overoptimization is an open research problem that is well outside the scope of this paper. While it is certainly a major issue that hinders many real-life applications (such as the provided medical imaging example), the issue of overoptimization also exists in language modeling to an equivalent degree ([Gao et. al., 2022](https://arxiv.org/abs/2210.10760)), yet it has certainly not prevented RLHF from having a huge impact in that domain ([Ouyang et. al., 2022](https://arxiv.org/abs/2203.02155); [Bai et. al., 2022](https://arxiv.org/abs/2212.08073)). Current techniques for combating overoptimization, such as KL-regularization or automated evaluation, do not completely eliminate the need for human evaluation in real-life systems.", "category": "Method_Disambiguation"}, {"question": " Was a classifier guidance baseline included in the evaluation, and how does it compare to the proposed method?", "answer": "This paper includes a comparison to classifier guidance in Appendix B. However, only the aesthetic predictor reward function is differentiable, making classifier guidance applicable in this case. In the subsequent experiment, universal guidance applied to the aesthetic predictor showed a non-negligible increase in aesthetic score, but it was much smaller than DDPO finetuning and significantly slower (2 minutes per image on an A100) due to the backpropagation of gradients through the large guidance network.", "category": "Experimental_Analysis"}, {"question": "How does the paper approach the implementation and effectiveness of the early stopping mechanism in their method?", "answer": "The authors found it straightforward to run their method for a fixed number of iterations (e.g., 15k reward queries) and recover good generations. They did not observe completely degenerated images in VLM or aesthetic quality experiments, except in compressibility and incompressibility tests when run for extended periods. Held-out metrics could be used to automatically determine a stopping point.", "category": "Method_Mechanics"}, {"question": "How does DDPO compare to DPOK in terms of performance when trained on multiple prompts, and what evidence supports the claim that DDPO has advantages over DPOK?", "answer": "The paper conducted an experiment directly comparing DDPO to DPOK tasks, and the results are included in Appendix C . The quantitative and qualitative results show that DDPO outperforms the numbers reported in the DPOK paper, even without KL regularization. Since DPOK has not released code, the exact cause of this performance difference is unclear, but it suggests that the design decisions in the paper's implementation of DDPO are more effective for RL training of diffusion models.", "category": "Method_Comparison"}, {"question": "Is the anonymized code provided by the authors written in PyTorch and successfully replicated on GPUs?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the bug in the aesthetic predictor caused by hardware-specific numerical instability that affected the scale of rewards and style of resulting images?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the results of the paper achievable only on customized devices like TPU?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the bug in the aesthetic predictor experiments affect the fairness of the quantitative comparison between different methods?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "6PmJoRfdaK", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models", "authors": ["Yukang Chen", "Shengju Qian", "Haotian Tang", "Xin Lai", "Zhijian Liu", "Song Han", "Jiaya Jia"], "abstract": "We present LongLoRA, an efficient fine-tuning approach that extends the context sizes of pre-trained large language models (LLMs), with limited computation cost.\nTypically, training LLMs with long context sizes is computationally expensive, requiring extensive training hours and GPU resources. For example, training on the context length of 8192 needs 16x computational costs in self-attention layers as that of 2048. In this paper, we speed up the context extension of LLMs in two aspects. On the one hand, although dense global attention is needed during inference, fine-tuning the model can be effectively and efficiently done by sparse local attention. The proposed shifted sparse attention effectively enables context extension, leading to non-trivial computation saving with similar performance to fine-tuning with vanilla attention. Particularly, it can be implemented with only two lines of code in training, while being optional in inference. On the other hand, we revisit the parameter-efficient fine-tuning regime for context expansion. Notably, we find that LoRA for context extension works well under the premise of trainable embedding and normalization. LongLoRA combines this improved LoRA with S^2-Attn. LongLoRA demonstrates strong empirical results on various tasks on Llama2 models from 7B/13B to 70B. LongLoRA extends Llama2 7B from 4k context to 100k, or Llama2 70B to 32k on a single 8x A100 machine. LongLoRA extends models' context while retaining their original architectures, and is compatible with most existing techniques, like Flash-Attention2. In addition, we further conduct supervised fine-tuning with LongLoRA and our long instruction-following LongAlpaca dataset. All our code, models, dataset, and demo are available at https://github.com/dvlab-research/LongLoRA.", "keywords": "Efficient fine-tuning;Long context;Large language model", "qa_pairs": [{"question": "What additional evaluations were conducted beyond perplexity and retrieval settings to assess the performance of the proposed model?", "answer": "Additional evaluations were conducted on the LongBench and L-Eval benchmarks[1]. The Llama2 7B[2] model was fine-tuned with the proposed method on long QA data and compared with GPT-3.5-Turbo and other Llama2-based long-context models, including Vicuna and LongChat models. The results, presented in Table 9 and Table 10, show that the 7B model achieved comparable or better performance than these models, with supervised fine-tuning taking about 4 hours and 0.3 billion tokens on a single 8x A100 machine.\n\nTable 1 - Evaluation on LongBench English tasks\n\n| Model | Avg | Single-Doc QA | Multi-Doc QA | Summarization | Few-shot Learning | Code | Synthetic |\n| --- | --- | --- | --- | --- | --- | --- | --- |\n| GPT-3.5-Turbo | 44.0 | 39.8 | 38.7 | 26.5 | 67.1 | 54.1 | 37.8 |\n| Llama2-7B-chat | 31.0 | 24.9 | 22.6 | 24.7 | 60.0 | 48.1 | 5.9 |\n| LongChat-v1.5-7B | 34.3 | 28.7 | 20.6 | 26.7 | 60.0 | 54.1 | 15.8 |\n| Vicuna-v1.5-7B | 31.9 | 28.0 | 18.6 | 26.0 | 66.2 | 47.3 | 5.5 |\n| Ours-7B | 36.8 | 28.7 | 28.1 | 27.8 | 63.7 | 56.0 | 16.7 |\n\nTable 2 - Evaluation on L-Eval open-ended tasks, i.e., comparing models to GPT-3.5-Turbo and judging win rates via GPT-4\n\n| Model | Win-rate | Wins | Ties |\n| --- | --- | --- | --- |\n| LongChat-7B | 33.68 | 36 | 56 |\n| LongChat-v1.5-7B | 33.59 | 38 | 53 |\n| Vicuna-v1.5-7B | 25.52 | 22 | 54 |\n| Ours-7B | 39.06 | 45 | 60 |\n\n[1] Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li: LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding. CoRR abs/2308.14508 (2023)\n\n[2] Chenxin An, Shansan Gong, Ming Zhong, Mukai Li, Jun Zhang, Lingpeng Kong, Xipeng Qiu: L-Eval: Instituting Standardized Evaluation for Long Context Language Models. CoRR abs/2307.11088 (2023)", "category": "Experimental_Exposition"}, {"question": "What ablation experiments were conducted to determine the optimal group size for different target sequence lengths, particularly for context lengths longer than 8192?", "answer": "Additional ablation experiments were performed using a context length of 16384. The results showed that setting the group size to 1/4 of the context length is optimal for efficiency-accuracy tradeoff.\n\n| Context Length | Full | 1/2 | 1/4 | 1/6 | 1/8 |\n| --- | --- | --- | --- | --- | --- |\n| 8192 | 8.02 | 8.04 | 8.04 | 8.10 | 8.16 |\n| 16384 | 7.82 | 7.84 | 7.86 | 7.94 | 7.98 |", "category": "Experimental_Exposition"}, {"question": "What method or formula was used to estimate the model FLOPs presented in Table 11?", "answer": "The context stage FLOPs of Llama2-7B were profiled using a batch size of 1 and various context lengths with the third-party tool torchprofile[1]. This tool traces the computation graph and sums up the FLOPs of each node in the graph, including Q/K/V/O projections, multi-head self-attention, fully-connected layers, and normalization layers.\n\n[3] torchprofile. https://github.com/zhijian-liu/torchprofile", "category": "Experimental_Setup"}, {"question": "What does it mean that the model is able to handle longer context in the retrieval-based evaluation section?", "answer": "Retrieval-based evaluation demonstrates that LongLoRA extends the context window of a pre-trained LLM. Specifically, Llama-2-7b fails to retrieve the passkey when the context window exceeds 4k, while LongLoRA maintains a high retrieval accuracy (60-90%) even at much longer context lengths of 33k-45k.", "category": "Concept_Understanding"}, {"question": "How is the performance of different attention patterns evaluated in terms of retaining the original standard self-attention at inference time, as shown in Table 6 ?", "answer": "In Table 6, each attention pattern's performance is evaluated under two protocols: using sparse attention in both training and testing, and using sparse attention only for training with standard full attention for testing. $S^2$-Attn achieves the best perplexity with the original full self-attention at inference time, as detailed in the table caption.", "category": "Experimental_Setup"}, {"question": "Was an additional baseline that trains LoRA + embeddings included in the experiments, and what were the results?", "answer": "Yes, an additional baseline that trains LoRA + embeddings was included in the experiments, as shown in Table 2. Finetuning the embedding layer provides larger benefits than normalization layers, with the results indicating improved performance.\n\n| Embed | x  | √  | √  |\n|-------|----|----|----|\n| Norm  | √  | x  | √  |\n| PPL   | 10.49 | 8.29 | 8.12 |\n\n\n", "category": "Experimental_Exposition"}, {"question": "How does the computational cost and efficiency of LongLoRA compare to training long-context language models from scratch?", "answer": "Training long-context LLMs from scratch is computationally expensive and unaffordable for most researchers. LongLoRA offers an efficiency advantage of orders of magnitude compared to training long-context LLMs from scratch, requiring significantly fewer resources. For example, Llama-2 models need training with 2 trillion tokens across hundreds of GPUs, while LongLoRA models are finetuned on about 2 billion tokens for 32k context length using just 8 A100 GPUs. Recent research by Meta, LLama2-Long[1], supports that long-context continual training on short-context models is more efficient and similarly effective compared to pretraining from scratch with long sequences, further emphasizing LongLoRA's importance.\n\n[1] Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, Madian Khabsa, Han Fang, Yashar Mehdad, Sharan Narang, Kshitiz Malik, Angela Fan, Shruti Bhosale, Sergey Edunov, Mike Lewis, Sinong Wang, Hao Ma: Effective Long-Context Scaling of Foundation Models. CoRR abs/2309.16039 (2023)", "category": "Method_Comparison"}, {"question": "Why do regular LoRA and LoRA+ (Table 12) use the same amount of memory, and does $S^2$-Attn reduce memory usage in addition to reducing flops?", "answer": "$S^2$-Attn does not reduce memory usage on top of LoRA (LoRA+) in Table 11 because FlashAttention-2[1] is used, which fuses all operators in multi-head self-attention (MHSA) into a single CUDA kernel, avoiding writing the attention score matrix to DRAM. The memory space required for MHSA is approximately $2 * B * N * C$, where B is the batch size, N is the sequence length, and C is the number of channels. Without FlashAttention-2, vanilla dense attention requires $2 * B * N * C + B * N^2 * C$ memory space, where $B * N^2 * C$ corresponds to attention scores. In comparison, $S^2$-Attn requires $2 * B * N * C + 1/4 * B * N^2 * C$ memory space. Empirically, $S^2$-Attn improves training speed by 2.1x and memory usage by 1.8x under a context length of 8192. At a 16k context length, dense attention runs out of memory, while $S^2$-Attn remains feasible. This comparison is included in Table 13 of the revision.\n\n| $S^2$-Attn | Train hours (8192) | Memory - GB (8192) | Train hours (16384) | Memory - GB (16384) |\n| --- | --- | --- | --- | --- |\n| x | 17.5 | 55.5 | OOM | OOM |\n| √ | 8.2 | 30.3 | 20.8 | 57.1 |\n\n[1] Tri Dao: FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. CoRR abs/2307.08691 (2023)", "category": "Experimental_Analysis"}, {"question": "What additional generative tasks requiring longer context were evaluated in the paper, and how does the proposed model perform on these tasks compared to other models?", "answer": "The paper includes evaluations on LongBench[1] and L-Eval[2], which consist of comprehensive generative tasks such as document QA, summarization, few-shot learning, code completion synthetic tasks, and other open-ended tasks in L-Eval. The proposed model, fine-tuned on Llama2 7B with long QA data, shows comparable or better performance than other models. \n\nTable 1 - Evaluation on LongBench English tasks.\n\n| Model | Avg | Single-Doc QA | Multi-Doc QA | Summarization | Few-shot Learning | Code | Synthetic |\n| --- | --- | --- | --- | --- | --- | --- | --- |\n| GPT-3.5-Turbo | 44.0 | 39.8 | 38.7 | 26.5 | 67.1 | 54.1 | 37.8 |\n| Llama2-7B-chat | 31.0 | 24.9 | 22.6 | 24.7 | 60.0 | 48.1 | 5.9 |\n| LongChat-v1.5-7B | 34.3 | 28.7 | 20.6 | 26.7 | 60.0 | 54.1 | 15.8 |\n| Vicuna-v1.5-7B | 31.9 | 28.0 | 18.6 | 26.0 | 66.2 | 47.3 | 5.5 |\n| Ours-7B | 36.8 | 28.7 | 28.1 | 27.8 | 63.7 | 56.0 | 16.7 |\n\nTable 2 - Evaluation on L-Eval open-ended tasks, i.e., comparing models to GPT-3.5-Turbo and judging win rates via GPT-4.\n\n| Model | Win-rate | Wins | Ties |\n| --- | --- | --- | --- |\n| LongChat-7B | 33.68 | 36 | 56 |\n| LongChat-v1.5-7B | 33.59 | 38 | 53 |\n| Vicuna-v1.5-7B | 25.52 | 22 | 54 |\n| Ours-7B | 39.06 | 45 | 60 |\n\n[1] Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li: LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding. CoRR abs/2308.14508 (2023)\n\n[2] Chenxin An, Shansan Gong, Ming Zhong, Mukai Li, Jun Zhang, Lingpeng Kong, Xipeng Qiu: L-Eval: Instituting Standardized Evaluation for Long Context Language Models. CoRR abs/2307.11088 (2023)", "category": "Method_Comparison"}, {"question": "What generative tasks were evaluated, and how does the proposed model perform compared to other models on these tasks?", "answer": "The paper has included the evaluation on LongBench[1] and L-Eval [2], as shown in Table 9 and Table 10. The paper fine-tunes the Llama2 7B model using its long QA data. In these benchmarks, there are comprehensive generative tasks, including document QA, summarization, few-shot learning, code completion synthetic tasks, and other open-ended tasks in L-Eval. The paper's model presents comparable or even better performance than other counterparts, with about 4 hours for fine-tuning on 8 A100 GPUs. It takes 60 million tokens per epoch, 5 epochs, and 0.3 billion tokens in total for the supervised fine-tuning.\n\nTable 1 - Evaluation on LongBench English tasks.\n\n| Model | Avg | Single-Doc QA | Multi-Doc QA | Summarization | Few-shot Learning | Code | Synthetic |\n| --- | --- | --- | --- | --- | --- | --- | --- |\n| GPT-3.5-Turbo | 44.0 | 39.8 | 38.7 | 26.5 | 67.1 | 54.1 | 37.8 |\n| Llama2-7B-chat | 31.0 | 24.9 | 22.6 | 24.7 | 60.0 | 48.1 | 5.9 |\n| LongChat-v1.5-7B | 34.3 | 28.7 | 20.6 | 26.7 | 60.0 | 54.1 | 15.8 |\n| Vicuna-v1.5-7B | 31.9 | 28.0 | 18.6 | 26.0 | 66.2 | 47.3 | 5.5 |\n| Ours-7B | 36.8 | 28.7 | 28.1 | 27.8 | 63.7 | 56.0 | 16.7 |\n\nTable 2 - Evaluation on L-Eval open-ended tasks, i.e., comparing models to GPT-3.5-Turbo and judging win rates via GPT-4.\n\n| Model | Win-rate | Wins | Ties |\n| --- | --- | --- | --- |\n| LongChat-7B | 33.68 | 36 | 56 |\n| LongChat-v1.5-7B | 33.59 | 38 | 53 |\n| Vicuna-v1.5-7B | 25.52 | 22 | 54 |\n| Ours-7B | 39.06 | 45 | 60 |\n\n[1] Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, Juanzi Li: LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding. CoRR abs/2308.14508 (2023)\n\n[2] Chenxin An, Shansan Gong, Ming Zhong, Mukai Li, Jun Zhang, Lingpeng Kong, Xipeng Qiu: L-Eval: Instituting Standardized Evaluation for Long Context Language Models. CoRR abs/2307.11088 (2023)\n\n", "category": "Method_Comparison"}, {"question": "Did the paper evaluate their model only on perplexity and retrieval settings?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "KOZu91CzbK", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "Retroformer: Retrospective Large Language Agents with Policy Gradient Optimization", "authors": ["Weiran Yao", "Shelby Heinecke", "Juan Carlos Niebles", "Zhiwei Liu", "Yihao Feng", "Le Xue", "Rithesh R N", "Zeyuan Chen", "Jianguo Zhang", "Devansh Arpit", "Ran Xu", "Phil L Mui", "Huan Wang", "Caiming Xiong", "Silvio Savarese"], "abstract": "Recent months have seen the emergence of a powerful new trend in which large language models (LLMs) are augmented to become autonomous language agents capable of performing objective oriented multi-step tasks on their own, rather than merely responding to queries from human users. Most existing language agents, however, are not optimized using environment-specific rewards. Although some agents enable iterative refinement through verbal feedback, they do not reason and plan in ways that are compatible with gradient-based learning from rewards. This paper introduces a principled framework for reinforcing large language agents by learning a retrospective model, which automatically tunes the language agent prompts from environment feedback through policy gradient. Specifically, our proposed agent architecture learns from rewards across multiple environments and tasks, for fine-tuning a pre-trained language model which refines the language agent prompt by summarizing the root cause of prior failed attempts and proposing action plans. Experimental results on various tasks demonstrate that the language agents improve over time and that our approach considerably outperforms baselines that do not properly leverage gradients from the environment.", "keywords": "Language Agent;AI Agent;Reinforcement Learning", "qa_pairs": [{"question": "How does the proposed Retroformer model perform in comparison to GPT-4, and what are the implications of these results for demonstrating the contribution of the work?", "answer": "The comparison with the GPT-4 model is presented in Table 2, Figure 4, and Figure 6 of the paper. The results show consistent improvements at episode 0 when verbal feedback is not applied, attributed to GPT-4's better decision-making capabilities. Additionally, Retroformer demonstrates consistent improvements under GPT-4, highlighting the effectiveness of verbal feedback with a well-trained actor (e.g., GPT-4) combined with a retrospective component tuned using the policy gradient method. ", "category": "Method_Comparison"}, {"question": "What are the specific improvements observed in HotPotQA, WebShop, and against traditional RL methods when using Retroformer?", "answer": "**(1) Improvements in HotPotQA**  \nThere are significant improvements in HotpotQA by using Retroformer. For example, as in Table 2, Retroformer (GPT-3, lora r=4) achieved a 14% improvement in success rate within just one episode. The improvement is also 6% higher than the Reflexion (GPT-3) method that uses a frozen model without the paper's proposed RLHF finetuning. The authors have also tested GPT-4 as an agent in the updated experiments and found that even GPT-4 gets stuck at a 54% success rate. It is believed that 54% is the upper bound of the success rate for existing LLMs. It deserves notice that HotPotQA, which is a text-based game, has an extremely large state and action space (it can be treated as a RL problem with a continuous action of 768 dimensions). It should take more than 500 episodes to observe certain improvements in some much simpler text-based environments in [1]. The paper's improvements (14%) by Retroformer in just one episode are significant.\n**(2) Improvements in WebShop**  \nThe difficulty of solving this web shop environment is well known in all recent language agent publications [2,3,4] that use this environment. The performance improvement by Retroformer over the frozen baselines is observed, but the improvements may be limited. This limitation of improvement was due to the fact that 'web browsing requires a significant amount of exploration with more precise search queries'. Nevertheless, the paper's Retroformer method is still performing better than Reflexion and React across episodes, which use a frozen model without the paper's proposed RLHF finetuning. The authors have added GPT-4 agents in the updated experiments and found similar results in the updated Figure 6 (b). The results probably indicate that the verbal feedback approach (Reflexion, Retroformer) is not an optimal method for this environment, but the paper's fine-tuning method still proves effective.\n**(3) Improvements against RL methods**  \nTo show the improvements against traditional RL methods, the authors added one method, i.e., Soft Actor-Critic [5], as the baseline in Table 2, and the implementation details of the RL method in Appendix C.2, which is quoted below. 'The paper includes one online RL algorithm, i.e., Soft Actor-Critic [5], or SAC as a baseline model for comparison. Given that the three environments are text-based games, inspired by [1], the authors do mean-pooling for the embeddings of the generated text outputs, such as 'Search[It Takes a Family]' as the agent actions. Therefore, the action space is continuous and is of 768 dimensions. The authors apply LoRA adapters with r=4 on the agent Action model instantiated from longchat-16k, and use SAC to do the online updates, with discount factor gamma=0.99, interpolation factor polyak=0.995, learning rate=0.01, entropy regularization alpha=0.2, and batch size=8.' For a fair comparison with the language agent methods in this paper, which all showed improvements in 5 episodes, the authors ran SAC for 5 episodes but did not observe significant improvement across all 3 environments in Table 2. This is mostly due to the fact that the three text-based games have an extremely large state and action space, and thus online exploration and learning within 5 online episodes cannot improve the LLM model with an extensive parameter count. Under this setting, verbal reinforcement with an LLM that is finetuned offline for generating effective reflective feedback (i.e., the paper's Retroformer) is much more effective.\n\n\n[1] Yuan, Xingdi, et al. \"Counting to explore and generalize in text-based games.\" arXiv preprint arXiv:1806.11525 (2018).\n[2] Yao, Shunyu, et al. \"React: Synergizing reasoning and acting in language models.\" arXiv preprint arXiv:2210.03629 (2022). \n[3] Shinn, Noah, Beck Labash, and Ashwin Gopinath. \"Reflexion: an autonomous agent with dynamic memory and self-reflection.\" arXiv preprint arXiv:2303.11366 (2023). \n[4] Liu, Xiao, et al. \"Agentbench: Evaluating llms as agents.\" arXiv preprint arXiv:2308.03688 (2023).\n[5] Haarnoja, Tuomas, et al. \"Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor.\" International conference on machine learning. PMLR, 2018.\n", "category": "Method_Comparison"}, {"question": "How does the proposed method compare against recognized RL methods like Soft Actor-Critic within the interaction-feedback-learning framework?", "answer": "The results in Table 2 show that SAC did not exhibit significant improvement across all three environments (HotPotQA, AlfWorld, and WebShop).\n\n| Method  | #Params | #Retries | HotPotQA | AlfWorld | WebShop |\n|---------|---------|----------|----------|----------|---------|\n| SAC     | 2.25M   | N=1      | 27       | 58.95    | 30      |\n|         |         | N=4      | 27       | 59.7     | 30      |\n", "category": "Method_Comparison"}, {"question": "How does Retroformer mitigate the risk of overfitting when fine-tuning on potentially repetitive data derived from sparse feedback in the agent-environment interaction?", "answer": "Retroformer addresses overfitting by collecting data offline for fine-tuning the Retrospective component, ensuring the ratings in the offline dataset are not sparse. The training set includes a balanced distribution of positive and negative ratings, with positive ratings accounting for around 40% and negative ratings for 60%. Detailed data collection methods are provided in Appendix C.1.For HotPotQA environment, the paper collected 3,383 reflection samples by running the base rollout policy for 3 trials (N = 3) for 3,000 tasks in the training set, in which 1,084 instruction-response pairs have positive ratings. For AlfWorld, the paper collected 523 reflection samples and for WebShop, the paper collected 267 reflection samples. Additionally, the offline training set's queries or task descriptions do not overlap with those used for reporting success rates, thus avoiding repetitive data and overfitting.", "category": "Experimental_Setup"}, {"question": "What are the reasons for the observed limited performance improvements in HotPotQA and WebShop?", "answer": "For HotPotQA, the complexity of the text-based game and its large state-action space limit the method's ability to solve the game within 5 episodes, but 14% success rate in one episode and 20% improvement within 5 episodes are significant. For WebShop, the need for extensive exploration and precise search queries affects performance, though Retroformer still outperforms Reflexion and React. The verbal feedback approach(Reflexion, Retroformer) is not optimal for WebShop, but fine-tuning remains effective.", "category": "Experimental_Analysis"}, {"question": "What evidence demonstrates that the finetuned reflection model (Retroformer) produces better prompts compared to the frozen language model (Reflexion), and how often does each model correctly identify an improved plan (e.g. have a human hand-label 100 prompts for Reflexion and 100 from the frozen LM and record their success rates.)?", "answer": "Manual labeling of 5 reflection responses in the HotPotQA dataset showed that the agent LM always succeeds in the next trial when human-labeled. For the frozen LM (Reflexion), 1 out of 5 tasks succeeded in the next trial, while Retroformer (fine-tuned by policy gradient) succeeded in 3 out of 5 tasks. For 100 tasks, as shown in Figure 4, under GPT-4 at episode 1, Reflexion had 48 successful tasks and Retroformer (r=4) had 51 successful tasks. This indicates that 8(48-40)  plans of the frozen LM and 11(51-40) plans of Retroformer identified improved plans, demonstrating the effectiveness of Retroformer.", "category": "Experimental_Analysis"}, {"question": "How is the 'f1 score' reward calculated and used in the HotPotQA environment?", "answer": "The 'f1 score' reward is used in the HotPotQA environment to compare the matching of a generated answer to a question against the ground truth answer. After removing stopwords from both answers, the number of common tokens is calculated. Precision is the number of common tokens divided by the number of tokens in the generated answer, and Recall is the number of common tokens divided by the number of tokens in the ground truth answer. The F1 score is then computed from these precision and recall values. This reward function was defined in the ReAct paper[1], which serves as the baseline model in this study.\n\n[1]Yao, Shunyu, et al. 'React: Synergizing reasoning and acting in language models.' arXiv preprint arXiv:2210.03629 (2022).", "category": "Method_Mechanics"}, {"question": "Is it typically a better strategy to finetune a small model or prompt a more capable cloud model, such as GPT-3 or GPT-4, for generating reflective responses in a given environment?", "answer": "The Reflexion (GPT-3) curve in Figure 4 and 6 represents the requested scenario of using GPT-3 as the reflection model. Additionally, the authors tested GPT-4 as the reflection model and found that Retroformer (GPT-3/4, r=4) outperformed it significantly in all environments. This demonstrates that fine-tuning a small model is more effective than prompting a more capable general-purpose model for generating reflective responses in a given environment.", "category": "Experimental_Exposition"}, {"question": "Can the paper provide the exact algorithm used for PPO finetuning, including any modifications made to adapt it for offline use with language models?", "answer": "The exact algorithm for PPO finetuning is provided in Appendix C.1 - Algorithm 1. It includes three steps: offline data collection, reward model learning, and policy gradient finetuning. The offline ratings data is used to train a reward model, which is then integrated into the PPO finetuning process.", "category": "Method_Mechanics"}, {"question": "Could tuning the agents from RLHF feedback have a more significant effect on the LLM compared to fixing the agents and tuning the prompts?", "answer": "The paper included an RL baseline, Soft Actor-Critic[1], which directly finetunes the LLM agent online with environment feedback. After running this method for 5 episodes, no improvement was observed across all 3 environments, indicating that tuning the agent prompts is more efficient than tuning the agent LLMs directly in the paper's problem setting.\n\n\n| Method | #Params | #Retries | HotPotQA | AlfWorld | WebShop |\n|--------|---------|----------|----------|----------|---------|\n| SAC    | 2.25M   | N=1      | 27       | 58.95    | 30      |\n|        |         | N=4      | 27       | 59.7     | 30      |\n\n\n[1] Haarnoja, Tuomas, et al. \"Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor.\" International conference on machine learning. PMLR, 2018.", "category": "Experimental_Exposition"}, {"question": "Does the inclusion of GPT-4 as a comparison agent demonstrate that prompt tuning is still beneficial for well-trained agents like GPT-4 because GPT-4 still requires exploration to eliminate incorrect answers and choose correct trajectories in subsequent rounds?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper provide detailed prompts used in multiple environments (HotPotQA, Alfworld, and Webshop) to support the claim of 'automatically tunes the language agent prompts'?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "Q1u25ahSuy", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression", "authors": ["Tim Dettmers", "Ruslan A. Svirschevski", "Vage Egiazarian", "Denis Kuznedelev", "Elias Frantar", "Saleh Ashkboos", "Alexander Borzunov", "Torsten Hoefler", "Dan Alistarh"], "abstract": "Recent advances in large language model (LLM) pretraining have led to high-quality LLMs with impressive abilities. By compressing such LLMs via quantization to 3-4 bits per parameter, they can fit into memory-limited devices such as laptops and mobile phones, enabling personalized use. Quantizing models to 3-4 bits per parameter can lead to moderate to high accuracy losses, especially for smaller models (1-10B parameters), which are suitable for edge deployment. To address this accuracy issue, we introduce the Sparse-Quantized Representation (SpQR), a new compressed format and quantization technique that enables for the first time \\emph{near-lossless} compression of LLMs across model scales while reaching similar compression levels to previous methods. SpQR works by identifying and isolating \\emph{outlier weights}, which cause particularly large quantization errors, and storing them in higher precision while compressing all other weights to 3-4 bits, and achieves relative accuracy losses of less than $1\\%$ in perplexity for highly-accurate LLaMA and Falcon LLMs. This makes it possible to run a 33B parameter LLM on a single 24 GB consumer GPU without performance degradation at 15\\% speedup, thus making powerful LLMs available to consumers without any downsides. SpQR comes with efficient algorithms for both encoding weights into its format, as well as decoding them efficiently at runtime. Specifically, we provide an efficient GPU inference algorithm for SpQR, which yields faster inference than 16-bit baselines at similar accuracy while enabling memory compression gains of more than 4x.", "keywords": "quantization;sparsity;large language models", "qa_pairs": [{"question": "Does SpQR's performance depend on specific hardware capabilities, and if so, how does it perform across different hardware types?", "answer": "SpQR has similar performance on all GPUs from the past 10 years, as the main bottleneck during inference is the memory bandwidth of the GPU rather than other factors like FLOPS performance. This means hardware capabilities are not important for GPUs. However, SpQR would be very slow on some hardware, such as TPUs, but since GPUs are widespread, the authors consider it fair to have an implementation optimized for GPUs but not TPUs.", "category": "Method_Disambiguation"}, {"question": "What is the overhead of memory accesses for activations during inference in the proposed method, and how does it vary with batch size?", "answer": "For a batch size of 1, the activations for Llama 2 7B are 4000x smaller than the weight matrix, resulting in only 0.025% overhead. For larger batch sizes, such as those common in deployments like ChatGPT, the overhead increases but remains at most about 7% for a 7B model. Therefore, the limitation is not significant.", "category": "Method_Disambiguation"}, {"question": "Does the unstructured sparse pattern of the outliers impose significant computational overhead on GPUs, and does this affect the practicality of the method?", "answer": "While there is an overhead due to the unstructured sparse pattern, the model still provides faster inference than an unquantized model. With the right implementation, a 7B model can generate about 115 tokens per second on a single RTX 4090 GPU, making it fast enough to be practical and not significantly disadvantageous.", "category": "Method_Disambiguation"}, {"question": "How does SpQR's methodology for outlier management and mixed-precision quantization differ from prior work?", "answer": "SpQR's methodology differs from prior work by focusing on large language models (LLMs), which exhibit very different outlier patterns compared to smaller models studied in earlier work. The analysis extends the findings of LLM.int8()[1] to more complex outlier structures, making it novel and distinct from earlier research.\n\n[1] Dettmers et al., LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale", "category": "Method_Comparison"}, {"question": "What aspects of the SpQR method are novel and how do they contribute to advancements in the field?", "answer": "The analysis and discovered structures in SpQR are novel and undocumented in the literature. The algorithm, which uses outlier-aware Hessian optimization for layer-by-layer quantization, is also novel. Additionally, SpQR is one of the first methods to quantize LLMs to 3-bits without significant performance degradation, representing major advances in analysis and algorithms while achieving state-of-the-art results.", "category": "Method_Disambiguation"}, {"question": "How do the efficiency improvements claimed in the paper vary across different model architectures and what is the impact of sparse representations on performance?", "answer": "The sparse representations vary between 0.35% to 1.0%, resulting in minimal changes to the fraction of outliers and consistent performance across architectures. The largest difference is observed in Falcon 180B, which requires about 1% outliers, while Llama v1/v2 require ~0.35-0.45%. These minimal differences indicate that the impact of sparse representations on performance is negligible. Additionally, the paper's implementation of SpQR is more effective than the one discussed in 'Sparse GPU Kernels for Deep Learning'[1].\n\n[1] Gale et al., Sparse GPU Kernels for Deep Learning", "category": "Method_Mechanics"}, {"question": "How is the 3-bit matrix multiplication executed on the A100 GPU architecture?", "answer": "3-bit integers are packed into 32-bit aligned float values. These values are loaded into registers, dequantized to 16-bit values, and then matrix multiplication is performed with 16-bit inputs and weights in registers. The sparse portion of the matrix multiplication is handled separately, and its results are added to the regular quantized matrix multiplication.", "category": "Experimental_Setup"}, {"question": "What was the primary focus of the optimization strategy in terms of memory access and computation, and under what conditions is activation compression beneficial?", "answer": "The primary focus was on minimizing the memory footprint of the weights rather than optimizing mean FLOPS utilization. This is particularly useful for single-user deployments, such as on a personal computer with a batch size of 1. In such cases, the majority of latency arises from loading weights from GPU memory into registers, which can be accelerated by compressing the weights. Activation compression is beneficial only when using a large batch size (>32) and 4-bit or 8-bit tensor cores.", "category": "Method_Disambiguation"}, {"question": "How does the performance of the quantized GEMM with 4-bit SpQR compare to the quantized GEMM with a normal 4-bit weight quantization method, particularly in terms of speed, memory savings, and model compression?", "answer": "The best 4-bit quantization methods are about 3.5x faster than 16-bit methods, while SpQR is 1.15x faster than 16-bit methods. This means 4-bit quantization without outliers is up to 3.04x faster than SpQR. However, SpQR offers better performance and stronger model compression, averaging close to 3.4 bits per parameter. At 3.4 bits, SpQR provides 1.3x better memory savings compared to 4-bit techniques and halves the perplexity gap between 16-bit and 4-bit methods. SpQR is close in speed to other methods, as achieving similar performance with regular quantization would require ~5 bits, increasing GPU memory usage by 50%. For example, SpQR allows running Falcon 180B on a single GPU, whereas regular quantization methods would require two GPUs for the same performance. The memory benefit of SpQR is significant, especially as models continue to scale.", "category": "Method_Comparison"}, {"question": "What is the difference in accuracy between SpQR and 4-bit group-wise methods like GPTQ[1], and how does this difference impact practical applications such as fitting models on hardware?\n\n[1] Frantar, Elias, et al. \"OPTQ: Accurate quantization for generative pre-trained transformers.\" The Eleventh International Conference on Learning Representations. 2022.", "answer": "GPTQ requires 0.5-0.6 more bits per parameter to match the performance of SpQR at 3.5 bits per parameter on average, and 0.3-0.4 more bits to match the lossless compression configuration. This difference is critical in practical applications; for example, SpQR allows fitting Llama-2-70b on a single V100 with 32Gb and space for KV cache, which is impossible for GPTQ at the same accuracy without weight offloading.\n\n| Model | Method | wbits | wikitext2 |  c4  |\n|:-----:|:------:|:-----:|:---------:|:----:|\n|   7   |  GPTQ  |  4.07 |  5.88   | 7.28 |\n|       |  SpQR  |  3.45 |  5.88   | 7.33 |\n|       |  GPTQ  |  5.07 |  5.72   | 7.12 |\n|       |  SpQR  |  4.63 |  5.73   | 7.13 |\n|   13  |  GPTQ  |  4.04 |  5.26   | 6.76 |\n|       |  SpQR  |  3.45 |  5.25   | 6.77 |\n|       |  GPTQ  |  5.04 |  5.13   | 6.64 |\n|       |  SpQR  |  4.63 |  5.13   | 6.64 |\n|   30  |  GPTQ  |  4.02 |  4.28   | 6.11 |\n|       |  SpQR  |  3.49 |  4.29   | 6.11 |\n|       |  GPTQ  |  5.01 |  4.14   | 6.01 |\n|       |  SpQR  |  4.69 |  4.14   | 6.01 |\n|   65  |  GPTQ  |  4.01 |  3.71   | 5.74 |\n|       |  SpQR  |  3.52 |  3.71   | 5.73 |\n|       |  GPTQ  |  5.01 |  3.58   | 5.65 |\n|       |  SpQR  |  4.71 |  3.57   | 5.64 |", "category": "Method_Comparison"}, {"question": "How is the memory efficiency of the model quantified and what is an example calculation for a 70B model?", "answer": "The memory efficiency is quantified by the average amount of bits per parameter. For a 70B model, with 3.4 bits per parameter, the calculation is 70B*(3.4 bits / 8 bits) = 70B*(0.425 bytes) = 29.75 GB.", "category": "Method_Mechanics"}, {"question": "What performance evaluations were conducted on trendy benchmarks for generative models, such as GSM8K, and HumanEval, and how do the results compare to other methods?", "answer": "Perplexity measures remain very important: Previous work [1] showed that perplexity evaluations have a very strong correlation (-0.94) with performance on a diverse set of zeroshot tasks. This means that the perplexity evaluations generalize to many different tasks.\n\nThe authors evaluate the paper's method with MetaMath-7B, a recent version of Llama-2 specialized on math problem-solving, and CodeLlama-7b, a popular code generation model. One can see that SpQR is close to fp16 model performance with 4 bit quantization while outperforming GPTQ by ~2%. Drops on HumanEval dataset are more significant, but SpQR is still much better than GPTQ at the same bitwidth. These results correlate well with perplexity measure differences. Pass@100 scores were not reported due to noise. RTN quantization turns out to be numerically unstable and produces models with random performance.\n\nTable 1. GSM-8k results\n\n| model      | method | wbits | Accuracy (%) |\n|:----------:|:------:|:-----:|:------------:|\n| MetaMath-7B|  None  |   16  |     66.9     |\n|            |   RTN  |   4   |      NaN     |\n|            |  GPTQ  |   4   |     64.7     |\n|            |  SpQR  |  3.98 |     66.5     |\n\nTable 2. HumanEval results\n\n| model       | method | wbits | pass@1 | pass@10 |\n|:-----------:|:------:|:-----:|:------:|:-------:|\n| CodeLLama-7b|  None  |   16  |  40.0  |   58.0  |\n|             |   RTN  |   4   |   NaN  |    NaN  |\n|             |  GPTQ  |   4   |  32.1  |   52.2  |\n|             |  SpQR  |  3.98 |  36.3  |   54.6  |", "category": "Method_Comparison"}, {"question": "How does the small group size in the two-stage quantization and the sparse representation of outliers impact the latency overhead in the paper's implementation?", "answer": "Small group sizes do not affect latency in the paper's implementation because every warp of threads on the GPU processes multiple values in the matrix multiplication, allowing multiple groups to be loaded with each warp. The total values loaded remain the same regardless of group size, so inference speed depends only on the average amount of bits in the representation. While the sparse representations do affect latency, the paper's implementation can still generate about 115 tokens per second from a 7B model on a single A100 GPU, making it practical despite the sparse representations.", "category": "Method_Mechanics"}]}
{"id": "mM7VurbA4r", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents", "authors": ["Xuhui Zhou", "Hao Zhu", "Leena Mathur", "Ruohong Zhang", "Haofei Yu", "Zhengyang Qi", "Louis-Philippe Morency", "Yonatan Bisk", "Daniel Fried", "Graham Neubig", "Maarten Sap"], "abstract": "*Humans are social beings*; we pursue social goals in our daily interactions, which is a crucial aspect of social intelligence. Yet, AI systems' abilities in this realm remain elusive. We present SOTOPIA, an open-ended environment to simulate complex social interactions between artificial agents and evaluate their social intelligence. In our environment, agents role-play and *interact* under a wide variety of scenarios; they coordinate, collaborate, exchange, and compete with each other to achieve complex social goals. We simulate the role-play interaction between LLM-based agents and humans within this task space and evaluate their performance with a holistic evaluation framework called SOTOPIA-Eval. With SOTOPIA, we find significant differences between these models in terms of their social intelligence, and we identify a subset of SOTOPIA scenarios, SOTOPIA-hard, that is generally challenging for all models. We find that on this subset, GPT-4 achieves a significantly lower goal completion rate than humans and struggles to exhibit social commonsense reasoning and strategic communication skills. These findings demonstrate SOTOPIA's promise as a general platform for research on evaluating and improving social intelligence in artificial agents.", "keywords": "Social;Interaction;Agent;Social intelligence;Large Language Models;Evaluation;Theory of Mind", "qa_pairs": [{"question": "How does the paper address the concern that using GPT-4 for both generating scenarios and evaluating them in SOTOPIA might introduce bias, given that the generated content could be present in GPT-4's training data?", "answer": "Whether GPT-4 would significantly prefer its own output during evaluation: This may be a factor, but it does not change the  paper's main conclusions as those are corroborated by human evaluations. In addition to GPT-4 based evaluation Table 2 and 3, the paper has human-annotated evaluation results in Table G.3 andG.4, and they show similar trends. Putting these together with results in Table 1, It can be concluded that GPT-4 based evaluation is a reasonable indication of models’ performance on SOTOPIA tasks. In addition, for future evaluations using SOTOPIA, it is also possible to reproduce the paper's human evaluation protocol, thus, the paper’s assumptions about the correlation between human and machine evaluation can be re-examined as more competitive non-GPT-4 models are released. In the paper’s pilot study, it was found that GPT-4 is the best proxy for human evaluation among all LLMs tested.\n\nAbout biases related to scenarios being more 'in-domain' for GPT-4 than other models: The authors not only use templates to steer GPT-4’s output, but also manually inspect and correct the generated scenarios and goals. As such, the language distribution of the paper's diverse scenarios significantly deviates from GPT-4's most probable language distribution [1]. Furthermore, the complexity of the task also depends on the partner agent during the interactions (Section 2.1). \n\nFor the point about the generated content possibly being present in the training data: The paper does not train any models. If the point about the generated scenarios might reflect GPT-4's training data: the paper's scenarios are novel generations and are not present in the public web corpora that one could browse (the authors searched for the paper's scenarios through n-gram matching with https://wimbd.apps.allenai.org/ in OpenWebText, C4, OSCAR, The Pile, LAION-2B-en and found 0% of them in there, while the authors don't know what GPT-4 is trained on, these datasets are representative of data used to train LLMs), so there is very little reason to believe that they are in the training data of GPT-4 or any other language model.\n\n[1] Kim, Hyunwoo, Jack Hessel, Liwei Jiang, Peter West, Ximing Lu, Youngjae Yu, Pei Zhou, et al. 2022. “SODA: Million-Scale Dialogue Distillation with Social Commonsense Contextualization.” EMNLP 2023", "category": "Method_Disambiguation"}, {"question": "How do the evaluation metrics 'believability' and 'social rules' specifically measure social intelligence in AI agents, given that they might appear to only assess dialogue fluency or rule memorization?", "answer": "The paper’s dimensions are specifically designed to evaluate social intelligence in AI agents: some of the paper’s metrics measure aspects specific to AI agents (e.g., believability), while others reflect human social intelligence (See section 3). The former, for example, led the authors to include ‘believability’ as a dimension needed specifically to measure AI social intelligence: it captures not only conversation naturalness but also the agent’s character consistency (i.e., whether they remain ‘in-character’ and consistent with their prompted persona); this shortcoming, common in many social AI systems [1], needs to be measured. This contrasts with humans, where persona consistency is less of a concern because most humans naturally have their own consistent persona. As for social rules, this dimension is not about just recalling rules, but about measuring whether AI agents can successfully follow or apply the relevant rules during the interactions, which often tends to be challenging in many social settings [2].\n\n[1] Shuster, Kurt, Jack Urbanek, Arthur Szlam, and Jason Weston. 2022. 'Am I Me or You? State-of-the-Art Dialogue Models Cannot Maintain an Identity.' In Findings of the Association for Computational Linguistics: NAACL 2022.\n\n[2] Jin, Zhijing, Sydney Levine, Fernando Gonzalez, Ojasv Kamal, Maarten Sap, Mrinmaya Sachan, Rada Mihalcea, Josh Tenenbaum, and Bernhard Schölkopf. 2023. 'When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment.' Neurips, 2023.", "category": "Method_Disambiguation"}, {"question": "What is the distribution of different types of social goal and relationship scenarios included in the evaluation framework?", "answer": "The evaluation framework includes 28 negotiation scenarios, 62 exchange scenarios, 7 competition scenarios, 17 collaboration scenarios, 12 accommodation scenarios, and 14 persuasion scenarios, with overlapping categories.", "category": "Method_Mechanics"}, {"question": "What is the rationale for not providing quantitative definitions for the metrics used in the evaluation, and how does the correlation between GPT-4 and human annotations compare when fine-grained descriptions are added?", "answer": "The authors provided the same rating instructions for GPT-4 and human annotators, citing previous work[1] that demonstrates LLMs' effectiveness in annotating subjective social tasks. While there are inherent challenges in rating social interactions, the paper's human eval shows that GPT-4 provides an evaluation for models that aligns with human annotations at the system level (Table G.4). They conducted experiments adding fine-grained descriptions of quantitative definitions for each range of the scale(e.g., *Relationship Deteriorates (-5 to -3): Scores from -5 to -3 indicate that the relationship is deteriorating. This range suggests a significant decline in the quality or strength of the relationship, with increasing conflicts, misunderstandings, or detachment*), but this did not significantly improve the correlation with human annotations and, in some cases, slightly worsened it. The results are shown in the provided table, with Pearson R values for different metrics before and after adding fine-grained descriptions.\n\n\n|                   | SEC  | KNO  | SOC  | BEL  | REL  | FIN  | GOAL |\n|-------------------|------|------|------|------|------|------|------|\n| Original Pearson R| 0.22 | 0.33 | 0.33 | 0.45 | 0.56 | 0.62 | 0.71 |\n| + Fine-grained Description Pearson R | 0.03 | 0.33 | 0.33 | 0.35 | 0.57 | 0.57 | 0.71 |\n\n[1] Can Large Language Models Transform Computational Social Science? Caleb Ziems, William Held, Omar Shaikh, Jiaao Chen, Zhehao Zhang and Diyi Yang Computational Linguistics, 2023", "category": "Experimental_Exposition"}, {"question": "How does SOTOPIA differ from existing social interactive benchmarks like SocialIQA, and what potential benefits does it offer for enhancing the social intelligence of AI agents?", "answer": " 1. The current benchmarks, such as SocialIQA, primarily consist of descriptive interactions paired with question-and-answer (QA) sets. In these cases, the model doesn't engage in real-time interactions. In contrast, Sotopia introduces a dynamic approach. It evaluates Large Language Models (LLMs) within a first-person interactive framework, allowing for direct engagement. This marks a significant shift from the static nature of existing benchmarks to a more dynamic and interactive evaluation environment in Sotopia. 2. With the rapid development of LLMs, some of the benchmarks gradually become saturated. While there are new adversarial benchmarks to evaluate social intelligence, which are harder than their predecessors, they still lack the dynamic nature of social interactions and the rich social context, which is deemed insufficient for evaluating social intelligence in AI systems [1]. 3. The paper's work is indeed inspired by several benchmarks and datasets in social intelligence, such as SocialIQA, Persona Chat, and others. For instance, many of the paper's scenarios come from SocialIQA, and the paper's character design is inspired by Persona Chat. 4. While the paper’s current focus is on evaluation rather than the improvement of agents’ social intelligence, the importance of the latter is acknowledged. The paper has ongoing work using SOTOPIA to enhance social intelligence in AI agents and there is likely a correlation between doing better at SocialIQA and being a better interaction agent.\n\n[1] Lee, Mina, Megha Srivastava, Amelia Hardy, John Thickstun, Esin Durmus, Ashwin Paranjape, Ines Gerard-Ursin, et al. 2023. “Evaluating Human-Language Model Interaction.” Transactions on Machine Learning Research. https://openreview.net/pdf?id=hjDYJUn9l1.", "category": "Method_Comparison"}, {"question": "How are the SOTOPIA metrics (GOAL, BEL, REL, KNO, SEC, SOC, & FIN) justified and connected to existing sociological, psychological, and economic literature?", "answer": "The SOTOPIA metrics are justified by drawing on multiple pieces of literature from sociological, psychological, and economic domains. Specifically, Goal Completion is based on Weber (1978)[1], Believability on Joseph (1994) and Park et al. (2023a)[2], Knowledge on Reiss (2004)[3] and Maslow (1943)[4], Secret on Reiss (2004)[5], Relationship on Maslow (1943)[4] and Benabou & Tirole (2006)[6], Social Rules on Maslow (1943)[4], and Financial and Material Benefits on Gilpin & Sandholm (2006)[7] and Burns et al. (2017)[8]. The metrics unify similar concepts from different domains with new definitions compatible with existing definitions and SOTOPIA scenarios.\n\n[1]Max Weber. The Nature of Social Action, pp. 7–32. Cambridge University Press, 1978. doi:\n10.1017/CBO9780511810831.005. URL https://classicalsociologicaltheory.files.\nwordpress.com/2016/06/max-weber-classical-sociological-theory.pdf.\n[2]Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and\nMichael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In In the\n36th Annual ACM Symposium on User Interface Software and Technology (UIST ’23), UIST ’23,\nNew York, NY, USA, 2023. Association for Computing Machinery.\n[3]Steven Reiss. Multifaceted nature of intrinsic motivation: The theory of 16 basic desires. Review of General Psychology, 8(3):179–193, 2004. doi: 10.1037/1089-2680.8.3.179. URL https:\n//doi.org/10.1037/1089-2680.8.3.179.\n[4]A H Maslow. A theory of human motivation. Psychol. Rev., 50(4):370–396, July 1943.\n[5]Steven Reiss. Multifaceted nature of intrinsic motivation: The theory of 16 basic desires. Review of General Psychology, 8(3):179–193, 2004. doi: 10.1037/1089-2680.8.3.179. URL https:\n//doi.org/10.1037/1089-2680.8.3.179.\n[6]Roland Benabou and Jean Tirole. Incentives and prosocial behavior. ´ American Economic Review, 96(5):1652–1678, December 2006. doi: 10.1257/aer.96.5.1652. URL https://www.aeaweb. org/articles?id=10.1257/aer.96.5.1652.\n[7]Andrew Gilpin and Tuomas Sandholm. A competitive texas hold’em poker player via automated abstraction and real-time equilibrium computation. In Proceedings of the 21st National Conference on Artificial Intelligence - Volume 2, AAAI’06, pp. 1007–1013. AAAI Press, 2006. ISBN 9781577352815.\n[8]Tom Burns, Ewa Roszkowska, Ugo Corte, and Nora Machado des Johansson. Sociological game theory: Agency, social structures and interaction processes. Optimum. Studia Ekonomiczne, pp. 187–199, 01 2017. doi: 10.15290/ose.2017.05.89.13.", "category": "Method_Mechanics"}, {"question": "How does the generalizability of the findings to other AI models, such as Llama2, or real-world social interactions compare to GPT-4, based on the correlation results with human evaluation?", "answer": "GPT-4 demonstrates the highest correlation with human evaluation among the tested models, as shown by the Pearson R values (e.g., 0.71 for GOAL). In contrast, Llama2 shows significantly lower correlation (e.g., 0.24 for GOAL), indicating its limitations as a proxy for human evaluation. This suggests that GPT-4 is more generalizable for such tasks compared to other models like Llama2.\n\n| Model       | SEC | KNO | SOC | BEL | REL | FIN | GOAL |\n|-------------|-----|-----|-----|-----|-----|-----|------|\n| GPT-4 Pearson R | 0.22| 0.33| 0.33| 0.45| 0.56| 0.62| 0.71 |\n| Llama2 Pearson R| nan | 0.05| nan | 0.08| 0.11| 0.13| 0.24 |\n", "category": "Experimental_Exposition"}, {"question": "What is the estimated cost of evaluating a given agent in SOTOPIA using GPT-4, including the number of API calls and the resulting budget?", "answer": "Evaluating one episode on average requires 2,000 input tokens and 500 output tokens. Based on the current GPT-4 pricing, this costs approximately $0.09 per episode. Therefore, evaluating a given agent would cost around $81 (0.09 * 900 episodes).", "category": "Method_Mechanics"}, {"question": "Did the study investigate whether humans could distinguish between interacting with an LLM or a human, and how potential negative AI bias might influence their behavior? Additionally, could the study reproduce the effect by explicitly informing participants that they are interacting with an AI or human, even if it is not always true?", "answer": "In the study, humans were randomly paired with either other humans or LLM agents, and they were informed that they could be interacting with either. Post-study interviews revealed that some annotators could not distinguish between human and AI interlocutors. Due to IRB limitations, the study could not deceive participants by always claiming they were interacting with an AI or human when it was not the case. ", "category": "Experimental_Setup"}, {"question": "Were the generated scenarios in SOTOPIA manually inspected and corrected to deviate from GPT-4's most probable language distribution?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the human-annotated evaluation results in Tables G.3 and G.4 show trends significantly different from those in the GPT-4 based evaluation Tables 2 and 3?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the evaluation of SOTOPIA scenarios using GPT-4 solely relied upon, without any human evaluation to corroborate the results?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is there evidence that the generated scenarios in SOTOPIA are present in the public web corpora commonly used for training large language models?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "xbjSwwrQOe", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Statistical Rejection Sampling Improves Preference Optimization", "authors": ["Tianqi Liu", "Yao Zhao", "Rishabh Joshi", "Misha Khalman", "Mohammad Saleh", "Peter J Liu", "Jialu Liu"], "abstract": "Improving the alignment of language models with human preferences remains an active research challenge. Previous approaches have primarily utilized online Reinforcement Learning from Human Feedback (RLHF). Recently, offline methods such as Sequence Likelihood Calibration (SLiC) and Direct Preference Optimization (DPO) have emerged as attractive alternatives, offering improvements in stability and scalability while maintaining competitive performance. SLiC refines its loss function using sequence pairs sampled from a supervised fine-tuned (SFT) policy, while DPO directly optimizes language models based on preference data, foregoing the need for a separate reward model. However, the maximum likelihood estimator (MLE) of the target optimal policy requires labeled preference pairs sampled from that policy. The absence of a reward model in DPO constrains its ability to sample preference pairs from the optimal policy. Meanwhile, SLiC can only sample preference pairs from the SFT policy. To address these limitations, we introduce a novel approach called Statistical Rejection Sampling Optimization (RSO) designed to source preference data from the target optimal policy using rejection sampling, enabling a more accurate estimation of the optimal policy. We also propose a unified framework that enhances the loss functions used in both SLiC and DPO from a preference modeling standpoint. Through extensive experiments across diverse tasks, we demonstrate that RSO consistently outperforms both SLiC and DPO as evaluated by both Large Language Models (LLMs) and human raters.", "keywords": "large language model;preference optimization;language model alignment", "qa_pairs": [{"question": "How does the proposed method address the computational challenges associated with directly computing the set $M = \\min\\{m | m\\pi_{sft}(y|x) \\geq \\pi_{r_{\\psi}}(y|x) \\text{ for all } y \\notin \\mathcal{Y}\\}$ and the inefficiency of rejection sampling?", "answer": "The proposed method addresses the computational challenges by estimating $\\frac{\\pi_{r_{\\psi}}(y|x)}{M\\pi_{\\text{sft}}(y|x)}$ using 64 sequences sampled through the SFT policy, rather than directly computing $M$. For rejection sampling, the method begins by efficiently sampling 64 sequences via the SFT policy, which is highly parallelizable, and then employs a sampling-without-replacement strategy in the rejection sampling algorithm. This strategy recalculates the maximum reward at the start of each round, ensuring that at least one optimal sequence is selected in each round. Specifically, the sequence whose reward is equivalent to the maximum reward is selected with a probability of one, thereby guaranteeing the selection of at least one optimal sequence in each round. This approach not only optimizes the selection process but also maintains the algorithm's effectiveness throughout its execution.", "category": "Method_Mechanics"}, {"question": "What is the purpose of Figure 3(b) and how do the results, including the overlapping confidence intervals, inform the selection of hyper-parameters and the performance of different loss functions?", "answer": "The primary objective of Figure 3(b) is to illustrate the process of hyper-parameter selection for the proxy reward model, with the win rate against SFT targets on the y-axis. The plot guides hyper-parameter choices and shows that hinge loss, though not statistically significantly better, ignores the SFT policy and trusts the proxy reward model's preference order, resulting in a high win rate. Despite overlapping confidence intervals, rso-sample-rank consistently outperforms sft-sample-rank across all loss functions. This is visually represented by the dotted lines (sft-sample-rank) remaining outside the shaded regions of all rso-sample-rank curves.The plot also challenges the notion that top-k-over-N is universally optimal. When beta equals zero, rso-sample-rank aligns with top-k-over-N. It is interesting to observe that a beta of 0.5 achieves a higher proxy reward model win rate than beta=0. Although this difference is not statistically significant, it indicates that a nuanced approach to beta selection could yield better results than the default top-k-over-N strategy. Additional baselines (RAFT[1] and ReST[2]) have been included to enhance the comparative framework and validate the findings.\n\n[1] Dong, Hanze, et al. \"Raft: Reward ranked finetuning for generative foundation model alignment.\" arXiv preprint arXiv:2304.06767 (2023).\n\n[2] Gulcehre, Caglar, et al. \"Reinforced self-training (rest) for language modeling.\" arXiv preprint arXiv:2308.08998 (2023).", "category": "Concept_Understanding"}, {"question": "How can the approach of aligning language models with human preferences be adapted to address issues of bias and fairness in language models?", "answer": "The paper's current research focuses on preference alignment in language models. In practical scenarios, reward scores are often multi-dimensional, and the aim of alignment is to attain a Pareto optimal frontier [1]. This allows for the introduction of additional objectives such as harmlessness, safety, and bias preference pairs. The method is adaptable, functioning with either weighted-averaged reward scores or through integration with multi-objective DPO loss functions[2]. Experimental studies have demonstrated that the RSO method effectively aligns with human preference pairs, and it has the potential to enhance fairness and reduce bias in language models when applied with appropriate human preference pairs. However, a comprehensive study of fairness and bias falls beyond the scope of this work. Additional discussions on this topic have been added in Appendix A.9.\n\n[1] Bai, Yuntao, et al. \"Constitutional ai: Harmlessness from ai feedback.\" arXiv preprint arXiv:2212.08073 (2022).\n\n[2] Zhou, Zhanhui, et al. \"Beyond One-Preference-for-All: Multi-Objective Direct Preference Optimization.\" arXiv preprint arXiv:2310.03708 (2023).\n", "category": "Method_Mechanics"}, {"question": "How does the computational efficiency of the proposed method compare to other methods, such as DPO, SLiC-HF-sample-rank, and RLHF?", "answer": "In terms of computational efficiency, the paper's approach is closer to SLiC-HF-sample-rank than DPO. The comparisons with other algorithms are as follows: Compared to DPO and SLiC-HF-direct, the paepr's approach requires an additional trained pairwise reward model, and also sample and rank stages. The pairwise reward model training is a standard text-to-text model without Bradley-Terry model assumption. The sample and rank are scalable and efficient with an increased number of servers, especially with shared prompts in SFT sample and short decoding length with reward model. Compared to SLiC-HF-sample-rank, the paper needs to sample more sequences from SFT policy as candidates, followed by a rejection sampling algorithm. Both steps are scalable and efficient. Compared to RLHF, RSO is offline with efficient memory usage (only one copy of policy network instead of four copies in PPO: policy, reference, value, reward), and the sample-rank stage is fully parallelizable (no need to conduct the online sample, which is not parallelizable).", "category": "Method_Comparison"}, {"question": "What are the potential reasons for the significant improvement in rso-sample-rank compared to sft-sample-rank on the Reddit TL;DR dataset, while the improvement is less noticeable on the AnthropicHH dataset?", "answer": "1. **Dataset Characteristics and Their Impact**:\n\n  **Reddit TL;DR Dataset**: This dataset uniquely combines SFT and human preference data sources. The human preference responses, generated through a mix of policies from OpenAI, including best-of-N rejection sampling, often yield higher quality outputs than SFT targets. Existing literature provides evidence to this effect. For example, in the right panel of Figure 2 in the DPO paper [1], it shows that Prefered-FT (SFT on the better responses from human preference data) is generally better than SFT. This divergence in quality potentially enlarges the gap between the SFT policy and the optimal policy, thereby making the improvements in rso-sample-rank more pronounced.\n\n  **AnthropicHH Dataset**: Contrarily, in the AnthropicHH dataset, both the SFT targets and reward model are derived exclusively from human preference data. The proximity in quality between the SFT policy and the reward model narrows the gap with the optimal policy, resulting in less noticeable improvements in rso-sample-rank.\n\n2. **Evaluation Methodology**:\n    The paper's human evaluation methodology, which deviates from the standard side-by-side comparison, might contribute to these observations. The paper asked raters to choose the best response among direct, sft-sample-rank, and rso-sample-rank options. While this approach has its merits, it may not fully capture the nuanced differences between sft-sample-rank and rso-sample-rank. \n\n**References**\n\n[1] Rafailov, Rafael, et al. \"Direct preference optimization: Your language model is secretly a reward model.\" arXiv preprint arXiv:2305.18290 (2023).", "category": "Experimental_Analysis"}, {"question": "What is the value of 'num_samples' in the experiments, and how many additional samples need to be generated compared to other methods like DPO, sft-sample-rank, and RAFT/ReST?", "answer": "In the experiments, 64 response candidates are sampled from the SFT policy, and 8 responses are selected using statistical rejection sampling. Compared to DPO, an extra 64 decodes are needed. Compared to sft-sample-rank, an additional 56 decodes are required (64-8=56). Compared to RAFT/ReST, no extra decodes are needed because they also use all 64 candidates to compute best-of-N and select those with high rewards.\nIn practice, batch decoding is efficient, scalable, and cost effective [1][2][3]. To further verify this, the authors conduct the following speed benchmarking on Llama2 7b q5-k-m with a single NVIDIA GeForce RTX 4090 GPU card with llama.cpp [4]:\n\n| Input length | Decode length | Number of decodes | Number of decoding tokens per second | Memory overhead | Model memory |\n| ------------ | ------------- | ----------------- | ------------------------------------ | --------------- | ------------ |\n| 1024         | 128           | 64                | 1135                                 | 5.7G            | 4.5G         |\n| 1024         | 128           | 16                | 596                                  | 1.9G            | 4.5G         |\n| 1024         | 128           | 8                 | 507                                  | 1.2G            | 4.5G         |\n| 1024         | 128           | 1                 | 120                                  | 0.7G            | 4.5G         |\n\nIn addition, the inference is scalable to the number of accelerators, because the parallelism can be applied across the whole inference dataset. One thing to notice is that for training data of size $N$, the inference will be done on those $N$ prompts and is much faster than training.\nSft-sample-rank can also use 64 response candidates, even with tournament ranking. In Table 2, the paper shows that rso-sample-rank is more effective than the sft-sample-rank setting with all candidates. This further demonstrates the effectiveness of the paper's approach even with the fixed computation cost.\n\n**References**\n\n[1] https://docs.mystic.ai/docs/mistral-ai-7b-vllm-fast-inference-guide\n\n[2] https://github.com/ggerganov/llama.cpp/issues/3479\n\n[3] Pope, Reiner, et al. \"Efficiently scaling transformer inference.\" Proceedings of Machine Learning and Systems 5 (2023).\n\n[4] https://github.com/ggerganov/llama.cpp/tree/master/examples/batched", "category": "Method_Comparison"}, {"question": "What are the key contributions and improvements of the proposed rejection sampling method in the context of offline human preference alignment algorithms?", "answer": "The key contributions and improvements of the proposed rejection sampling method include: (1) Bridging the gap between offline off-policy algorithms and approximately on-policy algorithms, addressing the distribution-shift issue that current approaches like DPO and SLiC-HF have not effectively tackled.The paper's work is pioneering in conducting a systematic study to approximate the optimal policy among offline human preference alignment algorithms. \n\n(2) Demonstrating that existing best-of-N or top-k-over-N rejection sampling methods are specific instances of the proposed comprehensive statistical rejection sampling algorithm, particularly when the beta parameter is set to zero. As beta approaches infinity, the method aligns with sampling from an SFT policy, balancing reliance on the SFT policy with the accuracy of the reward model. Unlike existing methods, the paper provides a nuanced approach that adapts dynamically to varying degrees of confidence in either the SFT policy or the reward model, thereby offering a more robust and adaptable solution for practical applications.\n\n(3) Enhancing efficiency by estimating $\\frac{\\pi_{r_{\\psi}}(y|x)}{M\\pi_{\\text{sft}}(y|x)}$ using 64 sequences sampled through the SFT policy, rather than computing $M$ directly. \n\n(4) Improving rejection sampling by sampling from the proposal distribution without replacement, ensuring efficiency in high-dimensional settings and guaranteeing the selection of at least one candidate with the highest reward. This modification guarantees efficiency, particularly when the target is a limited number of selected responses.\n\n(5) Grounding the algorithm in robust theoretical principles, simplifying the calculation of the threshold for a uniform random sample to exp((reward-max_reward) / beta). Overall, the method is a novel and effective solution for offline human preference alignment algorithms, offering both algorithmic efficiency and significant research advancements.", "category": "Method_Disambiguation"}, {"question": "What is the purpose of integrating $\\pi_{sft}(y|x)$ into Eq 10, and how does it contribute to the theoretical framework of SLiC-HF?", "answer": "1.The integration of $\\pi_{sft}(y|x)$ into Eq 10 aims to establish a more robust theoretical framework for the SLiC-HF approach, going beyond merely unifying DPO and SLiC. While it's true that SLiC-HF and DPO emerged around the same time, leading to some confusion about their connection, their starting points are fundamentally different. * SLiC-HF is grounded in the concept of likelihood calibration, where a better response is expected to have a higher likelihood than a less desirable response, by a defined margin. * On the other hand, DPO builds on the Reinforcement Learning from Human Feedback (RLHF) objective and the Bradley-Terry model, leading to the derivation of the sigmoid loss function as a Maximum Likelihood Estimation (MLE). \n\n2. The paper's contribution, as outlined in the paper, is to theoretically link these two approaches. By substituting the Bradley-Terry model with a Support Vector Machine (SVM) model, the paper obtains the hinge-norm of SLiC, effectively a normalized version. The original SLiC calibration loss overlooks the SFT policy normalization, relying solely on preference pairs labeled by the reward model. The hinge-norm loss, conversely, uses the SFT likelihood as a baseline to compute the relative likelihood ratio between the current policy and the SFT policy. This approach not only calibrates the likelihood ratio towards human preferences but also acts as an adaptive margin version of the SLiC loss. The margin, in this case, is influenced by the SFT likelihoods of the better and worse responses. Furthermore, this paper [1] highlights issues with the Bradley-Terry model and proposes the IPO loss, which aligns more closely with the hinge-norm loss than the original SLiC. \n\n3. In terms of implementation, DPO and SLiC variants share similarities by the choice of link functions and normalization. There isn't conclusive evidence yet to prefer one over the other. The paper provides a comprehensive framework for choosing between these options, underscoring the need for further research in this area. In conclusion, while the connection between DPO and SLiC might seem deliberate, especially in the context of Eq 10, it is part of a broader effort to establish a more theoretically sound approach to SLiC-HF, taking into account the nuances of different models and the importance of the SFT policy.\n\n[1] Azar, Mohammad Gheshlaghi, et al. \"A general theoretical paradigm to understand learning from human preferences.\" arXiv preprint arXiv:2310.12036 (2023).", "category": "Motivation_Analysis"}, {"question": "Why was ALMoST(https://arxiv.org/abs/2305.13735)  not included in the baseline comparisons, and how do RAFT(https://arxiv.org/abs/2304.06767) and ReST(http://arxiv.org/abs/2308.08998) align with RSO's methodology?", "answer": "ALMoST was not included in the baseline comparisons because its reliance on synthetic preference pairs diverges significantly from RSO's methodology, which could lead to an unfair comparison. In contrast, RAFT and ReST share more similarities with RSO in terms of objectives and methodologies. For RAFT, the SFT model is fitted using cross-entropy loss on the best response selected by tournament ranking using a pairwise reward model. For ReST, the parameters $\\tau=0.7, I=1, G=1$ were fixed as suggested by the ReST paper, the reward scores are normalized between 0 and 1, and responses with reward scores greater than 0.7 were chosen as new SFT targets. Multiple rounds of grow and improve stages were not included to ensure a fair comparison. RSO can also be done in multiple rounds to generate better rso-sample-rank preference pairs, and the results demonstrate that RSO outperforms these baselines across various metrics.", "category": "Method_Comparison"}, {"question": "Why is the policy induced from RSO in Eq 4 not considered strictly optimal?", "answer": "The policy induced from RSO in Eq 4 is not strictly optimal because it relies on a proxy reward model that labels responses without distinguish whether they're from the pre-trained or learned policy, and the reward model is learned from $D_p$ rather than $D_p^*$, introducing approximation errors and bias. ", "category": "Concept_Understanding"}, {"question": "What are the methods for alleviating the issue of reward hacking, which can arise due to errors within the proxy reward model?", "answer": "To mitigate the issue, the authors can collect preference data from the current best policy iteratively. In Llama2 [1], they collect human preference data from the current best policy on a weekly basis, as the policy keeps improving. The paper's work can follow the same approach as the paper improves the policy iteratively using rso-sample-rank.\n\n[1] Touvron, Hugo, et al. \"Llama 2: Open foundation and fine-tuned chat models.\" arXiv preprint arXiv:2307.09288 (2023).", "category": "Method_Mechanics"}, {"question": "Why does the pairwise reward model in the paper have significant advantages over the implicit pointwise reward model induced by DPO (Equation (5))?", "answer": " 1. **Pointwise vs Pairwise**: The paper's pairwise comparison approach, structured as “[Context] {context} [Response A] {response a} [Response B] {response b}”, aligns more naturally with human cognitive processes, compared with assigning a real-valued reward score for a response given a context. The pairwise approach not only simplifies the task but also leverages the extensive knowledge transfer from pre-trained language models in classification and comparison tasks. Unlike the Bradley-Terry model, which imposes a rank-1 approximation, the paper’s model directly represents the complexity of pairwise interactions without imposing such a simplifying assumption. 2. **Implicit vs Explicit**: The paper's model is an explicit text-to-text classification model, a task that is straightforward and less prone to the complexities associated with language generation. In contrast, the implicit reward model in DPO, based on likelihood scoring or generation, navigates a more intricate language output space. This complexity can hinder the effective transfer of knowledge from pre-trained models. The paper's approach, grounded in the principle that classification tasks are inherently more manageable than generation tasks, offers a more robust and reliable framework. The paper’s comprehensive experimental studies demonstrate the superiority of this model. For instance, rso-sample-rank shows to be significantly better than DPO. These results highlight not just the theoretical but also the practical advantages of the paper's approach.", "category": "Method_Disambiguation"}, {"question": "Has additional discussion on addressing bias and fairness in language models been included in the pape?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is a comprehensive study on fairness and bias within the scope of the current research as presented in the paper?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "2lDQLiH1W4", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Instant3D: Fast Text-to-3D with Sparse-view Generation and Large Reconstruction Model", "authors": ["Jiahao Li", "Hao Tan", "Kai Zhang", "Zexiang Xu", "Fujun Luan", "Yinghao Xu", "Yicong Hong", "Kalyan Sunkavalli", "Greg Shakhnarovich", "Sai Bi"], "abstract": "Text-to-3D with diffusion models has achieved remarkable progress in recent years.  However, existing methods either rely on score distillation-based optimization which suffer from slow inference, low diversity and Janus problems, or are feed-forward methods that generate low-quality results due to the scarcity of 3D training data. In this paper, we propose Instant3D, a novel method that generates high-quality  and diverse 3D assets from  text prompts in a feed-forward manner. We adopt a two-stage paradigm, which first generates a sparse set of four structured and consistent views from text in one shot with a fine-tuned 2D text-to-image diffusion model, and then directly regresses the NeRF from the generated images with a novel transformer-based sparse-view reconstructor. Through extensive experiments, we demonstrate that our method can generate diverse 3D assets of high visual quality within 20 seconds, which is two orders of magnitude faster than previous optimization-based methods that can take 1 to 10 hours. Our project webpage is: https://jiahao.ai/instant3d/.", "keywords": "text-to-3d;generative models;diffusion models;3D reconstruction;3D generation;sparse-view reconstruction", "qa_pairs": [{"question": "Are the training shapes canonicalized when training the view-conditioned image-to-triplane decoder, and how does transforming the camera poses affect the generated shapes? Consider a set of views V = [v1, v2, v3, v4] is input and a shape A is generated. Then, all the views are multiplied with a transformation matrix M, and a shape B is generated. Will shapes A and B be in the same canonicalized coordinate frame?", "answer": "The training shapes in Objaverse are not canonicalized, except for scaling all objects to a cube within [-1, 1]. The input to the transformer includes both images and their camera poses in the world coordinate system. If the images remain unchanged and only the camera poses are transformed, the output object will be correspondingly transformed. For example, rotating the camera poses by 90 degrees and use the rotated poses as input to reconstruct the shape from the original 4-view images results in a correspondingly rotated reconstructed shape.", "category": "Method_Mechanics"}, {"question": "How does the fine-tuning on the Objaverse dataset affect the model's ability to handle arbitrary text prompts, and what are the observed failure cases related to the domain gap?", "answer": "The light-weight fine-tuning on the Objaverse dataset for 10K steps allows the model to preserve the SDXL model's capability to handle a wide range of prompts, including complex compositional concepts. The second-stage sparse-view reconstruction model, trained on renderings from the Objaverse dataset, demonstrates robustness and generalization. However, the model fails to handle over-complicated prompts, such as those involving complex spatial arrangements of multiple subjects and complex scenes. Additionally, the generated assets are not as photorealistic as the 2D images generated by SDXL, which may be due to information loss during fine-tuning.", "category": "Experimental_Setup"}, {"question": "Which technical component in the proposed pipeline contributes to the improved diversity of generated images compared to prior methods like DreamFusion and ProlificDreamer?", "answer": "The improved diversity is attributed to the text-to-multiview generation network, fine-tuned from the 2D text-to-image diffusion model SDXL, which inherits the capability of generating diverse images from random samples. These diverse 2D images are then lifted to 3D using a sparse-view reconstructor. In contrast, previous methods like DreamFusion and ProlificDreamer use Score Distillation-based optimization, which tends to produce similar results even with different initializations, as discussed in Section A.5 of the DreamFusion paper [1].Comparisons on diversity with DreamFusion and ProlificDreamer, which generate 4 different shapes for each prompt using different seeds, show that the paper’s method generates more diverse results than the baseline methods. For example, for the prompt 'a chimpanzee dressed like a football player', the paper's method generates diverse players while DreamFusion and ProlificDreamer failed to do so.\n\n[1] Dreamfusion: Text-to-3D using 2D Diffusion, Poole et. al, https://arxiv.org/abs/2209.14988 (2022)", "category": "Method_Comparison"}, {"question": "What is the significance of DINO features compared to features from other pre-trained networks, and how does the choice of features impact performance?", "answer": "The authors provide insights from three perspectives:\n\n(a) Pre-trained vs from-scratch: The paper has trained the paper’s model without initializing the visual encoder from a pre-trained weight (i.e., trained it from scratch). It is found that a pre-trained model gives faster convergence in the beginning, but the gap keeps decreasing as the training budget increases. The complete elimination of the gap has not been observed under the limited compute budget, but it is guessed that the gap can possibly disappear with more data and longer training time.\n\n(b) Different pre-trained methods: For the same ViT architecture, there are multiple pre-trained methods, e.g., MAE, DINO, CLIP, and ImageNet-CLS. Among them, the paper has tested with CLIP-init visual encoder but the results are worse than DINO. It is guessed that this is because CLIP is only trained for its [CLS] token with img-text contrastive loss thus it is hard to preserve all raw spatial information. For the other two (i.e., MAE and IN-CLS), the paper has not given a test. But it is guessed that MAE might perform similar to DINO, and IN-CLS would perform similar to CLIP based on how their pre-training method is built: MAE supervises spatial tokens while IN-CLS only supervises the [CLS] token. DINO supervises [CLS] token as well but its spatial token is good according to its pre-training strategy and also shown in previous papers [1].\n\n(c) Different model architectures (e.g., ViT vs CNN): The paper has not experimented with CNN-based visual encoder, but it is thought that CNN should give similar results. The only consideration in the paper’s mind is that CNN usually needs a different training schedule than transformer thus makes the model tuning harder. Given this, the paper picks ViT in its exploration since easy tuning is crucial for large-scale training.\n\n[1] Deep vit features as dense visual descriptors, Amir, S., Gandelsman, Y., Bagon, S., & Dekel, T. (2021).", "category": "Method_Disambiguation"}, {"question": "What are the advantages of using tile-based generation for multi-view representations compared to generating views as separate channels, as in MVDream?", "answer": "The main advantage of tile-based generation is that it does not require changes to the architecture of the original diffusion model. Unlike multi-channel methods, which rely on specific attention modules and require additional parameters like view conditioning, tile-based generation can be applied to a wide range of network architectures without modifications or extra inputs. This approach offers several benefits: (a) It is more transferable across different generative methods and model architectures without specific adaptation. (b) Light-weight fine-tuning can be easily applied to new models. (c) Acceleration techniques developed for the base model can be directly applied, enabling faster deployment. (d) On-device or platform-specific code(e.g., cpp code; swift code) can be easily integrated, expanding potential applications. While multi-channel generation is an interesting future direction, tile-based generation provides significant practical advantages.", "category": "Method_Comparison"}, {"question": "Does the improvement in texture quality in the results shown in Figures 4 and 5 primarily result from the curated training dataset rather than enhancements in the texture generation method itself, such as improving the rendering method or adding extra photo-realistic losses?", "answer": "The improvement in texture quality is primarily attributed to the paper's method's use of normal feed-forward diffusion inference (DDIM), rather than the curated training dataset. The paper's model fine-tuned on non-curated data from Objaverse does not exhibit over-saturation problems, unlike optimization-based methods such as DreamFusion and ProlificDreamer, which use score distillation losses.", "category": "Method_Disambiguation"}]}
{"id": "UnUwSIgK5W", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "WizardCoder: Empowering Code Large Language Models with Evol-Instruct", "authors": ["Ziyang Luo", "Can Xu", "Pu Zhao", "Qingfeng Sun", "Xiubo Geng", "Wenxiang Hu", "Chongyang Tao", "Jing Ma", "Qingwei Lin", "Daxin Jiang"], "abstract": "Code Large Language Models (Code LLMs), such as StarCoder, have demonstrated remarkable performance in various code-related tasks. However, different from their counterparts in the general language modeling field, the technique of instruction fine-tuning remains relatively under-researched in this domain. In this paper, we present Code Evol-Instruct, a novel approach that adapts the Evol-Instruct method to the realm of code, enhancing Code LLMs to create novel models, WizardCoder. Through comprehensive experiments on five prominent code generation benchmarks, namely HumanEval, HumanEval+, MBPP, DS-1000, and MultiPL-E, our models showcase outstanding performance. They consistently outperform all other open-source Code LLMs by a significant margin. Remarkably, WizardCoder 15B even surpasses the well-known closed-source LLMs, including Anthropic's Claude and Google's Bard, on the HumanEval and HumanEval+ benchmarks. Additionally, WizardCoder 34B not only achieves a HumanEval score comparable to GPT3.5 (ChatGPT) but also surpasses it on the HumanEval+ benchmark. Furthermore, our preliminary exploration highlights the pivotal role of instruction complexity in achieving exceptional coding performance.", "keywords": "Large Language Models;Code;Instruction Fine-tuning", "qa_pairs": [{"question": "Why do the results from evol round 0 + 1 + 2 + 3 + 4 in Figure 3 show worse performance than round 0 + 1 + 2 + 3, and is there a point at which returns from Evol-Instruct diminish or turn negative?", "answer": "The empirical results indicate a potential saturation point after 4 rounds of Code Evol-Instruct. The decline in performance may be due to two factors: (1) continual evolution could lead to overly complicated and unreasonable instructions, and (2) beyond a certain round, instructions may become excessively complex, challenging the evolution executing model (i.e., gpt3.5-turbo) to generate suitable responses.", "category": "Experimental_Analysis"}, {"question": "What do the results in Table 4 demonstrate, and how do they confirm that the performance gains are not due to an increase in samples or tokens? ", "answer": "In Table 4, all models are fine-tuned separately from scratch, not sequentially. The models are fine-tuned with only the specific round data, which contains a similar number of samples (upper part) or tokens (lower part). The results show that with a similar number of samples or tokens, the models trained with evolved data achieve better pass@1.", "category": "Concept_Understanding"}, {"question": "What is the impact of using different evolving models, such as GPT-4 or open-source alternatives, on the performance of Evol-Instruct?", "answer": "The authors replaced GPT3.5 (gpt3.5-turbo) with GPT-4 for generating evolved rounds. HumanEval Pass@1 scores increased from 73.2 to 73.8 for 34B and from 59.8 to 62.2 for 15B. Furthermore, the authors explored using the open-source models (OSS) CodeLlama-Instruct-34B for generating evolved instructions. However, it demonstrated relatively low coding performance in response generation. To address this, the authors fine-tuned it using the code-alpaca dataset and utilized this model for response generation. The OSS-based Code Evol-Instruct tuned models also exhibited performance improvements compared to the code-alpaca fine-tuned model, increasing from 65.9 to 70.1 for 34B and from 45.7 to 55.5 for 15B. While GPT-4 demonstrates superior coding performance with a score of 88.4 compared to GPT3.5 (gpt3.5-turbo) with a score of 73.2, leveraging a more powerful model for generating evolved rounds does not result in a proportional performance gain (73.8 vs. 73.2). Conversely, despite CodeLlama exhibiting weaker performance than GPT3.5 (gpt3.5-turbo), the performance gap significantly diminishes when utilizing the paper's Code-Evol Instruct to generate evolved rounds (73.2 vs. 70.1). This underscores the observation that employing a stronger model for evolved rounds can somewhat enhance performance, but the extent of improvement is not strongly correlated to the evolution executing model's coding performance. The paper's Code Evol-Instruct emerges as the pivotal factor in driving performance enhancement.", "category": "Experimental_Exposition"}, {"question": "What is the difference between the two methods 'Add new constraints and requirements' and 'Replace a commonly used requirement' in the context of optimising evolutionary instructions in Section 3.1?", "answer": "The first method, 'Add new constraints and requirements,' involves including additional requirements in the instruction. For example, the original instruction `Write a MongoDB query to select all documents in a collection where the field 'category' is 'clothes' and the 'brand' field is not equal to 'Nike'` is expanded to include more constraints: `Write a MongoDB query to select all documents in a collection where the field 'category' is 'clothes' and the 'brand' field is not equal to 'Nike', and the 'price' field is greater than or equal to 100 and less than or equal to 500.` The second method, 'Replace a commonly used requirement,' involves substituting existing requirements in the instruction. For example, the original instruction is modified to replace the requirements: `Write a MongoDB query to select all documents in a collection where the field 'material' is 'cashmere' and the 'designer' field is not equal to 'Chanel'.`", "category": "Method_Disambiguation"}, {"question": "What is the purpose of the 4th instruction in the Code Evolution Heuristic Methods table, which involves providing a piece of erroneous code as a reference to increase misdirection, and why is this approach effective?", "answer": "The 4th instruction involves including an erroneous code snippet as a reference to challenge the code LLM. or example, the original instruction is: `Write a Python function to tell me what the date is today.` The evolved instruction is: `Write a Python function to tell me what the date is today, using the datetime module. The function should return a string in the format of 'Today is YYYY-MM-DD'. For example, if today is November 17, 2023, the function should return 'Today is 2023-11-17'. As a reference, here is a piece of erroneous code that does not work as intended:` ```python import datetime def get_date(): today = datetime.date.today() return 'Today is ' + today ``` This erroneous code serves as an adversarial sample designed to challenge the code LLM, akin to previous works focused on attacking pre-trained code models.", "category": "Concept_Understanding"}, {"question": "How does the claim about the pivotal role of instruction complexity in enhancing coding performance hold given the analysis in Table 4, where later rounds (assumed to have more complexity) lead to lower performance or statistically insignificant changes?", "answer": "The claim holds because models trained with any round of round 1-4 data (incorporating evolved complexity) exhibit significantly better performance than round 0 data (devoid of evolved complexity). Although rounds 1-4 are not monotonically increasing in complexity, they all outperform the low-complexity code-alpaca data. Figure 3 further supports this by showing that the combination of rounds 0-3 achieves the best performance, indicating the beneficial impact of introducing instruction complexity.", "category": "Experimental_Analysis"}, {"question": "What was the rationale behind the development of the five heuristic methods for increasing the complexity of code instructions, and how do they align with the characteristics of the code domain and prior research?", "answer": "The five heuristic methods were developed to increase the complexity of code instructions, aligning with the characteristics of the code domain and prior research on code models' limitations. The first three heuristics naturally escalate coding task difficulty by incorporating additional requirements, replacing common requirements, or introducing extra reasoning steps. This aligns with observed trends on platforms like LeetCode, where tasks with more requirements or intricate reasoning are considered more challenging. The fourth heuristic introduces erroneous code as an adversarial sample, inspired by prior research in attacking pre-trained code models [1][2], serving to assess the robustness of the code LLM. The fifth heuristic, emphasizing time and space complexity, aligns with the significance of optimization in the coding domain. Previous studies [3] indicate that tasks requiring optimized program performance pose heightened difficulty, making this direction a valuable avenue for complexity improvement.\n\n[1] Yang, Zhou et al. “Natural Attack for Pre-trained Models of Code.” 2022 IEEE/ACM 44th International Conference on Software Engineering (ICSE).\n\n[2] Jha, Akshita et al. “CodeAttack: Code-based Adversarial Attacks for Pre-Trained Programming Language Models.” AAAI 2022.\n\n[3] Madaan, A. et al. \"Learning Performance-Improving Code Edits.\" (arxiv, 2023)", "category": "Motivation_Analysis"}, {"question": "Why does the Pass@1 score in Table 4 decrease as the number of iterations increases?", "answer": "The Pass@1 scores in the upper part of Table 4 show fluctuations rather than a consistent decrease, with a notable decline in the fourth round. This decline is due to two factors: 1) continual evolution leading to overly complicated and unreasonable instructions, and 2) excessive complexity of instructions beyond a certain round, challenging the evolution executing model (gpt3.5-turbo) in generating suitable responses. In the lower part of Table 4, the decreasing scores are normal as more complex instructions increase sample length, reducing the number of samples per round to maintain the same token count. Individual metrics for specific rounds may not fully represent the overall impact on final model performance, as merging different complexity levels can improve performance, as shown in Figure 3.", "category": "Experimental_Analysis"}, {"question": "How does the performance of the model change when less capable models are used for generating evolving data, and can the same model be used to generate data for the next iteration?", "answer": "Using open-source models like CodeLlama-Instruct-34B for generating evolved instructions initially showed low performance, but fine-tuning with the code-alpaca dataset improved performance from 65.9 to 70.1 for 34B and from 45.7 to 55.5 for 15B. Despite CodeLlama's weaker performance compared to GPT3.5, the gap reduces with the paper's Code-Evol Instruct (73.2 vs. 70.1).. However, using the same model for the next iteration is challenging as it is trained for code responses, not evolved instructions.", "category": "Experimental_Exposition"}, {"question": "What was the rationale behind the development of the five heuristic methods?", "answer": "The five heuristic methods were developed to increase the complexity of code instructions, aligning with the characteristics of the code domain and previous research on code models' limitations. The first three heuristics escalate coding task difficulty by incorporating additional requirements, replacing common requirements, or introducing extra reasoning steps, which aligns with trends observed on platforms like LeetCode. The fourth heuristic introduces erroneous code as an adversarial sample, inspired by prior research in attacking pre-trained code models[1][2], to assess the robustness of the code LLM. The fifth heuristic emphasizes time and space complexity, reflecting the importance of optimization in the coding domain. Previous studies [3] indicate that tasks requiring optimized program performance pose heightened difficulty, making this direction a valuable avenue for complexity improvement.\n\n[1] Yang, Zhou et al. “Natural Attack for Pre-trained Models of Code.” 2022 IEEE/ACM 44th International Conference on Software Engineering (ICSE).\n\n[2] Jha, Akshita et al. “CodeAttack: Code-based Adversarial Attacks for Pre-Trained Programming Language Models.” AAAI 2022.\n\n[3] Madaan, A. et al. \"Learning Performance-Improving Code Edits.\" (arxiv, 2023)", "category": "Motivation_Analysis"}, {"question": "Why is the pass@1(%) result on MBPP not as strong as on HumanEval, and how does the diversity of benchmarks used address this concern?", "answer": "The pass@1(%) result on MBPP may not be as strong as on HumanEval due to the latter's narrower distribution and potential for overfitting. To address this, the paper's evaluation includes a diverse set of benchmarks such as MBPP, DS-1000, and MultiPl-E. The paper's models show substantial improvements across these benchmarks, with a 20-point improvement on HumanEval, a 5.2-point improvement on MBPP, a 7.4-point improvement on DS-1000 (insertion), and an average 9.3-point improvement on MultiPL-E, demonstrating the effectiveness of the paper's approach. Additionally, Codellama-instruct only achieves a 2-point improvement on MBPP compared to Codellama.", "category": "Experimental_Analysis"}, {"question": "What is the number of iterations used in the Code Evol-Instruct technique, and what criteria were used to determine when to stop the iterations?", "answer": "A total of 4 iterations were conducted. The decision to stop was based on the observation in Figure 3a, which shows that optimal performance on the dev set was achieved after 3 rounds of evolutions.", "category": "Experimental_Setup"}, {"question": "What model is used to generate evolved data in Section 3 of the Code Eval-Instruct iteration on the sample dataset, and does the quality of the pre-trained model outputs affect the results?", "answer": "GPT3.5 (gpt3.5-turbo) is used to generate evolved data, as mentioned in Section 4.2. Additional experiments replacing GPT3.5 with GPT-4 showed an increase in HumanEval Pass@1 scores from 73.2 to 73.8 for 34B and from 59.8 to 62.2 for 15B. While using a superior LLM like GPT-4 contributes to better evolved data, the improvement is not strongly correlated with the coding performance of the evolution executing model (88.4 for GPT4 and 73.2 for GPT3.5).", "category": "Experimental_Exposition"}, {"question": "Is the model in this paper trained using 24 V100 GPUs, with the 15B model requiring 15 hours and the 34B model requiring 32 hours?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "ruk0nyQPec", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore", "authors": ["Sewon Min", "Suchin Gururangan", "Eric Wallace", "Weijia Shi", "Hannaneh Hajishirzi", "Noah A. Smith", "Luke Zettlemoyer"], "abstract": "The legality of training language models (LMs) on copyrighted or otherwise restricted data is under intense debate. However, as we show, model performance significantly degrades if trained only on low-risk text (e.g., out-of-copyright books or government documents), due to its limited size and domain coverage. We present SILO, a new language model that manages this risk-performance tradeoff during inference. SILO is built by (1) training a parametric LM on the Open License Corpus (OLC), a new corpus we curate with 228B tokens of public domain and permissively licensed text and (2) augmenting it with a more general and easily modifiable nonparametric datastore (e.g., containing copyrighted books or news) that is only queried during inference. The datastore allows use of high-risk data without training on it, supports sentence-level data attribution, and enables data producers to opt out from the model by removing content from the store. These capabilities can foster compliance with data-use regulations such as the fair use doctrine in the United States and the GDPR in the European Union. Our experiments show that the parametric LM struggles on its own with domains not covered by OLC. However, access to the datastore greatly improves out of domain performance, closing 90% of the performance gap with an LM trained on the Pile, a more diverse corpus with mostly high-risk text. We also analyze which nonparametric approach works best, where the remaining errors lie, and how performance scales with datastore size. Our results suggest that it is possible to build high quality language models while mitigating legal risk.", "keywords": "language modeling; retrieval; legality of language modeling", "qa_pairs": [{"question": "What are the reasons for focusing on LM perplexity in a paper?", "answer": "The paper focus on LM perplexity in the paper was based on the following two reasons: 1. Prior work has shown that LM perplexity correlates well with downstream tasks (including zero-shot and few-shot in-context learning) across a large collection of tasks and all models [1, 2, 3]. So it’s a natural first step. 2. Prior work has shown adding a nonparametric datastore improves performance on downstream tasks [4, 5, 6] even when pretrained LMs are already trained on in-domain data. Gains are expected to be larger in SILO where pretrained LMs are not trained on in-domain data (this is the case in perplexity, as reported in Appendix D.2 & Figure 5). The paper's text classification experiments confirm these findings. Furthermore, perplexity has been the primary evaluation criterion in many recent ICLR papers that innovated on language modeling [7, 8, 9, 10].\n\n[1] Kaplan et al. \"Scaling Laws for Neural Language Models\". 2020. https://arxiv.org/abs/2001.08361\n[2] Hernandez et al. Scaling Laws for Transfer. 2021. https://arxiv.org/abs/2102.01293\n[3] Xia et al. “Training Trajectories of Language Models Across Scales”. ACL 2023. https://arxiv.org/abs/2212.09803\n[4] Asai et al. “ACL 2023 Tutorial: Retrieval-based Language Models and Applications”. https://acl2023-retrieval-lm.github.io/\n[5] Khandelwal et al. \"Nearest Neighbor Machine Translation\". ICLR 2021. https://arxiv.org/abs/2010.00710\n[6] Jiang et al. \"Active Retrieval Augmented Generation\" EMNLP 2023. https://arxiv.org/abs/2305.06983\n[7] https://openreview.net/forum?id=HklBjCEKvH ICLR 2020\n[8] https://openreview.net/forum?id=nnU3IUMJmN ICLR 2022\n[9] https://openreview.net/forum?id=R8sQPpGCv0 ICLR 2022\n[10] https://openreview.net/forum?id=BS49l-B5Bql ICLR 2022", "category": "Motivation_Analysis"}, {"question": "How does the paper address the trade-offs between legal risk, runtime, and performance?", "answer": "The paper considers SILO as a method to control trade-offs between legal risk, runtime, and performance. It highlights the novel connection between the legal risk of non-permissive training data and a nonparametric datastore, which is expected to motivate further research on efficient inference with such architectures. ", "category": "Method_Disambiguation"}, {"question": "Does using open license data with SILO completely eliminate the risk of infringing intellectual property (IP) from others?", "answer": "No, SILO does not completely eliminate the risk of infringing intellectual property (IP) from others. Legal risks are highly context-dependent, and there are uncertainties about how courts will interpret copyright protections. SILO does not remove the need for obtaining permission to use copyrighted content, but its opt-out capabilities can strengthen fair use defense. Additionally, SILO’s attribution features make it easier to identify copied content. SILO is designed to lower legal risk and provide more control to model developers and users in balancing risk and performance, rather than completely removing legal risk.", "category": "Method_Disambiguation"}, {"question": "What is the rationale for focusing on the different licenses PD, SW, and BY in the paper, given their apparent lack of significance in the experiments?", "answer": "There are two reasons: 1. The OLC dataset is designed to have varying levels of permissiveness, as permissiveness is a spectrum rather than a binary decision. This allows LM developers to define their own boundaries and control the trade-off between training data size, legal risk, and performance. 2. The general empirical findings are shown to hold across different possible boundaries, as demonstrated in Appendix D.2 and Figure 3.", "category": "Motivation_Analysis"}, {"question": "What is the novelty of the proposed approach, given that it builds upon the kNN-LM framework by Khandelwal et al., 2020[1]?\n\n[1]Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In Proceedings of the International Conference on Learning Representations, 2020.", "answer": "The study contributes to technological innovation by being among the first to provide technical mitigation of legal risk in language models. It introduces a novel training recipe that separates data for training and data for a datastore, an approach not explored in prior work. The novelty lies in using this separation to mitigate legal risk, an underexplored area, and demonstrating empirical findings on recovering performance close to existing language models.", "category": "Method_Disambiguation"}, {"question": "Why does the paper focus on LM perplexity and zero-shot text classification instead of exploring the influence of SILO on instruction-tuning or task fine-tuning processes?", "answer": "The paper focuses on LM perplexity and zero-shot text classification because prior work has shown that LM perplexity correlates well with downstream tasks across a large collection of tasks and all models[1][2][3]. Additionally, prior work has demonstrated that adding a nonparametric datastore improves a range of tasks in both fine-tuning[4, 5] and instruction-tuning scenarios[6], even when pretrained LMs are already trained on in-domain data. Gains are expected to be larger in SILO, where pretrained LMs are not trained on in-domain data, as evidenced by the results reported in Appendix D.2 and Figure 5.\n\n[1] Kaplan et al. \"Scaling Laws for Neural Language Models\". 2020. https://arxiv.org/abs/2001.08361\n[2] Hernandez et al. Scaling Laws for Transfer. 2021. https://arxiv.org/abs/2102.01293\n[3] Xia et al. “Training Trajectories of Language Models Across Scales”. ACL 2023. https://arxiv.org/abs/2212.09803\n[4] Guu et al. 2020. “REALM: Retrieval-Augmented Language Model Pre-Training”. https://arxiv.org/abs/2002.08909\n[5] Izcard et al. \"Atlas: Few-shot Learning with Retrieval Augmented Language Models\". 2022 https://arxiv.org/abs/2208.03299\n[6] Lin et al. 2023. \"RA-DIT: Retrieval-Augmented Dual Instruction Tuning\" https://arxiv.org/abs/2310.01352", "category": "Motivation_Analysis"}, {"question": "What are the specific computational costs, in terms of GPU hours, for training the SILO and Pythia models?", "answer": "Pythia 1.4B, which was used for comparison as SILO has 1.3B parameters, was trained for 7,120 GPU hours. The main SILO model, named PDSW, was trained for 6,613 GPU hours. Both models used A100 GPUs with 40GB of memory, as described in the Pythia paper [1]\n\n[1] Biderman et al. \"Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling\". 2023. https://arxiv.org/abs/2304.01373", "category": "Experimental_Setup"}, {"question": "Does SILO completely eliminate the legal risk of IP infringement when using open license data?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is it claimed in the paper that SILO entirely removes the need for obtaining permission to use copyrighted content?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "dsUB4bst9S", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Teaching Arithmetic to Small Transformers", "authors": ["Nayoung Lee", "Kartik Sreenivasan", "Jason D. Lee", "Kangwook Lee", "Dimitris Papailiopoulos"], "abstract": "Large language models like GPT-4 exhibit emergent capabilities across general-purpose tasks, such as basic arithmetic, when trained on extensive text data, even though these tasks are not explicitly encoded by the unsupervised, next-token prediction objective. This study investigates how even small transformers, trained from random initialization, can efficiently learn arithmetic operations such as addition, multiplication, and elementary functions like square root, using the next-token prediction objective. We first demonstrate that conventional training data is not the most effective for arithmetic learning, and simple formatting changes can significantly improve accuracy. This leads to sharp phase transitions as a function of training data scale, which, in some cases, can be explained through connections to low-rank matrix completion. Building on prior work, we then train on chain-of-thought style data that includes intermediate step results. Even in the complete absence of pretraining, this approach significantly and simultaneously improves accuracy, sample complexity, and convergence speed. We also study the interplay between arithmetic and text data during training and examine the effects of few-shot prompting, pretraining, and parameter scaling. Additionally, we discuss the challenges associated with length generalization. Our work highlights the importance of high-quality, instructive data that considers the particular characteristics of the next-word prediction loss for rapidly eliciting arithmetic capabilities.", "keywords": "Machine Learning;Transformers;Large Language Models", "qa_pairs": [{"question": "Does the model train on a fixed set of examples, and how does training for multiple epochs on varying dataset sizes affect its generalization behavior for arithmetic operations?", "answer": "For all experiments, the model is trained on a fixed set of examples. The sample efficiency curve demonstrates how performance varies when training for multiple epochs on training datasets of different sizes. This approach allows the paper to investigate the model's ability to generalize arithmetic operations on digits from one position to another where it has never seen it.", "category": "Experimental_Analysis"}, {"question": "Why was a character-level tokenizer chosen for the experiments, and how does this choice affect the generalizability of the results?", "answer": "The authors chose a character-level tokenizer to simplify the experimental setting and enable systematic ablation of different factors. While the results may not completely generalize, experiments with BPE tokenization (tiktoken) for GPT-2 and findings from Chain-of-Thought data and models of varying scales suggest that most results remain applicable.", "category": "Motivation_Analysis"}, {"question": "Does the model generate a chain of thought during inference when trained with chain-of-thought data, and is there a correlation between the validity of the chain of thought and the correctness of the answer?", "answer": "Yes, the model generates output in a chain-of-thought style during inference when trained with CoT data from scratch. There is a strong correlation between the validity of the chain of thought and the correctness of the answer, as measured by the `scratchpad match score`, which evaluates the fraction of correct intermediate computations. Errors in intermediate steps often propagate to subsequent steps, leading to incorrect answers, though some less accurate models, such as NanoGPT trained on 250 samples, may still arrive at the correct answer despite intermediate errors.", "category": "Experimental_Exposition"}, {"question": "Is the improved performance of the detailed chain of thought model due to the length of the data format, or is it attributed to the intermediate steps?", "answer": "The improved performance of the detailed scratchpad format is attributed to the intermediate steps rather than the length of the data format. Evidence includes experiments in Section B.3 and B.4 of the Appendix, which examine the effect of recording correct digit-wise sum `(A)` and carry `(C)` information compared to randomizing `A` and `C` in the simplified scratchpad format, showing that addition is best learned with accurate `A` and `C` values. Additionally, an experiment replacing the intermediate steps with random tokens while keeping the data format length constant showed significantly worse performance, confirming the importance of the intermediate steps.", "category": "Experimental_Exposition"}, {"question": "What is the intuitive explanation for the improvement in accuracy when certain numbers are excluded from the training data, and how does this relate to the regularization effect mentioned in Table 1?", "answer": "The columns where 100 and 500 numbers are excluded for plain and all 3 cases for reverse, respectively, show higher overall accuracy compared to the case when no numbers are excluded. This is referred to as a regularization effect, as it can be thought of as analogous to data augmentation with cropped images used during the training of vision models. A more reasonable explanation for this is a generalization of the matrix completion argument discussed in Section 5. Learning how to add two-digit numbers can be viewed as completing a fourth-order, rank-4 tensor (rank-4 in the CP sense). $ab + cd = 10a + b + 10c + d$, which can then be represented as the sum of fourth-order rank-1 tensors (rank-1 in the CP sense): $10* T_1 + T_2 + 10*T_3 + T_4$ where $T_1$ is $\\mathbf{N}\\circ \\mathbf{1} \\circ \\mathbf{1} \\circ \\mathbf{1}$, $T_2$ is $\\mathbf{1}\\circ \\mathbf{N}\\circ \\mathbf{1}\\circ \\mathbf{1}$ and so on. $\\mathbf{N} \\in \\mathbb{R}^{10}$ denotes the vector $[0, 1, 2, \\dots, 9]$ and $\\mathbf{1} \\in \\mathbb{R}^{10}$ denotes the vector of 1s. While this argument is still informal, it now shows that the sample complexity of learning the 'addition-map' is $O(\\#digits)$ rather than $O(10^{\\#digits})$. It also reveals that seeing each digit in each position enough times to learn $T_i$ might be sufficient.\n", "category": "Concept_Understanding"}, {"question": "What is the novelty of the study, given its reliance on previously established methods and datasets, particularly reasoning-augmented data?", "answer": "Although the paper's study does not introduce new datasets or techniques, the novelty of the paper's research lies in the extensive ablation studies and the thorough exploration of various factors influencing model performance in a well-specified setting. The research provides a comprehensive understanding of how transformers develop arithmetic skills, highlighting the importance of data sampling and formatting, such as reversing outputs. These findings contribute to a deeper study of emergent phenomena in models and have broader implications for synthetic training data generation for LLM training, which is valuable for the research community and timely given the current state of research.", "category": "Method_Disambiguation"}, {"question": "Can the proposed method be extended to models beyond GPT-3?", "answer": "The methodology, which focuses on data sampling and formatting techniques, is applicable to models of various scales. Experiments on finetuning GPT-3(Table 2) indicate that the findings are applicable from NanoGPT to GPT-2 and GPT-3, suggesting it should also extend to larger models. Ongoing experiments on GPT-4 aim to further validate this. ", "category": "Experimental_Analysis"}, {"question": "How does the paper address the concern that their findings on simple arithmetic settings do not generalize to natural language scenarios?", "answer": "The paper's primary objective is to analyze and understand how these models learn arithmetic skills and, if possible, determine more efficient ways to elicit such learning. This is expected to help better understand the general case of emergent properties in large language models. While the ideas of reversing the input/output may seem specific and not directly applicable to real data, the paper's findings reveal the importance of data formatting in accelerating the learning process. Demonstrating the impact of a simple method like reversing highlights the significance of properly structuring data to enable the language model to learn a compositional function using its individual components more effectively. These findings are believed to have significant implications for synthetic training data generation for LLMs—a topic that is rapidly growing in importance. The paper also analyzes the token-efficiency of these different formats in Section 6, which is particularly relevant for practical applications. The paper is the first work to study these aspects of CoT-style data for arithmetic applications. While it might be challenging to draw firm conclusions, this is primarily due to the inherent complexity of studying LLMs. Given the scale and the multitude of variables involved, isolating individual aspects of the problem is a daunting task. This is the reason why the paper focused on a simple problem like arithmetic and conducted extensive ablations to gain insights into data formatting, sampling, and their impact on teaching skills to small transformer models.", "category": "Method_Disambiguation"}, {"question": "To what extent is the finding on reversing digit order in the output specific to the decoder-only architecture used in the study, and how does this finding apply to encoder-only and encoder-decoder architectures?", "answer": "The study intentionally focuses on decoder-only architectures to understand emergent properties in LLMs, which are typically decoder-only transformers. For encoder-only models (e.g., BERT-style architectures), reversing digit order is likely irrelevant because the model produces all digits simultaneously. For encoder-decoder architectures, reversing digit order allows the model to learn a simpler function, as it still produces digits from left-to-right. ", "category": "Method_Mechanics"}, {"question": "What is the correct probability of selecting a 1-digit number in the case of 3-digit addition, and how is this probability calculated?", "answer": "The probability of selecting a 1-digit number in 3-digit addition is $0.01\\%$. This is calculated based on the fact that there are only $100$ samples containing 1-digit numbers among the $10^6$ possible 3-digit addition samples. While the probability of selecting a single 1-digit operand is $1\\%$, the probability that either operand is 1-digit is significantly lower.", "category": "Concept_Understanding"}, {"question": "How does the model perform when the operands have different digit lengths, such as in the cases of 2 + 312 or 546 + 7?", "answer": "The following results present different digit length combinations, where each row and column represents the number of digits in the first and second operand, respectively. For plain addition, performance is lower for lower-digit combinations and is not necessarily symmetric in operand lengths. For example, 1-digit + 2-digit results in 42.2% performance, while 2-digit + 1-digit results in 27.8% performance It is conjectured that this is because, despite the balanced sampling approach, the number of samples of $(1, 3)$ digit or $(2,1)$ digit sums are relatively few in the paper's training data. \n\n| Plain   | 1-digit | 2-digit | 3-digit |\n| ----   | ----    | ----    | ----    |\n| 1-digit| 72.8%   | 42.2%   | 77.0%   |\n| 2-digit| 27.8%   | 60.0%   | 91.4%   |\n| 3-digit| 38.6%   | 61.2%   | 97.0%   |\n\nFor reverse addition, performance is consistently high across different digit combinations, with most combinations achieving over 94.0% accuracy.\n\n| Reverse| 1-digit | 2-digit | 3-digit |\n| ----   | ----    | ----    | ----    |\n| 1-digit| 100.0%  | 99.8%   | 99.0%   |\n| 2-digit| 96.4%   | 100.0%  | 100.0%  |\n| 3-digit| 94.0%   | 99.4%   | 100.0%  |", "category": "Experimental_Exposition"}, {"question": "To what extent does the model learn the commutative property of addition, specifically whether predictions for A+B and B+A are the same, and how does this behavior evolve during training?", "answer": "The model learns the commutative property of addition to a large extent, especially after being fully trained. Using NanoGPT trained on the plain data format, the authors tested whether $(A+B) = (B+A)$. Out of 9900 test examples, 8799 samples commute. While most of these are trivially commutative (since the model gets the answer correct), out of the 164 samples where the model gets both $(A+B)$ and $(B+A)$ wrong, it still commutes on 99 of them. This indicates that the model learns to commute 60.3% of the time even when it gets the answers wrong. These findings are summarized in the table below.\n\n| P(A+B=B+A, both are correct answer) | P(A+B and B+A are both incorrect) | P(A+B=B+A \\| both are incorrect) |\n| ----    | ----      | --- |\n| 0.878 (8700 samples)   | 0.016 (164 samples)  | 0.603 (99/164 samples) |", "category": "Experimental_Exposition"}, {"question": "Do the results of the proposed method generalize to the addition of decimal numbers, such as 1.21 + 13.12?", "answer": "Yes, the results generalize to the addition of decimal numbers. The authors trained the model on decimal numbers by converting integer numbers (up to 3-digit) to decimal numbers with 1 digit and 2 floating points (e.g., 1-digit number `a` becomes `0.0a`, 2-digit number `ab` becomes `0.ab`, and 3-digit number `abc` becomes `a.bc`). The results align with findings on integer addition, showing that addition is learned more efficiently with reversed output. Surprisingly, the performance on decimal numbers is better than on 3-digit integer addition, likely due to the decimal number formatting behaving like zero-padding with fixed digit length (Appendix B.1).", "category": "Experimental_Analysis"}, {"question": "How do the results from Sections 3 and 4, which suggest the model learns a specific addition algorithm, reconcile with the results from Section 5, which frame addition as a memorization+interpolation problem?", "answer": "The results from Sections 3 and 4 indicate that the model learns the specified addition algorithm when presented with reversed outputs, but the limited length generalization suggests it does not 'truly' learn this algorithm. Section 5 shows that the generalization properties of NanoGPT when certain numbers are excluded from the training data, surpass Matrix Completion and therefore are not simply interpolating the low-rank addition matrix. While the exact algorithm learned by the model is not precisely characterized, its tendencies and behaviors, particularly its sensitivity to input data format, are documented.", "category": "Experimental_Analysis"}, {"question": "How does the inclusion of the `$` symbol and the `\\n` character as a prefix affect the test accuracy of plain addition and reverse tasks, as shown in Figure 8?", "answer": "The inclusion of the `$` symbol allows plain addition to approach 100% test accuracy, while the reverse task without `$` reaches almost 90%. Additionally, introducing a `\\n` character as a prefix significantly enhances performance for models trained without the `$` delimiter, indicating that the inclusion of a delimiter during testing is crucial for model accuracy.", "category": "Experimental_Exposition"}, {"question": "What is the rationale behind using the `$`-wrapped plain addition as the baseline in Figure 2, and Which examples have been adopted by the structured sampling strategy?", "answer": "The `$`-wrapped plain addition was chosen as the baseline because it achieves nearly 100% accuracy within a reasonable number of samples. The structured sampling strategy involves selecting 100 instances of 1-digit, 900 instances of 2-digit (9% of the total 10,000 instances), and 9,000 instances of 3-digit numbers. While this strategy performs better than random sampling, it may need adjustments to achieve 100% accuracy for 2-digit addition when $n=3$.", "category": "Experimental_Analysis"}, {"question": "How do the findings on simple settings contribute to understanding emergent properties in large language models and their applicability to natural language?", "answer": "The primary objective of this study is to analyze how models learn arithmetic skills and determine efficient ways to elicit such learning, which contributes to understanding emergent properties in large language models. While the specific method of reversing input/output may not directly apply to real data, the findings reveal the importance of data formatting in accelerating the learning process. Demonstrating the impact of a simple method like reversing highlights the significance of properly structuring data to enable the language model to learn a compositional function using its individual components more effectively. These insights have implications for synthetic training data generation for LLMs, a rapidly growing area of research. Additionally, the study analyzes token-efficiency in Section 6, which is relevant for practical applications. The paper is the first work to study these aspects of CoT-style data for arithmetic applications.", "category": "Method_Disambiguation"}, {"question": "How well do the findings from this study, which pinpoint factors contributing to the fast emergence of arithmetic capabilities through ablations, transfer to large language models (LLMs) at scale, and are these findings properly backed by experiments?", "answer": "The findings do transfer to large language models (LLMs) at scale, as demonstrated by a full table in Section 8 that includes models ranging from NanoGPT to GPT3, along with additional experiments in Section F that confirm the scalability of the findings.", "category": "Experimental_Analysis"}, {"question": "Do the findings from training Transformers on arithmetic data transfer to models trained primarily on language with some arithmetic data, and how does the simplicity of the cross-entropy loss on next 'token' affect the learning of underlying arithmetic algorithms?", "answer": "The simplicity of the cross-entropy loss on next 'token' does not fully explain the learning of underlying arithmetic algorithms, as arithmetic in the token space differs from arithmetic in the embedding space. The surprising aspect is the generalization to unseen arithmetic examples. Experiments intermingling arithmetic tasks with textual data (Section 7 and Appendix E) suggest that the findings are transferable to settings with text, and few-shot prompting enhances performance in these scenarios.", "category": "Experimental_Analysis"}, {"question": "Why is the learning curve between 1k and 4k examples referred to as a phase transition, given that the task has a deterministic output and 100% accuracy is achievable, and the model is only trained for 6k examples?", "answer": "For all experiments, the paper trains on a fixed set of examples. The sample efficiency curve shows how performance varies when training for multiple epochs on training datasets of varying sizes. The term 'phase transition' is used to describe the qualitative shift in the model's capability to perform addition, from an inability to grasp addition with insufficient examples to learning perfect addition as the number of samples increases. Between 1k and 4k examples, the model transitions from 0% accuracy to almost 100% accuracy, which is why this range is referred to as a phase transition. Additionally, this range represents an extremely small fraction of all possible 3-digit combinations, and training stops at 6k examples because further training is unnecessary once the model has learned addition perfectly. The phase transition would appear sharper if the x-axis were extended to the 10^6 possible samples.", "category": "Concept_Understanding"}, {"question": "Is the overall performance on decimal number addition better than 3-digit integer addition, potentially due to the decimal formatting behaving like zero-padding?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the probability of selecting a single 1-digit operand in a 3-digit addition task 1%?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the obtained results on decimal number addition align with the findings on integer addition, showing more efficient learning with reversed output?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the model trained on decimal numbers by converting integer numbers to a specific decimal format (1 digit + 2 floating points)?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the results generalize to the addition of two decimal numbers such as 1.21 + 13.12?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper show that the reverse data format improves performance significantly both with and without the `$` delimiter, as mentioned in Appendix B?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "qV83K9d5WB", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Large Language Models as Tool Makers", "authors": ["Tianle Cai", "Xuezhi Wang", "Tengyu Ma", "Xinyun Chen", "Denny Zhou"], "abstract": "Recent research has highlighted the potential of large language models (LLMs)\nto improve their problem-solving capabilities with the aid of suitable external\ntools. In our work, we further advance this concept by introducing a closed-\nloop framework, referred to as LLMs A s Tool Makers (LATM), where LLMs\ncreate their own reusable tools for problem-solving. Our approach consists of two\nphases: 1) tool making: an LLM acts as the tool maker that crafts tools for a set\nof tasks, where a tool is implemented as a Python utility function. 2) tool using:\nanother LLM acts as the tool user, which applies the tool built by the tool maker\nfor problem-solving. The tool user can be either the same or a different LLM\nfrom the tool maker. On the problem-solving server side, tool-making enables\ncontinual tool generation and caching as new requests emerge. This framework\nenables subsequent requests to access cached tools via their corresponding APIs,\nenhancing the efficiency of task resolution. Beyond enabling LLMs to create their\nown tools, our framework also uncovers intriguing opportunities to optimize the\nserving cost of LLMs: Recognizing that tool-making requires more sophisticated\ncapabilities, we assign this task to a powerful, albeit resource-intensive, model.\nConversely, the simpler tool-using phase is delegated to a lightweight model. This\nstrategic division of labor allows the once-off cost of tool-making to be spread\nover multiple instances of tool-using, significantly reducing average costs while\nmaintaining strong performance. Furthermore, our method offers a functional\ncache through the caching and reuse of tools, which stores the functionality of\na class of requests instead of the natural language responses from LLMs, thus\nextending the applicability of the conventional cache mechanism. We evaluate\nour approach across various complex reasoning tasks, including Big-Bench tasks.\nWith GPT-4 as the tool maker and GPT-3.5 as the tool user, LATM demonstrates\nperformance equivalent to using GPT-4 for both roles, but with a significantly\nreduced inference cost.", "keywords": "large language models;tool making;tool using;serving efficiency", "qa_pairs": [{"question": "Why does CoT achieve a score of 0.0 on the 'Chinese Remainder Theorem' problem?Additionally, how does vanilla GPT perform on this task, and what causes its failure if it fails?", "answer": "CoT achieves a score of 0.0 on the 'Chinese Remainder Theorem' problem because direct few-shot prompting without CoT was applied, as readily available demonstrations were unavailable. Vanilla GPT fails because the task requires many steps of intermediate calculation, which GPT-4 cannot perform effectively internally, even with prompting. Adding CoT demonstrations would be extremely long, exacerbating context length limitations.", "category": "Experimental_Analysis"}, {"question": "What is the substantial contribution of the paper, and how does it address the novelty bar for methods involving prompting large language models (LLMs)?", "answer": "The substantial contribution of the paper is the proposal of a novel framework, LATM, which combines multiple language models to enable task-solving ability close to the stronger model’s performance while maintaining the weaker models’ serving cost. This framework is novel and pragmatic, improving efficiency and task performance in a way that was not obvious prior to this work. Although more extensive evaluation across a diverse range of tasks would further demonstrate the capabilities and limitations of the paper's approach, the paper's focus was to validate the potential of this novel technique on a small but representative set of reasoning tasks as an initial proof of concept study. The paper also demonstrates the feasibility of this approach on a representative set of reasoning tasks from the Big-Bench Hard subset, showing significant performance improvements, achieving 80% to 100% accuracy compared to GPT-3.5’s 20% to 60% accuracy while maintaining similar inference costs. The work provides novel insights into improving the efficiency of LLM systems and advances their capabilities on complex reasoning tasks.", "category": "Method_Disambiguation"}, {"question": "Is the performance of the self-verification process stable when using different sets of few-shot validation examples?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the price for the GPT-3.5 Turbo tool maker module 1$/1M tokens for input and 2$/1M tokens for output?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "5Nn2BLV7SB", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization", "authors": ["Yidong Wang", "Zhuohao Yu", "Wenjin Yao", "Zhengran Zeng", "Linyi Yang", "Cunxiang Wang", "Hao Chen", "Chaoya Jiang", "Rui Xie", "Jindong Wang", "Xing Xie", "Wei Ye", "Shikun Zhang", "Yue Zhang"], "abstract": "Instruction tuning large language models (LLMs) remains a challenging task, owing to the complexity of hyperparameter selection and the difficulty involved in evaluating the tuned models. To determine the optimal hyperparameters, an automatic, robust, and reliable evaluation benchmark is essential. However, establishing such a benchmark is not a trivial task due to the challenges associated with evaluation accuracy and privacy protection. In response to these challenges, we introduce a judge large language model, named PandaLM, which is trained to distinguish the superior model given several LLMs. PandaLM's focus extends beyond just the objective correctness of responses, which is the main focus of traditional evaluation datasets. It addresses vital subjective factors such as relative conciseness, clarity, adherence to instructions, comprehensiveness, and formality. To ensure the reliability of PandaLM, we collect a diverse human-annotated test dataset, where all contexts are generated by humans and labels are aligned with human preferences. Our findings reveal that PandaLM-7B offers a performance comparable to both GPT-3.5 and GPT-4. Impressively, PandaLM-70B surpasses their performance. PandaLM enables the evaluation of LLM to be fairer but with less cost, evidenced by significant improvements achieved by models tuned through PandaLM compared to their counterparts trained with default Alpaca's hyperparameters. In addition, PandaLM does not depend on API-based evaluations, thus avoiding potential data leakage.", "keywords": "LLM evaluation", "qa_pairs": [{"question": "How does PandaLM perform in terms of generalization when evaluating generations from models not included in its training set, such as LLaMA-2-7B?", "answer": "PandaLM demonstrates notable generalization capabilities when evaluating generations from models not included in its training set. Experimental results show that PandaLM aligns closely with human benchmarks, maintaining high performance even on unseen models like LLaMA-2-7B. This indicates that PandaLM's effectiveness is not limited to models it was trained with, reinforcing its utility as a versatile and reliable evaluation tool for LLMs. The detailed analysis is provided in Appendix K of the paper.\n\n| Model Comparison       | PandaLM (P, R, F1) | Human (P, R, F1) | Metrics (P, R, F1)          |\n|------------------------|--------------------|------------------|-----------------------------|\n| llama1-7b vs llama2-7b | (23, 61, 16)       | (23, 70, 7)      | (0.7061, 0.7100, 0.6932)    |\n| llama1-13b vs llama2-13b| (18, 73, 9)        | (20, 68, 12)     | (0.7032, 0.6800, 0.6899)    |\n| llama1-65b vs llama2-70b| (20, 66, 14)       | (34, 56, 10)     | (0.7269, 0.6600, 0.6808)    |\n", "category": "Experimental_Exposition"}, {"question": "How does PandaLM perform when evaluating larger language models, specifically LLaMA-13B and LLaMA-70B, and what insights does this provide about scalable oversight?", "answer": "\n| Model Comparison       | PandaLM (P, R, F1) | Human (P, R, F1) | Metrics (P, R, F1)          |\n|------------------------|--------------------|------------------|-----------------------------|\n| llama1-7b vs llama2-7b | (23, 61, 16)       | (23, 70, 7)      | (0.7061, 0.7100, 0.6932)    |\n| llama1-13b vs llama2-13b| (18, 73, 9)        | (20, 68, 12)     | (0.7032, 0.6800, 0.6899)    |\n| llama1-65b vs llama2-70b| (20, 66, 14)       | (34, 56, 10)     | (0.7269, 0.6600, 0.6808)    |\n\nPandaLM performs effectively when evaluating larger language models, such as LLaMA-13B and LLaMA-70B. The evaluation demonstrates that as the pre-training data for instruction-tuned models increases, their ability to understand and follow instructions improves. This highlights the significance of extensive pre-training in developing language models that are more skilled at understanding and correctly responding to various instructions.", "category": "Experimental_Exposition"}, {"question": "Why is it important to include the performance of LLaMA in Table 1 to evaluate the effectiveness of the proposed LLM tuning framework for PandaLM?", "answer": "\n| **Model**                         | **Accuracy** | **Precision** | **Recall** | **F1 score** |\n|-----------------------------------|--------------|---------------|------------|--------------|\n| LLaMA-7B 0-shot (log-likelihood)  | 12.11        | 70.23         | 34.52      | 8.77         |\n| LLaMA-30B 0-shot (log-likelihood) | 31.43        | 56.48         | 43.12      | 32.83        |\n| LLaMA-7B 5-shot (log-likelihood)  | 24.82        | 46.99         | 39.79      | 25.43        |\n| LLaMA-30B 5-shot (log-likelihood) | 42.24        | 61.99         | 51.76      | 42.93        |\n| Vicuna-7B (log-likelihood)        | 15.92        | 57.53         | 34.90      | 14.90        |\n| Vicuna-13B (log-likelihood)       | 35.24        | 57.45         | 43.65      | 36.29        |\n| PandaLM-7B                        | **59.26**        | 57.28         | 59.23      | 54.56        |\n| PandaLM-7B (log-likelihood)       | **59.26**        | **59.70**         | **63.07**      | **55.78**        |\n\n\nIncluding LLaMA's performance in Table 1 helps to rule out the possibility that PandaLM's performance is solely due to the backbone model rather than the proposed LLM tuning framework. Appendix G provides evidence that tailored tuning significantly affects performance in zero and few-shot scenarios, showing that a well-tuned smaller model can surpass a larger, untuned model. Additionally, experiments with finetuned LLaMA (Vicuna) demonstrate the improved evaluation capabilities of instruction-tuned models, further validating the tuning framework's effectiveness.", "category": "Experimental_Analysis"}, {"question": "What is a realistic higher-bound on the expected F1/accuracy scores for the model, and how does it compare to the inter-annotator agreement (IAA) among human evaluators?", "answer": "The paper's test set was annotated by three individuals with inter-annotator agreement (IAA) exceeding 0.85. To refine the model's performance assessment compared to human evaluators, the paper can use the inter-annotator agreement (IAA) of 0.85 as a benchmark. If the paper's model exceeds this, it indicates strong performance. However, setting a realistic target slightly above this human IAA, say around 0.90, offers a challenging yet achievable goal.", "category": "Experimental_Analysis"}, {"question": "Given that PandaLM-7B achieves 47% for LSAT and 51% for BioASQ in Table 3, which is close to the assumed chance accuracy of 50%, how do these results validate the model's performance, and what additional steps were taken to ensure its validity?", "answer": "The task is a three-category classification (win/lose/tie), where random guessing would yield around 33% in precision, recall, and F1 score. PandaLM-7B's results of 47% and 51% are significantly above this level. To further validate PandaLM-7B, a human evaluation was conducted with 30 samples on the BioASQ evaluation. The results showed that both PandaLM-7B and GPT-4 consistently favored Vicuna over Alpaca, indicating a reliable trend in their evaluations. ", "category": "Experimental_Analysis"}, {"question": "How does the performance of the model vary with adjustments in learning rate and epochs during hyperparameter optimization using PandaLM?", "answer": "The authors explored a range of learning rates (2e-6 to 2e-4),optimizers, schedulers, and epochs, saving model checkpoints at the end of each epoch. Performance was assessed through pairwise comparisons between checkpoints, counting win rounds for each model. The analysis suggests a tendency towards a learning rate of 2e-5, though this preference was not uniformly clear across all models. Peak performance often occurred around the fourth or fifth epoch. The results highlight the complex interplay of hyperparameters with model performance, influenced by data distribution, optimizer, and scheduler choices. There is no universally optimal setting, but a learning rate around 2e-5 and an epoch count near 4 might be beneficial in some cases, these results, however, are not conclusive. These findings emphasize the need for tailored hyperparameter optimization for different models, as it allows for a more nuanced understanding of model performance across various scenarios.", "category": "Experimental_Analysis"}, {"question": "What does the analysis of perplexity in Appendix C reveal about the relationship between perplexity and model performance?", "answer": "The analysis reveals that lower perplexity, which indicates better predictive ability in pretraining, does not always correlate with better overall performance of instruction-tuned models. For example, LLaMA-PandaLM has a higher perplexity than LLaMA-Alpaca but outperforms it in both pairwise comparisons (PandaLM, GPT, Human) and traditional tasks (lm-eval). In some cases, lower perplexity might even suggest potential overfitting, which could diminish the model's generalizability.", "category": "Experimental_Analysis"}, {"question": "How does the implementation of early stopping using Pandalm affect the fine-tuning of large models like LLaMA?", "answer": "The paper implemented an early stopping strategy using Pandalm, focusing specifically on LLaMA. The paper’s experiments revealed that in some cases, the model’s performance at epoch 3 was inferior to that at epoch 2, but subsequent epochs showed performance improvements. This suggests that early stopping may not always be suitable for large model fine-tuning, as it could prematurely halt training before reaching optimal performance. This analysis is presented in Appendix J of the paper.", "category": "Experimental_Exposition"}, {"question": "How does PandaLM's broader assessment for hyperparameter optimization compare to models optimized by loss or perplexity?", "answer": "Models optimized by loss or perplexity might not always correlate with high performance in practical tasks, as perplexity measures predictive ability during pretraining but does not encompass all aspects of language understanding. PandaLM's broader assessment is more effective for hyperparameter optimization because it evaluates models in complex, instruction-based scenarios.", "category": "Method_Comparison"}, {"question": "How was the early stopping strategy implemented, and what impact did it have on the overall performance of the LLaMA model?", "answer": "The early stopping strategy was implemented using Pandalm, which determined if the current training checkpoint surpassed the previous one. If inferior, training was stopped. For LLaMA, training halted at epoch 5, consistent with earlier observations from comprehensive pairwise comparisons of tuned checkpoints, considering variations in learning rate, optimizer, and scheduler. This strategy reduced the time to identify optimal epochs but still required experimentation with different learning rates, optimizers, and schedulers for full performance optimization. The analysis is detailed in Appendix J.", "category": "Experimental_Setup"}, {"question": "What are the contents and implications of Table 3, and how do the models 'alpaca' and 'vicuna' correlate in terms of their performance on datasets such as LSAT, PubMedQA, and BioASQ? ", "answer": "Table 3 evaluates responses in the legal and biological domains using GPT-4 as the gold standard due to the difficulty of hiring domain experts for annotation. The performance of 'vicuna' and 'alpaca' is analyzed on datasets like LSAT, PubMedQA, and BioASQ, with PandaLM as the evaluator and GPT-4's evaluations as the standard. Given that the paper is dealing with a three-class classification problem (win, tie, lose), a random guess would typically result in approximately 33% accuracy in terms of precision, recall, and F1 score. The results show that PandaLM achieves accuracy significantly higher than the 33% expected from random guessing, indicating strong consistency with GPT-4's evaluations. A human evaluation on a subset of 30 samples from BioASQ further validated that both PandaLM and GPT-4 tend to favor Vicuna over Alpaca. ", "category": "Concept_Understanding"}, {"question": "Does the paper claim that the efficacy of PandaLM alone can be inferred from the superior performance of models trained using PandaLM?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Table 1 include the performance metrics of LLaMA as a judge to compare with PandaLM?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do models optimized by loss or perplexity consistently show high performance in practical tasks according to the experimental results?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does perplexity measure all aspects of language understanding crucial for practical applications?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the results for models selected based on optimal perplexity included in the paper?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "EHg5GDnyq1", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors", "authors": ["Weize Chen", "Yusheng Su", "Jingwei Zuo", "Cheng Yang", "Chenfei Yuan", "Chi-Min Chan", "Heyang Yu", "Yaxi Lu", "Yi-Hsin Hung", "Chen Qian", "Yujia Qin", "Xin Cong", "Ruobing Xie", "Zhiyuan Liu", "Maosong Sun", "Jie Zhou"], "abstract": "Autonomous agents empowered by Large Language Models (LLMs) have undergone significant improvements, enabling them to generalize across a broad spectrum of tasks. However, in real-world scenarios, cooperation among individuals is often required to enhance the efficiency and effectiveness of task accomplishment. Hence, inspired by human group dynamics, we propose a multi-agent framework AgentVerse that can effectively orchestrate a collaborative group of expert agents as a greater-than-the-sum-of-its-parts system. Our experiments demonstrate that AgentVerse can proficiently deploy multi-agent groups that outperform a single agent. Extensive experiments on text understanding, reasoning, coding, tool utilization, and embodied AI confirm the effectiveness of AgentVerse. Moreover, our analysis of agent interactions within AgentVerse reveals the emergence of specific collaborative behaviors, contributing to heightened group efficiency. We will release our codebase, AgentVerse, to further facilitate multi-agent research.", "keywords": "large language mode;agent;multi-agent", "qa_pairs": [{"question": "How has the evaluation module addressed the concerns of relying solely on LLMs for generating feedback and evaluation scores?", "answer": "The evaluation module has incorporated external feedback across various scenarios to enhance reliability and effectiveness. Specifically: - **Minecraft Environment:** The evaluator receives in-game environmental factors and player states for real-time, context-specific assessment. - **Humaneval Benchmark:** The evaluator leverages model-generated unit test results to validate outcomes, as detailed in Appendix A. - **Commongen Benchmark:** The evaluator uses coverage test results to gauge the effectiveness of generated responses, also elaborated in Appendix A. - **Tool Usage Scenarios:** The evaluator integrates outcomes from different agents to deliver accurate and comprehensive feedback or reports to users.", "category": "Method_Mechanics"}, {"question": "How does the number of agents influence the performance of the AgentVerse framework in quantitative experiments for GPT-3.5-turbo on Humaneval and MGSM datasets?", "answer": "The paper runs each setting on Humaneval (coding) and MGSM (math) for 3 runs, and reports the averaged performance. (Solo means 1 role assigner agent + 1 decision-making agent + 1 evaluation agent, Group is the same except that there are multiple decision-making agents.)\n\n|                        | CoT              | Solo             | Group-2          | Group-3          | Group-4          |\n| ---------------------- | ---------------- | ---------------- | ---------------- | ---------------- | ---------------- |\n| Mathematical Reasoning | $81.2_{\\pm 0.6}$ | $83.2_{\\pm 1.5}$ | $82.3_{\\pm 0.2}$ | $82.8_{\\pm 0}$   | $81.6_{\\pm 1.3}$ |\n| Programming            | $72.4_{\\pm 0.3}$ | $73.4_{\\pm 1.9}$ | $74.8_{\\pm 0.7}$ | $74.6_{\\pm 1.1}$ | $74.8_{\\pm 0.6}$ |\n| Average Performance    | $76.8_{\\pm 0.2}$           | $78.3_{\\pm 1.0}$           | $78.5_{\\pm 0.5}$           | $78.7_{\\pm 0.5}$           | $78.2_{\\pm 0.9}$     |\n\n Where `Group-x` indicates `x` decision-making agents. Generally, using AgentVerse framework with one to three decision-making agents gives satisfying results. The averaged performance is highest when there are 3 decision-making agents. While these datasets primarily test specific agent abilities, not fully utilizing the diversity of a multi-agent setup, an upward trend in average performance was still observed with an increase in agents. The diminishing returns upon further scaling can be attributed to communication inefficiencies, as discussed in Section 3.1. AgentVerse’s true potential is best observed in more complex challenges. The paper's case studies, tool utilization experiments, and Minecraft game-playing scenarios are prime examples where the multi-agent framework's capabilities are more pronounced and beneficial.\n", "category": "Experimental_Exposition"}, {"question": "What is the primary motivation for using multiple agents instead of single agents, given that some single-agent methods like self-refine/self-ensemble may achieve better performance?", "answer": "As single-agent capabilities continue to evolve, a key area of research is enabling these agents to collaborate beyond solitary operation, particularly in complex real-world challenges. This paper focuses on this collaborative potential among AI agents. In single-agent research, the focus often lies on benchmark performance (similar to Sections 3.1 and 3.2 of this paper). However, the essence of multi-agent study extends beyond these metrics, delving into the autonomous behaviors and communication patterns within agent dialogues—crucial yet often overlooked aspects of agent interaction. Therefore, rather than solely emphasizing benchmark performance, this paper delves into the autonomous coordination and communication that remains largely untapped. The paper’s research, particularly in Sections 3.3 (Tool Using) and 4 (Minecraft Game Playing), demonstrates the capabilities of LLM agents in autonomous coordination and collaborative task accomplishment. These scenarios highlight the need to consider LLM agents not just as assistants but as interactive participants in a multi-agent environment. Thus, while multi-agent systems may not always outperform single agents in certain benchmarks, their unique aspects, like emergent behaviors and communication patterns, are valuable research areas in their own right.", "category": "Motivation_Analysis"}, {"question": "Why is the complexity and cost of the multi-agent framework justified despite its inefficiencies and high token usage for OpenAI API calls?", "answer": "The increased complexity and cost of the paper’s multi-agent system compared to single-agent systems are inherent in the paper’s research question: ‘As single-agent capabilities continue to evolve, how can we enable these agents to collaborate beyond solitary operation, especially in complex real-world challenges?’ Drawing parallels with human society, where collaboration leads to greater efficiency and the ability to tackle more complex tasks, the paper’s work delves into the potential and challenges of multi-agent collaborations. Indeed, inefficiencies in agent collaboration, especially in communication, have been observed. However, as one of the pioneering studies in autonomous multi-agent systems, the paper explores the potential of collaborative agents and demonstrates this potential with several promising cases achieved using the paper’s framework. Although communication challenges exist, highlighting these challenges can pave the way for future research to develop more efficient multi-agent systems with advanced communication capabilities. In light of this, the complexity and cost of the framework are acceptable. Identifying the challenging issues and providing the right direction for further research is as important as proposing incremental methods to optimize an autonomous multi-agent system. This approach allows the paper to highlight both the great potential and some limitations of agents in collaborative settings.", "category": "Method_Mechanics"}, {"question": "What are the key factors driving the performance of the framework in quantitative experiments, and how are they supported by evidence?", "answer": "The performance of the framework is driven by three key factors: 1. **Role Assignment**: Assigning expert identities to agents enhances accuracy, as supported by related work (e.g., https://arxiv.org/abs/2305.14930) and evidenced in the consulting case (Figure 2). An experiment with GPT-3.5-Turbo CoT tasks shows improved results with role assignment, with averaged results and standard deviations reported for Math Reasoning and Coding tasks. \n\n|  | w/o role | w/ role |\n| :------------------- | :-----------------: | :---------------: |\n| Math Reasoning |       72.4±0.3       |      73.7±0.5     |\n| Coding         |       81.2±0.6       |      82.4±0.7     |\n\n\n2. **Environmental Feedback**: Utilizing environmental feedback, such as coverage tests in Commongen and unit tests in Humaneval, significantly boosts performance, as explained in Appendix A, illustrating the impact on datasets where such feedback is applicable.\n\n3. **LLM Communication Capabilities**: The ability of LLMs like GPT-3.5-turbo and GPT-4 to handle contradictions during decision-making stages is crucial. GPT-4's superior handling of contradictions indicates the importance of robust communication capabilities for maximizing the benefits of a multi-agent system.", "category": "Experimental_Analysis"}, {"question": "How can AgentVerse be extended to explore future research topics?", "answer": "As the capability of single agent evolves, exploring how to enable these agents to move beyond working in isolation is considered an important research topic. AgentVerse, with its modular design, facilitates experimentation with various agent teams and tasks. \nAgentVerse can be extended in several ways:\n1.  **More Capable Agents and More Challenging Scenarios.** The paper can easily incorporate more capable agents from recent research and apply them to challenging scenarios, such as more sophisticated embodied agent scenarios. The paper's experiments in the Minecraft environment, a sandbox game, demonstrate their applicability for agents in **the** virtual world. This success suggests that extending AgentVerse to other embodied simulations, or even real-world physical scenarios, is not only feasible but a promising avenue for future exploration. Minecraft serves as a proof of concept, showing how AgentVerse's principles can be adapted to environments where agents interact in the more physically grounded settings.\n2.  **Enhanced Communication.** There are several topics that can be further explored in multi-agent communication. For example, the communication structure. Though the paper only designs two straightforward and intuitive communication structures, e.g., horizontal and vertical, further exploration can be made based on the paper's code and experiments by customizing and replacing the module. Also, the research on communication language, protocol, etc. can all be considered within AgentVerse.\n3.  **Leverage Emergent Behaviors and Mitigate Safety Issues.** The emergent behaviors observed in the paper's tool utilization and Minecraft scenarios provide valuable insights. **How to harness** the beneficial emergent behaviors will be a research topic for multi-agent research. Also, the harmful behaviors the paper observed also indicate the necessity to study the agent safety within the multi-agent context. They may be new scenarios that need to be considered in alignment.\n", "category": "Method_Disambiguation"}, {"question": "How does the proposed AgentVerse framework compare to existing LLM-based agent frameworks across multiple dimensions of indicators?", "answer": "Comparison Table:\nMultiple Agent\nExpert Recruitment\nAutonomous Interaction Among Agents\nCollaborative Decision-Making\nDistributed Action Execution\nInteract with external world\n\n|    | Agent Type | Expert Recruitment | Collaborative Decision-Making (Multi-agent Communication) | Distributed Action Execution | Interact with external world |\n|----------|------------|--------------------|-----------------------------------------------------------|------------------------------|------------------------------|\n| Camel    | Multiple   | x                  | ✓                                                           | x                            | x                            |\n| AutoGPT  | Single     | x                  | x                                                          | x                            | ✓ (human, environment)       |\n| XAgent   | Multiple   | x                  | x                                                          | x                            | ✓ (human, environment)       |\n| METAGPT  | Multiple   | x                  | ✓                                                          | x                            | ✓ (human)                    |\n| AutoGen  | Multiple   | x                  | ✓                                                          | ✓                            | ✓ (human, environment)       |\n| AutoAgents | Multiple  | ✓                  | ✓                                                          | x                            | ✓ (human)                    |\n| AgentVerse | Multiple  | ✓                  | ✓                                                          | ✓                            | ✓ (human, environment)       |\n\n\n", "category": "Method_Comparison"}, {"question": "How does AGENTVERSE perform compared to a single ReAct agent when the number of tasks increases or when the tasks are heterogeneous?", "answer": "The authors acknowledge the limitations of existing datasets like ToolBench (https://arxiv.org/abs/2307.16789) and APIBench (https://arxiv.org/abs/2305.15334) for evaluating complex multi-tool tasks. They highlight technical challenges with ToolBench and the simplicity of APIBench. However, they refer to Appendix B, which demonstrates AGENTVERSE's superior performance over a single ReAct agent in complex tool utilization scenarios. Based on these findings, they are confident in AGENTVERSE’s ability to outperform single-agent setups in such scenarios.", "category": "Experimental_Exposition"}, {"question": "What does the data in Table 2 regarding coding capabilities represent?", "answer": "The data in Table 2 represents the Pass@1 performance of GPT-3.5-turbo and GPT-4 on the Humaneval dataset, showing that the multi-agent setting consistently outperforms both the solo and CoT settings for both models.", "category": "Concept_Understanding"}, {"question": "Do the authors have additional quantitative results, including uncertainty measures, to support their claims about the AgentVerse framework beyond the results presented in Table 1 and Table 2? Given the severe lack of quantitative results in this paper, the case studies provide interesting insights into how the AgentVerse framework handles problems, but these insights are not yet sufficiently significant to fully validate the framework’s performance.", "answer": "The paper includes five benchmark datasets, a number that can be considered substantial for evaluating multi-agent systems. These benchmarks were chosen to be diverse and evaluate different aspects of the paper’s framework. Furthermore, the emergent behaviors and communication patterns shown in the case studies are non-trivial. These aspects, though less tangible than numerical benchmark results, are central to advancing the understanding and capabilities of multi-agent systems. The case studies in the paper’s work are designed to mirror real-world challenges more closely than conventional benchmark datasets, thereby providing valuable insights into the practical applicability of AgentVerse. The consideration of these qualitative aspects as equally, if not more, important than quantitative experiments is significant. The nuanced understanding they provide into agent interaction dynamics is vital for the advancement of multi-agent systems.\n\nThe authors have conducted further experiments with GPT-3.5-turbo, running each experiment three times, and report the averaged performance and the standard deviation on each dataset. The results are as follows:\n\n\n|                                    | CoT             | Solo            | Group           |\n| ---------------------- | --------------- | --------------- | --------------- |\n| Conversation           | $83.7_{\\pm1.0}$ | $84.0_{\\pm1.7}$ | $86.8_{\\pm1.2}$ |\n| Creative Writing       | $79.5_{\\pm0.2}$ | $92.4_{\\pm0.2}$ | $90.2_{\\pm0.4}$ |\n| Mathematical Reasoning | $81.2_{\\pm0.6}$ | $83.2_{\\pm1.5}$ | $82.8_{\\pm0}$   |\n| Coding                 | $72.4_{\\pm0.3}$ | $73.4_{\\pm1.9}$ | $74.8_{\\pm0.7}$ |\n", "category": "Method_Disambiguation"}, {"question": "What are the advantages of framing the problem as a multi-agent system compared to using sophisticated prompting for a single agent, as discussed in Sections 3.3 and 4?", "answer": "The multi-agent framework offers distinct advantages over sophisticated prompting for a single agent: 1. **Efficiency of Multi-Agent System:** An individual agent, even with sophisticated prompting, may face limitations in efficiency when tackling complex tasks. Multi-agent systems excel in this aspect, especially in real-world scenarios involving intricate multi-tool utilization (Section 3.3) or collaborative goals in gaming environments (Section 4). By coordinating and collaborating, multiple agents can operate more efficiently than a single agent in these contexts since they have a clear division of labor, and less distracting information in their memory. 2. **Complementarity of Single and Multi-Agent Approaches:** Advanced prompting techniques can indeed create more capable agents. This does not contradict the multi-agent system approach. Multi-agent systems can also benefit from more capable agents. This synergy allows for leveraging the strengths of single-agent efficiency and multi-agent collaboration. 3. **Exploring Multi-Agent Dynamics:** The paper's research delves into some of the intricate properties of multi-agent communication, shedding light on how heterogeneous agents can collaborate effectively. This exploration reveals the potential of multi-agent systems, a domain where single-agent frameworks fall short. The dynamics within multi-agent interactions are not just an extension of single-agent capabilities but constitute a distinct and valuable area of research. The paper's work represents a significant step in demonstrating the potential and worth of multi-agent studies based on LLMs.", "category": "Method_Comparison"}]}
{"id": "uREj4ZuGJE", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "In-context Autoencoder for Context Compression in a Large Language Model", "authors": ["Tao Ge", "Hu Jing", "Lei Wang", "Xun Wang", "Si-Qing Chen", "Furu Wei"], "abstract": "We propose the In-context Autoencoder (ICAE), leveraging the power of a large language model (LLM) to compress a long context into short compact memory slots that can be directly conditioned on by the LLM for various purposes. ICAE is first pretrained using both autoencoding and language modeling objectives on massive text data, enabling it to generate memory slots that accurately and comprehensively represent the original context. Then, it is fine-tuned on instruction data for producing desirable responses to various prompts. Experiments demonstrate that our lightweight ICAE, introducing about 1% additional parameters, effectively achieves $4\\times$ context compression based on Llama, offering advantages in both improved latency and GPU memory cost during inference, and showing an interesting insight in memorization as well as potential for scalability. These promising results imply a novel perspective on the connection between working memory in cognitive science and representation learning in LLMs, revealing ICAE's significant implications in addressing the long context problem and suggesting further research in LLM context management. Our data, code and models are available at https://github.com/getao/icae.", "keywords": "large language model;context compression;in-context autoencoder;pretraining;fine-tuning;Llama;GPT;memorization", "qa_pairs": [{"question": "What are the reasons for the U-shaped trend observed in the loss values in the rightmost image of Figure 4?", "answer": "The U-shaped trend in the loss values is attributed to two main factors: 1) The pretraining samples are typically long (~500 tokens), making short sequences like 100 or 200 tokens less common in the training data, which results in higher loss for these shorter sequences. 2) The loss is generally higher for earlier context, such as the 10th word having only 9 preceding words to condition on, leading to higher loss, whereas the 300th word benefits from 299 preceding words, providing a richer context and thus easier prediction with lower loss.", "category": "Experimental_Analysis"}, {"question": "What is the impact of full model fine-tuning compared to LoRA on the final quality, and what are the practical advantages of using LoRA?", "answer": "Under the same training steps (10k), full model fine-tuning yields better results than LoRA. However, LoRA has the practical advantage of adding only 1% extra memory usage, making it more resource-efficient for practical applications.", "category": "Method_Comparison"}, {"question": "What were the results of pretraining an ICAE from scratch compared to using a pretrained LLM, and is there a possibility of a better training objective for ICAE?", "answer": "Pretraining an ICAE from scratch using a GPT-style model with a similar number of parameters resulted in weaker performance compared to using a pretrained LLM. While no more efficient and scalable objective than AE/LM has been found, there may be a better training objective for pretraining ICAE.", "category": "Experimental_Exposition"}, {"question": "What are the results of System 1 (Llama2-7b-chat without ICAE) and System 2 (GPT-4) ?", "answer": "The results for System 1 (Llama2-7b-chat without ICAE) and System 2 (GPT-4) are as follows:\n |     System 1    | System 2 |  Win | Lose |  Tie |\n|:---------------:|:--------:|:----:|:----:|:----:|\n| Llama-2-7b-chat |   GPT-4  | 19.3 | 24.5 | 56.2 |", "category": "Experimental_Exposition"}, {"question": "What range of values for N was used in the experiments for text continuation training?", "answer": "In the experiments for text continuation training, N ranges from 1 to 4.", "category": "Experimental_Setup"}, {"question": "What is the approximate training cost (time and money) for the model, and is it necessary to use 8 A100 GPUs with 80GB?", "answer": "Training does not require 8 A100 GPUs with 80GB; it can be done with just 1 A100 due to the use of LoRA, which eliminates the need for GPU memory control through parallelism techniques. The use of 8 A100 GPUs was only for speeding up the pretraining process, which takes about 4 days on 26.2 billion tokens.", "category": "Method_Disambiguation"}, {"question": "How is the encoder LLM initialized in relation to the decoder LLM, and how does the approach address the difference between autoregressive attention in the decoder and potential bidirectional attention in the encoder?", "answer": "The encoder LLM in authors' paper is a frozen LLM (specifically Llama) combined with a learnable LoRA and a learnable embedding for memory tokens. It does not use bidirectional attention but causal attention, similar to the decoder. The encoder and decoder are not the same as in the traditional encoder-decoder architecture.", "category": "Method_Mechanics"}, {"question": "Why was an encoder-decoder language model such as T5 not used for compressing contexts in this study?", "answer": "The encoder and decoder in this study refer to the encoding and decoding modules of the autoencoder (ICAE), which are distinct from the encoder/decoder in an encoder-decoder architecture like T5. The goal was to compress contexts for an existing Large Language Model (LLM), specifically Llama. Building the ICAE on the same LLM (Llama) ensures better alignment with the LLM, enabling the generation of a more compressed context.", "category": "Experimental_Analysis"}, {"question": "What are the practical benefits of ICAE compression when using the same context multiple times, and how does it handle the overhead for different contexts?", "answer": "When the same context is used multiple times (e.g., cached in advance), ICAE compression significantly reduces latency/memory overhead, achieving over 7x speedup, as discussed in Section 3.3.2. This is common and practical, such as a customized ChatGPT for a specific user can compress the user's chat history in advance and use the compressed memory during inference to maintain the long-term memory without online latency overhead for compression. For different contexts, an additional compression step is required, but as shown in Table 7, this additional compression still brings significant end-to-end speedup compared to using the original context when generating a long sequence, making it still a highly practical method.", "category": "Experimental_Exposition"}, {"question": "How many training steps or examples are required for the model to converge in the autoencoding task with a 4x compression setting (e.g., using 128 tokens to restore 512 tokens)?", "answer": "The validation loss continues to decrease even after 150,000 training steps in the 4x compression setting (e.g., using 128 tokens to restore 512 tokens).", "category": "Experimental_Setup"}, {"question": "What is the format of memory slots, how are they generated from the model, and what is observed when mapping each memory slot vector to the most likely word?", "answer": "Memory slots are vectors in the embedding space that are fed into the decoder as input, and are the final output of the ICAE encoder (before softmax) in the positions of memory tokens. When mapped to the vocabulary space, memory slots retain keyword information from the original context and are concentrated in the front part of the context. This can be explained by the memorization intuition we discussed in our paper, where the model will only remember keywords that provide crucial hints.", "category": "Concept_Understanding"}, {"question": "What experimental evidence supports the claim that black-box memories are more compressive than summarization models, as mentioned in Figure 2?", "answer": "The experimental evidence includes: (1) Figure 4 and Figure 5, which show that 128-length black-box memories can almost losslessly restore the original 512-length context (with BLEU>99%), a result not achievable with a 128-length summary; and (2) the last row of Table 6, which compares 128-length black-box memory and 128-token summary, confirming that black-box memory is more compact and informative than a summary of equal length.", "category": "Experimental_Exposition"}, {"question": "Is it feasible to implement an adaptive size of memory slots to achieve more compact compression, considering that the amount of information varies based on the content of the context rather than solely its length?", "answer": "Yes, it is feasible. As shown in Table 3, texts with lower perplexity (PPL) have greater compression potential due to higher certainty and ease of memorization. The authors are currently working on adaptive memory slots to dynamically adjust the compression rate based on the text's difficulty (PPL) to optimize compression performance without compromising results.", "category": "Experimental_Analysis"}, {"question": "Is the encoder initialized from the same LLM checkpoint as the decoder, and what modifications are made to the encoder?", "answer": "Yes, both the encoder and decoder are initialized from the same LLM checkpoint, which is Llama. The encoder additionally includes a learnable LoRA module and a learnable embedding for memory tokens, as shown in Figure 3.", "category": "Method_Disambiguation"}, {"question": "Why does the first row of Table 1 show a PPLX increase despite the absence of compression?", "answer": "The PPLX increase occurs because the original natural language tokens were replaced with compressed memory slots, leading to some inherent loss. This is due to the difference in pretraining data: memory slots are pretrained on 26.2 billion tokens, whereas natural language tokens are pretrained on trillions of tokens by Llama.", "category": "Experimental_Analysis"}, {"question": "Are both the pretrained ICAE and non-pretrained model variants trained on the same instruction fine-tuning set?", "answer": "Yes, both variants are trained on the same instruction fine-tuning set. The only difference is that the pretrained ICAE is pretrained with AE+LM before instruction fine-tuning, while the non-pretrained model is directly trained on the instruction fine-tuning data.", "category": "Experimental_Setup"}, {"question": "Are the llama-2-chat models in column#2 of rows 2, 3, 4, and 5 in table 4 fine-tuned on the PwC dataset?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the results of System 1 - Llama2-7b-chat (without ICAE) and System 2 - GPT-4 included in Table 4 for reference?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is it necessary to use 8 A100 GPUs with 80GB for training, or can it be done with a single A100 GPU?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the two pre-training tasks interleaved?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the number of training tokens for the pretraining tasks specified as 26.2 billion tokens in the paper's Appendix A?", "answer": "True", "category": "Claim_Verification"}, {"question": "According to the demonstrations in Figures 4 and 5 and Table 6, are memory slots more compact and informative than summary baselines of the same length?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the reason for performing the two pre-training tasks in an interleaved manner the catastrophic forgetting problem, given that sequential training performed poorly in the paper’s setup?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the training process on 8 A100 GPUs take approximately 4 days to process 26.2 billion tokens?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "yzfi15eVI7", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "iHyperTime: Interpretable Time Series Generation with Implicit Neural Representations", "authors": ["Elizabeth Fons", "Alejandro Sztrajman", "Yousef El-Laham", "Andrea Coletta", "Alexandros Iosifidis", "Svitlana Vyetrenko"], "abstract": "Implicit neural representations (INRs) have emerged as a powerful tool that provides an accurate and resolution-independent encoding of data. Their robustness as general approximators has been shown across diverse data modalities, such as images, video, audio, and 3D scenes. However, little attention has been given to leveraging these architectures for time series data. Addressing this gap, we propose an approach for time series generation based on two novel architectures: TSNet, an INR network for interpretable trend-seasonality time series representation, and iHyperTime, a hypernetwork architecture that leverages TSNet for time series generalization and synthesis.\nThrough evaluations of fidelity and usefulness metrics, we demonstrate that iHyperTime outperforms current state-of-the-art methods in challenging scenarios that involve long or irregularly sampled time series, while performing on par on regularly sampled data. Furthermore, we showcase iHyperTime fast training speed, comparable to the fastest existing methods for short sequences and significantly superior for longer ones. Finally, we empirically validate the quality of the model's unsupervised trend-seasonality decomposition by comparing against the established STL method.", "keywords": "Time series generation;implicit neural representations", "qa_pairs": [{"question": "What is the procedure for removing data in experiments with irregular time-series, and is the missing data applied to training, testing, or both?", "answer": "Following existing literature [1,2,3], Irregular time-series datasets are created by randomly dropping 30, 50, or 70% of observations from the original datasets. Each observation is dropped uniformly at random and independently for each time series, and independently for each time series {$(t_i, y_i)$}$_{i=0}^N$ in the original dataset. The removed data is consistent across all models and repeats, and is applied to both training and testing. \n[1] - Jeon, Jinsung, et al. \"GT-GAN: General Purpose Time Series Synthesis with Generative Adversarial Networks.\" Advances in Neural Information Processing Systems 35 (2022): 36999-37010.\n\n[2] - Kidger, Patrick, et al. \"Neural controlled differential equations for irregular time series.\" Advances in Neural Information Processing Systems 33 (2020): 6696-6707.\n\n[3] - Tang, Xianfeng, et al. \"Joint modeling of local and global temporal dynamics for multivariate time series forecasting with missing values.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 04. 2020.\n", "category": "Experimental_Setup"}, {"question": "How do RNN encoder-decoder models with training strategies such as weak decoder, contrastive loss, or attractive loss compare with iHyperTime in terms of generating clustered interpretable latent representations and sequences?", "answer": "Classical RNN approaches have shown lower performance in generative tasks compared to GAN-based approaches like TimeGAN, even when trained with teacher-forcing or professor-forcing. Even when they are trained using teacher-forcing (T-Forcing) [2] as well as professor-forcing (P-Forcing) [3] to regularize the generation. In fact, autoregressive approaches condition the model using previously generated samples, and even small prediction errors can compound over the sequences. A major difference is that iHyperTime can generate realistic time-series for longer sequences over 700 time steps, whereas autoregressive RNN approaches suffer from compounding prediction errors. iHyperTime also achieves state-of-the-art performance in generation tasks with an interpretable time-series representation, unlike existing RNN-based approaches that do not directly target time-series representation for generation. The authors are surveying approaches using contrastive loss, attractive losses, and weak decoder, and will discuss the differences between iHyperTime's interpretable latent representation for trend, seasonality, and residuals, and clustered interpretable latent representations in RNN-based models.\n[2] - Alex Graves. \"Generating sequences with recurrent neural networks.\" arXiv preprint arXiv:1308.0850, 2013.\n\n[3] - Alex M Lamb, et al. \"Professor forcing: A new algorithm for training recurrent networks.\" In Advances In Neural Information Processing Systems, pages 4601–4609, 2016.", "category": "Method_Comparison"}, {"question": "How does the linear interpolation of latent codes in time series generation affect the diversity and expressiveness of the generated time series, and what happens qualitatively in the latent space during this process?", "answer": "The linear interpolation of latent codes in time series generation uses lambda values randomly sampled from the interval [0.0, 0.3] to balance diversity and fidelity to the original data characteristics. This range ensures a trade-off between realism and diversity, supported by quantitative and qualitative metrics. Users can adjust lambda (e.g., 0.5) for more diversity. The embeddings are divided into trend, seasonality, and residuals, allowing for granular and targeted interpolation. For example, authors can interpolate between two time series while keeping one of the embeddings (e.g. trend) fixed. This gives place to a constrained generation process which maintains certain characteristics of the original series fixed (i.e. trend), while varying others (i.e. seasonality and residuals), as demonstrated in Appendix H.", "category": "Experimental_Analysis"}, {"question": "How is the diversity of the generated time series evaluated both qualitatively and quantitatively in the study?", "answer": "The diversity of the generated time series is qualitatively evaluated using t-SNE plots in Figure 5, which show that iHyperTime synthetic data mostly overlap with real data, indicating sufficient diversity. Quantitatively, a novel metric called $sym$-Recall, which extends the $\\beta$-Recall  from Alaa et al. (2022), is used to measure the extent to which synthetic samples cover the full variability of real samples.", "category": "Experimental_Setup"}, {"question": "How does the method address the potential issue of data non-exchangeability in real-world datasets?", "answer": "The method addresses data non-exchangeability by adjusting the dataset slightly to retain the test's validity. For example, in cases where non-exchangeability arises from ascending IDs (e.g., in SQuAD and HumanEval), the data can be adjusted by either removing these IDs or permuting the examples while keeping IDs constant.", "category": "Method_Mechanics"}, {"question": "What quantitative and qualitative results are provided to evaluate the reconstruction quality of iHyperTime for different datasets?", "answer": "Table 5 in the supplementary material reports quantitative metrics for iHyperTime's reconstruction quality on Energy24, Stock24, Stock72, and Stock360 datasets. Figures 6 and 7 show qualitative results, displaying reconstructed time-series for Stock 72 and Stock 360, indicating good reconstruction quality and a balanced trade-off between reconstruction and generation performance.", "category": "Experimental_Exposition"}, {"question": "Why does LS4 have a low marginal score on the Temp Rain dataset in Table 5, while IHyperTime outperforms LS4 in the other two metrics?", "answer": "The Temp Rain dataset has complex temporal dynamics with stiff transitions, where most data points lie around the x-axis (y=0) with sharp spikes. A model can achieve a low marginal score by generating data close to the x-axis, but such data are easily distinguishable from real data. LS4 generates data with less sharp spikes than real data, resulting in a better marginal score but a lower Classification score compared to IHyperTime. The less sharp spikes also hinder the predictive power of models trained on synthetic data but tested on real data, explaining LS4's high predictive score.", "category": "Experimental_Analysis"}, {"question": "Can you explain the process by which the model performs unconditional generation of new time series using the set encoder?", "answer": "The model generates new time series by randomly selecting pairs of time series from the training set and performing a linear interpolation between their embeddings $Z^{(1)}$ and $Z^{(2)}$: $ Z^{\\text{gen}} = Z^{(1)} + \\lambda (Z^{(2)} - Z^{(1)})$, where $\\lambda$ is randomly sampled. Optionally, interpolation can be done on individual embedding components $Z_T$, $Z_S$, $Z_R$ for conditional generation.", "category": "Method_Mechanics"}, {"question": "What are the time coordinates $t_i$ used in the model, and how are long time series encoded—as a single set or divided into windows? Additionally, which datasets in Table 1 are univariate and multivariate, and what are the varying lengths represented by the values 24, 72, and 360?", "answer": "The time coordinates $t_i$ are represented as a single floating point value, scaled to the interval $[-1, 1]$ during data pre-processing. Regardless of sequence length, the time series is encoded as a single set, resulting in a single embedding $Z$. In Table 1, the Stock and Energy datasets are multivariate, with 6 and 28 features, respectively, and are sliced into 24 time-step windows. The univariate version of the Stock dataset is created by selecting the Close Price feature and sliced into windows of 72 (Stock72) and 360 (Stock360) time steps. Additionally, Section 4.4 includes four univariate datasets from the Monash repository with varying lengths: FRED-MD, NN5 Daily, and Temperature Rain have approximately 750 time steps, while Solar Weekly has 52 time steps.", "category": "Experimental_Setup"}, {"question": "Can the model effectively disambiguate between season, trend, and residual components in real-world datasets with strong seasonality or trend, and can this be visualized?", "answer": "Authors have now conducted an analysis on three real-world datasets to show that iHT is able to disambiguate between trend, season, and residual components after training. **Datasets:** Authors used two well-known real-world time series datasets that exhibit strong seasonality and trend: Atmospheric CO2 and Airline Passenger, and used the Stock72 dataset described in section B of the supplementary material that exhibits strong trend but low seasonality. The Atmospheric CO2 dataset corresponds to monthly data (January 1959 to December 1987) with a total of 348 observations with a seasonality period of 12 [1]. Authors sliced the data into time series of sequence length 60, obtaining a total of 288 time series. The Airline Passenger dataset corresponds to monthly data (January 1948 to December 1960) with a total of 144 observations with a seasonality period of 12 [2]. Authors sliced the data into time series of sequence length 36, obtaining a total of 108 time series. To estimate the presence of trend and seasonality of each dataset authors use two metrics defined in [3]: the strength of trend and strength of seasonality. As these datasets do not include ground-truth trend and seasonality, authors estimate them using STL [1]. Given a time series with additive decomposition in trend, seasonality and residual: $y_t = T_t+ S_t + R_t$, authors can define the strength of trend as: $$ F_T = \\max\\left(0, 1 - \\frac{\\text{Var}(R_t)}{\\text{Var}(T_t+R_t)}\\right) $$ This measures the relative variance of the residual component $R_t$ to the variance of the trend with the residual component. The value ranges from 0 to 1, with 0 indicating no trend and 1 indicating a strong trend. Strength of seasonality is defined by: $$ F_S = \\max\\left(0, 1 - \\frac{\\text{Var}(R_t)}{\\text{Var}(S_t+R_t)}\\right) $$ Which measures the relative variance of the residual component $R_t$ to the variance of the seasonality with the residual component. A value close to 0 indicates little to no seasonality, and a value close to 1 indicates strong seasonality. Table below shows these metrics for the three datasets. As expected, Stock72 doesn't exhibit strong seasonality (as stock data usually have not such component), while the other two datasets exhibit strong trends and seasonality, with all values on average above 0.98. $$ \\begin{array}{cccc} \\hline \\text{Component} & \\text{CO2} & \\text{Air Pass} & \\text{Stock 72} \\\\ \\hline \\text{Trend strength} & 0.99 \\pm 0.01 & 0.98 \\pm0.01 & 0.99 \\pm 0.01 \\\\ \\text{Seasonal strength} & 0.99 \\pm 0.00 & 0.99 \\pm0.01 & 0.32 \\pm 0.09 \\\\ \\hline \\end{array} $$ **Results:** Authors trained iHT on each dataset and authors show  the reconstructed time series and the output of each of the individual blocks for the CO2 and Air Passenger datasets, respectively. For the output of the trend, seasonality and residual block, authors compare with the STL method. Authors observe that there is a good agreement in the trend and seasonality components in both datasets. The STL method requires as parameter the period of the seasonality, which is known for both datasets, while our approach is able to estimate the seasonality from the data. In the case of the Stock72 dataset, for the STL comparison, authors estimated the period using a Fourier transform and retrieving the dominant frequency. Figure 10 shows the decomposition generated by iHT and compared against STL. Authors can see in this case that the seasonal component is quite small in magnitude compared to the residuals. [1] Cleveland, R. B., Cleveland, W. S., McRae, J. E. & Terpenning, I. (1990). STL: A Seasonal-Trend Decomposition Procedure Based on Loess (with Discussion). Journal of Official Statistics. [2] Downloaded from https://www.kaggle.com/datasets/rakannimer/air-passengers [3] Wang, X., Smith, K. A., & Hyndman, R. J. (2006). Characteristic-based clustering for time series data. Data Mining and Knowledge Discovery, 13(3), 335–364.", "category": "Experimental_Analysis"}, {"question": "How does the iHT model handle each dimension of a multidimensional function f(t)?", "answer": "The iHT model handles each dimension of a multidimensional function f(t) independently by modeling it as $f_T(t) + f_S(t) + f_R(t)$. Specifically, the trend block outputs $m$ trends, the seasonality block outputs $m$ time series corresponding to the seasonality of each feature, and the residual block does the same. These features are then added together to obtain an $m$-dimensional time series.", "category": "Method_Mechanics"}, {"question": "How does the proposed model synthesize new time series samples without requiring {t, f(t)} as input?", "answer": "The model synthesizes new time series by randomly selecting pairs of time series and performing a linear interpolation between their embeddings $Z^{(1)}$ and $Z^{(2)}$: $Z^{\\text{gen}}= Z^{(1)} + \\lambda (Z^{(2)} - Z^{(1)})$, where $\\lambda$ is also sampled randomly. Optionally, the interpolation can be performed on individual components $Z_T$, $Z_S$, $Z_R$ of the embeddings, enabling the conditional generation of time series.In the current setting, the model cannot synthesize new samples without {t, f(t)}, but this could be achieved in a variational setting or by incorporating adversarial training.", "category": "Method_Mechanics"}, {"question": "How does the model handle the influence from other dimensions in the case of multivariate data, particularly in the context of the Trend-Seasonality-Residual (TSR) components?", "answer": "The model processes the Trend-Seasonality-Residual (TSR) components separately but computes and encodes different dimensions together in the case of multivariate data. In the first block of the architecture (set encoder), both the TSR components and the channels are processed together, enabling the network to generate embeddings that encode interactions between the different components and signal channels. This approach contrasts with methods like Fourier Flows[1], where individual channels are processed independently.\n[1] A. Alaa, A. Chan, M. van der Schaar. Generative time-series modeling with fourier flows. In ICLR, 2021.", "category": "Method_Mechanics"}, {"question": "How does the proposed method address the limitation of not achieving unconditional generation, and what are the practical applications of its conditional generation capabilities?", "answer": "The proposed method addresses the limitation of not achieving unconditional generation by focusing on conditional generation, which allows the construction of interpretable counterfactuals of existing time series by interpolating between sub-vectors of embeddings (trend, seasonality, and residuals). This is analyzed in Appendix H, showing the synthesis of new time series with constraints on the trend component. Practical applications include generating new scenarios for evaluating algorithmic trading strategies in finance, such as creating alternative COVID-19 seasonality patterns with observed trends. The manuscript will be updated to clarify the conditional nature and discuss a potential extension to a variational autoencoder framework for unconditional generation. The method contributes significantly to time series generation by providing realistic, interpretable samples with advantages in realism, computational time, and usefulness compared to state-of-the-art approaches.", "category": "Method_Mechanics"}, {"question": "Does the paper use a Fourier mapped perceptron (FMP) to implement fr(t)?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is a single layer SIREN used to model the Residual component in TSNet as clarified in the rebuttal?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "QAwaaLJNCk", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Improving Factuality and Reasoning in Language Models through Multiagent Debate", "authors": ["Yilun Du", "Shuang Li", "Antonio Torralba", "Joshua B. Tenenbaum", "Igor Mordatch"], "abstract": "Large language models (LLMs) have demonstrated remarkable capabilities in language generation, understanding, and few-shot learning in recent years. An extensive body of work has explored how their performance may be further improved through the tools of prompting, ranging from verification, self-consistency, or intermediate scratchpads. In this paper, we present a complementary approach to improve language responses where multiple language model instances propose and debate their individual responses and reasoning processes over multiple rounds to arrive at a common final answer. Our findings indicate that this approach significantly enhances mathematical and strategic reasoning across a number of tasks. We also demonstrate that our approach improves the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to. Our approach may be directly applied to existing black-box models and uses identical procedure and prompts for all tasks we investigate. Overall, our findings suggest that such \"society of minds\" approach has the potential to significantly advance the capabilities of LLMs and pave the way for further breakthroughs in language generation and understanding.", "keywords": "Large Language Models;Factuality;Reasoning;Multiagent Reasoning", "qa_pairs": [{"question": "Why does summarization improve model performance despite the potential for information loss, and have similar experiments been conducted on other tasks?", "answer": "Summarization improves model performance by removing extraneous repeated information and condensing relevant information, especially when context lengths are long and LLMs struggle to process the entire message. Additional experiments on GSM and MMLU tasks showed performance improvements with summarization: GSM performance increased from 85.0 to 87.0, and MMLU performance increased from 71.1 to 73.0. These results are included in Appendix A.1.", "category": "Experimental_Exposition"}, {"question": "How does authors' multi-model approach compare in performance to the simple majority voting method used in the Minerva model, and have authors included this comparison in the paper?", "answer": "Authors' multiagent majority baseline, included in Table 1 and 2, corresponds to the multiagent majority (auto-debate) baseline in Minerva. Authors generate multiple answers using a single language model instance and use majority voting. Authors have updated Section 3.1 to clarify this and added a reference to the Minerva paper. Authors also compare approach with ensembling self-refining agents in Table VII, showing it outperforms this method, and provide qualitative illustrations in Figure XXXI and Figure XXXII.", "category": "Method_Comparison"}, {"question": "What are the key differences and similarities between the multiagent majority voting baselines and the proposed multiagent debate approach, particularly in terms of the number of agents and rounds of communication?", "answer": "The primary difference is that the multiagent debate approach uses multiple rounds of communication between agents to achieve the final answer, while multiagent majority voting takes the majority vote at the end of the first round without any discourse between models. Both approaches have a flexible number of agents $N$, but the rounds of debate are 1 in multiagent majority voting and $T$ in multiagent debate. Additionally, an experiment in Table V demonstrates that the multiagent debate approach outperforms majority voting with 50 agents.", "category": "Method_Comparison"}, {"question": "What is the intrinsic reason why the multi-agent debating method improves the reasoning ability of large language models (LLMs), and how does it go beyond simply combining ensemble methods and self-reflection?", "answer": "The multi-agent debating method improves LLMs' reasoning ability by tightly coupling two sources of performance gains: (1) the wisdom of the crowds, where multiple LLM instances jointly improve performance through majority voting, and (2) self-reflection, where LLMs use previous generations to enhance their performance. The method goes beyond simply combining these aspects by enabling LLMs to reflect not only on their own thought processes but also on those of other agents. This allows LLMs to stitch together correct reasoning paths from multiple responses, enabling the debate procedure to recover fully correct answers in subsequent rounds by fixing misunderstandings or errors in reasoning using responses from other agents. Additionally, a new baseline comparing debate to ensembling self-reflection results shows that debate outperforms this baseline, indicating that discourse between agents is more effective than simply ensembling answers from self-reasoning models.", "category": "Method_Disambiguation"}, {"question": "How does the performance of multi-agent debate with 10 agents over 2 rounds compare to majority voting with 50 agents across Arithmetic, GSM, and MMLU tasks?", "answer": "Multi-agent debate with 10 agents over 2 rounds outperforms majority voting with 50 agents across Arithmetic, GSM, and MMLU tasks. Specifically, the results are: Majority Vote (50 agents) - Arithmetic: 92.0±2.7, GSM: 85.0±3.6, MMLU: 67.0±4.7; Debate (10 agents, 2 rounds) - Arithmetic: 96.0±1.9, GSM: 89.0±3.1, MMLU: 71.0±4.5.\n | Model | Arithmetic | GSM  | MMLU |\n | -------- | ------- | ------- | ------- |\n | Majority Vote (50 agents) | 92.0+- 2.7  | 85.0 +- 3.6 |  67.0 +- 4.7 | \n| Debate (10 agents 2 rounds) | 96.0+- 1.9 |  89.0 +- 3.1 | 71.0 +- 4.5 |", "category": "Experimental_Exposition"}, {"question": "Why were different model bases not used to study multi-agent debate, and what additional results have been included to address this limitation?", "answer": "The paper have included additional results using the chat-Llama 2 7B model, demonstrating that multi-agent debate is also effective on open-source language models. The performance metrics for various models, including single agent, single agent with reflection, multi-agent majority, and multi-agent debate, have been reported and added to Appendix A.1.\n | Model | Arithmetic | GSM  | MMLU |\n | -------- | ------- | ------- | ------- | \n|Single Agent  |  9.0 +- 1.6  | 20.7 +- 2.3 | 41.0 +- 2.8 |\n |Single Agent (Reflection) | 10.7 +- 1.7 | 21.0 +- 2.3  | 39.7 +- 2.8 |\n |Multi-Agent (Majority) | 11.0 +- 1.8 |  25.7 +- 2.5 | 43.3 +- 2.9| \n|Multi-Agent (Debate) | 13.3 +- 1.9  | 29.3 +- 2.6 | 47.7 +- 2.9  | \n\n", "category": "Experimental_Exposition"}, {"question": "How does the multiagent debate process improve model performance by allowing agents to reference each other's answers and synthesize correct reasoning, particularly in cases where initial answers contain errors?", "answer": "The multiagent debate process improves model performance by enabling agents to use different lines of reasoning from separate agents to stitch together an overall response. Each agent provides initial answers with errors in different parts of their reasoning. Through debate, agents can combine the correct parts of their own reasoning with the correct parts of other agents' reasoning, turning initially incorrect answers into fully correct ones. For example, in Figure 11, the chatGPT agent initially fails to realize Carla needs to fully redownload a file, while the Bard agent correctly identifies this but miscalculates the download time. Through debate, the chatGPT agent integrates the correct reasoning from Bard to synthesize a correct solution. Similar examples are shown in Figures XXXI and XXXII, both agents make different mistakes in evaluating the arithmetic expression, where the first agent incorrectly adds each component together at the end and the second agent carries over the wrong answer from a previous computation. Both agents stitch the reasoning of the other agent with their original reasoning in the previous round to get the final answer correctly.", "category": "Method_Mechanics"}, {"question": "Under what conditions are model agents more 'agreeable'? Specifically, are they more agreeable when there are more similar answers (regardless of correctness) or when the answer is correct? Additionally, can the relationship between 'agreeable' behavior and instruction tuning or RLHF be verified using pre-aligned checkpoints?", "answer": "The experiments show that model agents are more agreeable when there are similar answers from other agents, regardless of whether the answer is correct. To verify the relationship between 'agreeable' behavior and instruction tuning or RLHF, the authors compared the LLama-2 7B model and the RLHF chat-LLama-2 7B model. On the GSM8K task, the LLama-2 7B model achieved a consensus of 51.3% after debate, while the aligned chat-LLama 2 model achieved 62.7%. On the MMLU task, the LLama-2 7B model achieved a consensus of 74.3%, while the aligned chat-LLama 2 model achieved 100.0%. These results support the conclusion that RLHF models are more agreeable than unaligned checkpoints. This experiment has been added to Appendix A.1.", "category": "Experimental_Exposition"}, {"question": "What is the primary novelty of the proposed methodology, and how does it differ from prior works that use ensembles of models or self-reflection of a single model?", "answer": "The primary novelty of the paper is the methodology of multi-agent debate as an inference-time approach to improving the performance of LLMs across both factuality and reasoning. Unlike prior works that explore ensembles of models or self-reflection of a single model, this methodology combines both ideas, allowing each agent to reflect on the generations of the original language model and other models. This is illustrated in new Figure XXXI and Figure XXXII, which show how debate can self-correct initially incorrect answers. Additionally, an experiment with the chat-Llama 7B model demonstrates that debate improves final performance, as shown in the provided table.\n | Model | Arithmetic | GSM  | MMLU | \n| -------- | ------- | ------- | ------- | \n|Single Agent  |  9.0 +- 1.6  | 20.7 +- 2.3 | 41.0 +- 2.8 |\n |Single Agent (Reflection) | 10.7 +- 1.7 | 21.0 +- 2.3  | 39.7 +- 2.8 | \n|Multi-Agent (Majority) | 11.0 +- 1.8 |  25.7 +- 2.5 | 43.3 +- 2.9|\n |Multi-Agent (Debate) | 13.3 +- 1.9  | 29.3 +- 2.6 | 47.7 +- 2.9  | ", "category": "Method_Disambiguation"}, {"question": "How does the proposed method address the issue of being resource-intensive and potentially exceeding context limits in lengthy conversations?", "answer": "The proposed method, though computationally expensive, allows a cheaper model like GPT-3.5 to perform similarly to a more expensive model like GPT-4, offering computational savings. Additionally, the multiagent debate procedure can be distilled into another model for faster inference and self-improvement.", "category": "Method_Mechanics"}, {"question": "How does the paper clarify the effect of multiagent debate and differentiate it from other multiagent approaches, and what additional experiments or illustrations have been added to support these clarifications?", "answer": "The paper clarifies the effect of multiagent debate by comparing it with ensembling self-refining agents in Table VII, showing superior performance. Two qualitative illustrations ([Figure XXXI](https://ibb.co/gr69Z4m) and [Figure XXXII](https://ibb.co/FxsGK75)) demonstrate debate-induced changes in responses when original answers are incorrect. Table 3 and a method section discussion differentiate the approach from multiagent majority baselines. Additional experiments in Table V show outperformance over majority voting with 50 agents, application to the chat-Llama 2 7B model, and improvements in agreeableness after RLHF finetuning. Summarization's role in enhancing performance on other tasks is illustrated in Appendix Section A.1.", "category": "Experimental_Exposition"}]}
{"id": "GDdxmymrwL", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Corex: Pushing the Boundaries of Complex Reasoning through Multi-Model Collaboration", "authors": ["Qiushi Sun", "Zhangyue Yin", "Xiang Li", "Zhiyong Wu", "Xipeng Qiu", "Lingpeng Kong"], "abstract": "Large Language Models (LLMs) are evolving at an unprecedented pace and have exhibited considerable capability in the realm of natural language processing (NLP) with world knowledge. Benefiting from ultra-large-scale training corpora, a single LLM can manage typical NLP tasks competently. However, its performance in executing reasoning tasks is still confined by the limitations of its internal representations. To push this boundary further, we introduce Corex in this paper, a suite of novel general-purpose strategies that transform LLMs into autonomous agents pioneering multi-model collaborations for complex task-solving. Inspired by human behaviors, Corex is constituted by diverse collaboration paradigms including Debate, Review, and Retrieve modes, which collectively work towards enhancing the factuality, faithfulness, and reliability of the reasoning process. These paradigms foster task-agnostic approaches that enable LLMs to ''think outside the box,'' thereby overcoming hallucinations and providing better solutions. Through extensive experiments across four different types of reasoning tasks, we demonstrate that orchestrating multiple LLMs to work in concert yields substantially better performance compared to existing methods. Further results and in-depth analysis demonstrate the cost-effectiveness of our method, facilitating collaboration among different LLMs and promoting annotation efficiency. Our code and data are available at https://anonymous.4open.science/r/Corex.", "keywords": "Large Language Models;Complex Reasoning;Multi-Model Collaboration", "qa_pairs": [{"question": "How were the 5 different opinions solicited in the multi-model collaboration, and were different temperatures or multiple samples from one LLM used to create heterogeneous agents? Additionally, is using different LLMs to play the judge roles sufficient to establish a heterogeneous setting with multiple agents?", "answer": "1. The 5 different opinions were solicited by employing different meta instructions to enable the models to assume distinct roles, as shown in Table 21. For problem-solving, prompts from existing works were followed to ensure a fair comparison (Appendix F). 2. Temperature was not adjusted during the generation process to ensure reproducibility and stability of the results, with details provided in Appendix B. 3. The heterogeneous settings in Debate focus on varying the judge's role, as the judge acts as a 'Hub' for information aggregation when discrepancies arise. Altering other models in different teams would introduce too much complexity and randomness, so the analysis focuses on the LLMs playing the judge role to derive direct and reliable conclusions.", "category": "Method_Mechanics"}, {"question": "What are the key distinctions between the proposed multi-agent debate approach and previous works, particularly in terms of addressing context length limitations, reliability of consensus, and risk of monopolization?", "answer": "The proposed multi-agent debate approach differs from previous works[1,2] in several ways: 1. **Context Length Limitations**: It overcomes context length constraints by distributing the debate to different sets of models, ensuring the entire process is captured without exceeding context limits. 2. **Reliability of Consensus**: It incorporates mechanisms to critically evaluate the debate process and outcomes, reducing the risk of incorrect consensus or biases. 3. **Risk of Monopolization**: It ensures balanced participation among models, preventing stronger models from dominating the debate. Additionally, the Review mode integrates external insights, encourages incremental modifications, and considers more scenarios, distinguishing it from self-reflection or self-refinement methods. 1. Integration of External Insights: Unlike the approach in [3], Corex encourages each participant in the Review mode to incorporate external insights. This integration allows for a broader perspective and mitigates the echo chamber effect often seen in self-refinement processes [4]. 2. Incremental Modifications: Participants make incremental modifications on top of their predecessor’s work. This iterative process enables a cumulative improvement of solutions, going beyond the limitations of self-correcting methods. 3. Consideration of More Scenarios: Review mode takes into account both NL and code, enhancing its applicability in a variety of tasks, especially those involving calculations. Review has shown superior results compared to previous strong baselines, and recent works[4,5,6] have also demonstrated that methods solely based on self-correcting have limited effectiveness.The 'retrieve' method, while motivated similarly to RAG, differs in implementation by generating and scoring candidate answers based on reasoning chains, a novel approach in this context.\n[1] Improving factuality and reasoning in language models through multiagent debate., arXiv:2305.14325\n\n[2] Examining the inter-consistency of large language models: An in-depth analysis via debate\n\n[3] Self-refine: Iterative refinement with self-feedback. arXiv:2303.17651.\n\n[4] Can Large Language Models Really Improve by Self-critiquing Their Own Plans?, arXiv: 2310.08118\n\n[5] Large Language Models Cannot Self-Correct Reasoning Yet, arXiv: 2310.01798\n\n[6] GPT-4 Doesn't Know It's Wrong: An Analysis of Iterative Prompting for Reasoning Problems, arXiv: 2310.12397\n", "category": "Method_Comparison"}, {"question": "How are the components debate, review, and retrieve conceptually related, and how do they integrate into a single framework despite being used separately in experiments?", "answer": "The components debate, review, and retrieve are inspired by human problem-solving strategies and are not isolated but interrelated. Debate emulates the exchange and challenge of ideas, review focuses on critically examining and refining outputs, and retrieve involves sourcing and selecting appropriate solutions. The framework intentionally separates these components, mimicking human problem-solving where different methods are employed at different stages. However, at the application level, strategic selection and routing allow for integration, enabling a collective problem-solving process. This approach is task-agnostic and can be extended to scenarios like robotics planning and social simulations.", "category": "Method_Disambiguation"}, {"question": "Why is the performance improvement of Corex over stronger baselines like CoT-SC(10) not very significant, and have the authors analyzed variants such as CoT-SC(20/30) to determine if they lead to different conclusions? Additionally, have the total tokens consumed by different methods been compared to ensure a fair evaluation?", "answer": "The performance improvement of Corex over baselines like CoT-SC(10) may appear less significant due to page limit constraints, which led to a focus on stronger baselines. However, Corex is more cost-effective, with implementation costs less than half of these baselines. Analysis of variants like CoT-SC(5, 10, 20, 40, 80) and ComplexCoT (10, 20, 40) is presented in section 5.3 and appendix C.3 (Figures 11 and 12), where results are visually displayed to avoid overburdening tables. The experiments include one task each from math, commonsense, and symbolic reasoning, and the total tokens consumed by different methods are also presented on a log scale.", "category": "Experimental_Analysis"}, {"question": "What specific limitations in large language models does Corex aim to address, and how is its effectiveness in improving AI reasoning evaluated in the paper?", "answer": "Corex aims to address the limitations of large language models (LLMs) that rely solely on internal representations, which often result in unreliable responses in complex reasoning tasks. It introduces a collaborative approach inspired by human social interactions, enhancing models' general reasoning abilities through model collaborations such as idea exchange and bug fixing. The effectiveness of Corex is evaluated across 18 tasks in 4 categories, showing significant improvements in AI reasoning compared to previous strong baselines.", "category": "Motivation_Analysis"}, {"question": "What are the reasons for not including the baselines from Table 1 and other similar modules for debating, retrieval, and review in the evaluation?", "answer": "The baselines from Table 1 and similar modules were not included due to significant challenges in evaluating them across the wide range of benchmarks covered. Specific issues include PHP's specificity to mathematical tasks, CoK's effectiveness only for commonsense reasoning requiring additional knowledge base configurations, and the lack of official prompts for MAD and ToT for other tasks, which complicates fair comparison. Additionally, for tasks like semi-structured understanding, GPT-3.5-Turbo is unavailable for these methods due to their long meta-instructions, prompts, and demonstrations exceeding the context window. Instead, more broadly applicable baselines like CoT-SC(10, 20, ...) and ComplexCoT(10, 20, ...) were used for direct insights into the method's effectiveness.", "category": "Experimental_Analysis"}, {"question": "What distinguishes Corex from prior work in terms of its novelty and contributions to the field of reasoning?", "answer": "The primary novelty of Corex is its introduction of multi-model collaboration into reasoning, inspired by human problem-solving behavior. Corex is also distinguished by its task-agnostic framework, which supports a reference-free approach applicable to a wide range of reasoning tasks. This contrasts with prior works that are often tailored to specific problem types or rely on predefined knowledge bases, references, or prompts. Corex introduces a set of flexible and versatile paradigms.", "category": "Method_Comparison"}, {"question": "What are the main contributions of Corex, and how do they address the issues identified in the introduction and Figure 1?", "answer": "The main contributions of Corex include: 1. Approach to debate distinctly differs from previous works [1]. Debate Mode: A novel group discussion format designed to balance factuality and diversity, addressing issues like monopolization and wrong consensus. 2. Review Mode: A task-agnostic framework[2] that integrates external insights and allows incremental modifications on predecessor efforts, applicable to both NL and code scenarios.  This approach has demonstrated superior results over strong baselines, and it addresses limitations found in methods based purely on self-correction, as evidenced by recent works [3,4,5]. The integration of external insights, coupled with the flexibility of handling multiple contexts, underlines the novelty of authors' design. 3. Retrieve Mode: A strategy that generates and scores candidate answers based on reasoning processes, focusing on reasoning value rather than majority voting. These components collectively enhance the overall reasoning capabilities of LLMs and address limitations found in prior methods.\n[1] Du, Yilun, et al. \"Improving Factuality and Reasoning in Language Models through Multiagent Debate.\" arXiv preprint arXiv:2305.14325 (2023).\n\n[2] Madaan, Aman, et al. \"Self-refine: Iterative refinement with self-feedback.\" arXiv preprint arXiv:2303.17651 (2023).\n\n[3] Large Language Models Cannot Self-Correct Reasoning Yet, arXiv: 2310.01798\n\n[4] Can Large Language Models Really Improve by Self-critiquing Their Own Plans?, arXiv: 2310.08118\n\n[5] GPT-4 Doesn't Know It's Wrong: An Analysis of Iterative Prompting for Reasoning Problems, arXiv: 2310.12397", "category": "Method_Mechanics"}, {"question": "What are the novel aspects of Corex's three modes (Debate, Review, and Retrieve) compared to existing works [1], [2], and [3]?\n[1] Du, Yilun, et al. \"Improving Factuality and Reasoning in Language Models through Multiagent Debate.\" arXiv preprint arXiv:2305.14325 (2023).\n\n[2] Madaan, Aman, et al. \"Self-refine: Iterative refinement with self-feedback.\" arXiv preprint arXiv:2303.17651 (2023).\n\n[3] Yang, Chengrun, et al. \"Large language models as optimizers.\" arXiv preprint arXiv:2309.03409 (2023).", "answer": "Debate Mode introduces a format combining group discussions to balance factuality and diversity, with a distinct approach using collective intelligence and a judging mechanism for enhanced decision-making. Review Mode is task-agnostic, encouraging incremental modifications and considering both NL and code scenarios, showing superior results compared to baselines. Retrieve Mode focuses on generating candidates and scoring them based on the faithfulness of their reasoning processes, prioritizing reasoning chains over majority voting, which is a novel approach in this context.", "category": "Method_Comparison"}, {"question": "What is the explanation for the varying performance improvements across Corex-Debate, Corex-Review-NL, Corex-Review-Code, and Corex-Retrieve, and why does Corex-Review-Code perform better in specific scenarios?", "answer": "The performance improvement varies based on the specific scenarios each algorithm is designed for. Corex-Review-Code benefits from using programming languages (PL), which inherently offer an advantage in mathematical problem-solving. Additionally, Corex helps rectify errors introduced during the NL2Code process, such as misunderstandings of the problem statement or program errors, making the Python solutions more reliable and enhancing outcomes. In the Repeat Copy task, using code to generate strings reduces uncertainty, which further improves Corex-Review-Code's performance compared to other variants.", "category": "Method_Disambiguation"}, {"question": "What explains the notable performance gap of Corex-Review-Code compared to other Corex variants in the GSM-Hard task (Table 2) and the Repeat Copy task (Table 4)?", "answer": "The performance gap is primarily due to the use of NL2Code + Task delegation, which helps alleviate the limitations of large language models (LLMs) in handling large number computations. For example, in the GSM-Hard task, LLMs struggle with directly computing large numbers, but using code ensures accurate calculations. : *James decides to run 1793815 sprints 1793815 times a week. He runs 60 meters each sprint. How many total meters does he run a week?* Large models struggle with directly computing such large numbers. However, with the help of code, authors can ensure accurate computations: ```python def solution(): sprints_per_day = 1793815 days_per_week = 1793815 meters_per_sprint = 60 total_sprints = sprints_per_day * days_per_week total_meters = total_sprints * meters_per_sprint result = total_meters return result ``` Corex-Review-Code enhances the reliability of these Python solutions by rectifying errors introduced during the NL2Code process, as discussed in Section 3.2. Similarly, in the Repeat Copy task, using code to generate strings reduces uncertainty, leading to improved performance compared to other variants.", "category": "Experimental_Analysis"}, {"question": "Did authors explore the impact of using different large language models (LLMs) as agents in the same mode, such as GPT-3.5-turbo, GPT-4, and Claude-Instant-1.2 in Retrieve mode, and how does this affect diversity compared to using the same LLMs?", "answer": "In authors' analysis (section 5.2), authors' discussed the synergies between different models in Debate and Retrieve modes. In Debate, authors used group discussions to avoid the diversity issues seen in previous works where a single model dominated, thus enhancing diversity even with the same LLMs. In Review mode, external insights were used to enhance diversity. For Retrieve mode, the results showed that a mix of stronger and weaker models yielded similar performance to using the same agents, indicating stable performance regardless of output diversity.", "category": "Experimental_Analysis"}, {"question": "Why were open-source models such as Llama not included in the results, and what challenges exist for incorporating them into the Corex framework?", "answer": "Open-source models were not included in the results due to their performance constraints, which pose a lower bound to the competencies necessary for collaboration. Corex is designed to enhance weaker models (e.g., GPT-3.5-Turbo) to surpass stronger ones (e.g., GPT-4), but open-source models face challenges such as weaker coding capabilities and insufficient long-text understanding, limiting their applicability in modes like review-code. However, with advancements in open-source large models, there is potential to incorporate them into the Corex framework in the future.", "category": "Experimental_Analysis"}, {"question": "What evidence or case studies demonstrate that Corex enhances factuality, faithfulness, and reliability?", "answer": "1. Improvement in commonsense reasoning tasks, as shown in experimental results, reflects enhanced factuality and faithfulness. Corex's ability to produce accurate conclusions (e.g., writing reviews, fixing errors in debate) further supports this. 2. Reliability is enhanced through the Review mode, which reduces erroneous reasoning chains and buggy code. For example, in the GSM8k task, the number of errors due to incorrect code decreased from 136 cases (10%) to 74 cases (5.6%) after applying the review mode. While quantitative evaluation is challenging due to the lack of reference sets, direct observation and empirical evidence from evaluations indicate improvement.", "category": "Experimental_Analysis"}, {"question": "Are there quantitative metrics specifically designed to measure faithfulness in reasoning chains used in the evaluation of Corex?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "Oho3UxCkKr", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "SCREWS: A Modular Framework for Reasoning with Revisions", "authors": ["Kumar Shridhar", "Harsh Jhamtani", "Hao Fang", "Benjamin Van Durme", "Jason Eisner", "Patrick Xia"], "abstract": "Large language models (LLMs) can improve their accuracy on various tasks through iteratively refining and revising their output based on feedback. We observe that these *revisions* can introduce errors, in which case it is better to roll back to a previous result. Further, revisions are typically homogeneous: they use the same reasoning method that produced the initial answer, which may not correct errors.\nTo enable exploration in this space, we present SCREWS, a modular framework for reasoning with revisions.\nIt is comprised of three main modules: *Sampling*, *Conditional Resampling*, and *Selection*, each consisting of sub-modules that can be hand-selected per task. We show that SCREWS not only unifies several previous approaches under a common framework, but also reveals several novel strategies for identifying improved reasoning chains. We evaluate our framework with state-of-the-art LLMs (ChatGPT and GPT-4) on a diverse set of reasoning tasks and uncover useful new reasoning strategies for each: arithmetic word problems, multi-hop question answering, and code analysis. \nHeterogeneous revision strategies prove to be important, as does selection between original and revised candidates.", "keywords": "Reasoning;LLMs", "qa_pairs": [{"question": "What are the novel contributions of the paper, specifically in terms of refinement strategies, and how do they address limitations in previous work?", "answer": "The paper introduces two novel components: (1) heterogeneous sampling as an approach to refinement and (2) the option to 'roll back' by choosing between the refined answer and the original prediction in case of errors during refinement. These components are absent in all previous work. Additionally, the proposed framework enables the integration of different strategies to optimize refinement for specific tasks, with a focus on improving reasoning during refinement. This framework, along with achieving 93+ accuracy, represents the main contribution of the work.", "category": "Method_Mechanics"}, {"question": "What are the novel contributions of this paper, specifically in terms of methodological advancements that address the improvement of existing selection methods?", "answer": "The paper introduces two novel components: heterogeneous sampling as an approach to refinement and the ability to choose between the refined answer and the original prediction to 'roll back' in case of errors in refinement. These components are not present in previous work. Additionally, the proposed framework allows for the combination of different strategies to achieve the best refinement strategy for a given task. This framework is a significant contribution, as it enables the integration of various components from previous methods.", "category": "Method_Disambiguation"}, {"question": "How does the conditional resampling in StrategyQA's tool use experiment remain beneficial when the LLM is initially configured to access retrieved facts?", "answer": "When the LLM already knows the facts, it avoids factual errors, achieving 90% accuracy in StrategyQA by providing facts in the first steps, which is higher than refinement. The experiment aims to test the LLM's ability to judge its mistakes and avoid repeated factual errors by providing factual information, thereby distinguishing between errors due to tools and those due to the LLM itself.", "category": "Experimental_Setup"}, {"question": "Is the direct provision of relevant facts to the conditional resampler, without using a Web search tool for fact retrieval, the primary reason for the observed task improvement? Additionally, would the same improvement be observed in a more realistic scenario involving an external tool (considering potential noise and long retrieved passages)?", "answer": "Providing facts directly to the LLM ensures it does not make factual errors, leading to high accuracy (e.g., 90% for StrategyQA). However, the primary focus of this study is to test if LLMs can judge their own mistakes, and providing factual information helps avoid errors. Testing with an external tool was out of scope, as it would complicate the distinction between errors due to the tool and errors due to the LLM.", "category": "Experimental_Analysis"}, {"question": "Given the mixed statistical significance in Tables 1 and 2 and the higher performance of independent sampling in Table 2, does SCREWS demonstrate clear advantages over naive CoT in terms of accuracy and cost-effectiveness, and can ablation studies confirm this?", "answer": "SCREWS is cost-effective as it involves resampling only selected samples, which is less expensive than resampling all samples. Fig. 3a shows that SCREWS achieves higher accuracy (83) with fewer samples (3) compared to CoT (75 accuracy with 5 samples), demonstrating its effectiveness.", "category": "Method_Comparison"}, {"question": "Does the improvement from sampling and conditional resampling approaches (73 to 73.99) in Table 1 significantly close the gap to the state-of-the-art performance on GSM8K, which is 90+ percent?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "AKJLnDgzkm", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Welfare Diplomacy: Benchmarking Language Model Cooperation", "authors": ["Gabriel Mukobi", "Hannah Erlebach", "Niklas Lauffer", "Lewis Hammond", "Alan Chan", "Jesse Clifton"], "abstract": "The growing capabilities and increasingly widespread deployment of AI systems necessitate robust benchmarks for measuring their cooperative capabilities. Unfortunately, most multi-agent benchmarks are either zero-sum or purely cooperative, providing limited opportunities for such measurements. We introduce a general-sum variant of the zero-sum board game Diplomacy—called Welfare Diplomacy—in which players must balance investing in military conquest and domestic welfare. We argue that Welfare Diplomacy facilitates both a clearer assessment of and stronger training incentives for cooperative capabilities. Our contributions are: (1) proposing the Welfare Diplomacy rules and implementing them via an open- source Diplomacy engine; (2) constructing baseline agents using zero-shot prompted language models; and (3) conducting experiments where we find that baselines using state-of-the-art models attain high social welfare but are exploitable. Our work aims to promote societal safety by aiding researchers in developing and assessing multi-agent AI systems. Code to evaluate Welfare Diplomacy and reproduce our experiments is available at https://anonymous.4open.science/r/welfare-diplomacy-72AC.", "keywords": "multiagent systems;cooperative AI;AI agents;language models", "qa_pairs": [{"question": "What does the term 'zero-shot' mean in the context of the paper's discussions on 'zero-shot prompted language model,' 'zero-shot baseline,' and 'zero-shot evaluations'?", "answer": "The term 'zero-shot' refers to scenarios where no examples of gameplay are provided to the model.", "category": "Concept_Understanding"}, {"question": "How does the theoretical analysis in Section 3.2.1 simplify the real WD game, and what are the key differences (gaps) between the theoretical model and the real WD game?", "answer": "The theoretical analysis simplifies the real WD game by using a toy version with fewer states, more symmetries between players, and a focus on a bargaining problem that captures the core challenge of WD. These modifications make the model more tractable to analyze while still addressing the central dynamics of the game.", "category": "Method_Mechanics"}, {"question": "What are the reasons why most models, except advanced LLMs like GPT-4, cannot have policies significantly better than the random baseline?", "answer": "The reasons include qualitative features of LLM policies: Claude-1.2 gets very low root mean Nash welfare because some players hold their units throughout the game, resulting in 0 Welfare Points. GPT-3.5 and Llama-2 tend not to acquire many more SCs but sometimes disband units, acquiring some Welfare Points.", "category": "Experimental_Analysis"}, {"question": "How does the paper's characterization of Standard Diplomacy (SD) as a zero-sum game align with the observation that it involves cooperation among players, as noted in [1]?\n[1] \"Human-level play in the game of Diplomacy by combining language models with strategic reasoning\", Science 2022.", "answer": "The paper's characterization of Standard Diplomacy as a 'zero-sum game' refers to the technical sense where the sum of players' scores remains constant, indicating a single winner. While SD does involve cooperation among subsets of players, it lacks opportunities for global rational cooperation and skilled play is not globally cooperative. ", "category": "Method_Mechanics"}, {"question": "What is the meaning of $\\pi^k$ as a Nash equilibrium, and how does Theorem 1 relate to the main claim of the paper?", "answer": "Theorem 1 demonstrates that in the toy environment, there are Nash equilibria that Pareto-dominate others, and these Pareto-dominating equilibria require more cooperative capabilities. The edited paragraph after Theorem 1 explicitly connects this to the main criterion (A) of the paper, which emphasizes the existence of opportunities for global, rational cooperation. A forward reference to Theorem 1 has also been added in Section 2.2 to further clarify this connection.", "category": "Method_Disambiguation"}, {"question": "What is the justification for constructing the exploiter in the paper, and does there exist an agent with better exploitative power?", "answer": "The exploiters were designed based on the authors' intuitions in playing WD and were not optimized for the best exploitation. The goals were to (i) present a proof of concept for measuring exploitability, a key property alongside social welfare, and (ii) assess the exploitability of the models, for which these exploiters were sufficient. It is acknowledged that better exploiters likely exist.", "category": "Experimental_Analysis"}, {"question": "Is human analysis of LLM's policy included in the paper to understand its performance?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "B6t5wy6g5a", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Aligning Large Multimodal Models with Factually Augmented RLHF", "authors": ["Zhiqing Sun", "Sheng Shen", "Shengcao Cao", "Haotian Liu", "Chunyuan Li", "Yikang Shen", "Chuang Gan", "Liangyan Gui", "Yu-Xiong Wang", "Yiming Yang", "Kurt Keutzer", "Trevor Darrell"], "abstract": "Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in ``hallucination'', generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench\ndataset with the 96% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement of 60% on MMHAL-BENCH over other baselines", "keywords": "AI Alignment;Large Multimodal Models;RLHF", "qa_pairs": [{"question": "What quantitative evaluation is provided to measure the reduction in hallucination, and what are the results?", "answer": "The quantitative evaluation of hallucination reduction is conducted using the proposed MMHAL-BENCH dataset. LLaVA-RLHF improves by 60% over other baselines and reduces the hallucination rate of the original LLaVA by 27%. Detailed results are provided in Table 6, and the dataset is described in Appendix C.", "category": "Experimental_Exposition"}, {"question": "What evidence supports the claim that Fact-RLHF reduces hallucinations, given that LLaVA-RLHF did not significantly differ from LLaVA-SFT in hallucination rates and even performed worse in the 13B model?", "answer": "Fact-RLHF reduces hallucinations by encouraging detailed and informative responses, which are often longer than those from SFT models or human annotators. However, GPT-4's evaluation on MMHal-Bench is biased against RLHF models because it mistakenly flags valid but unannotated information as hallucinations. This bias leads to an overestimation of the hallucination rate for RLHF models. When accounting for this bias, RLHF is shown to effectively reduce hallucinations.", "category": "Experimental_Analysis"}, {"question": "Can authors provide a detailed description of the experimental setup, including the datasets, tasks, and evaluation metrics used in the study?", "answer": "The experimental setup includes three supervised fine-tuning datasets: VQA-v2, AK-VQA, and Flickr30k, as listed in Section 2.2. VQA-v2 uses 'Yes' or 'No' queries (83k), A-OKVQA uses multiple-choice questions (16k), and Flickr30k uses grounded captions (23k). Additionally, 10k human preference data are paired outputs from the base 7B LLaVA model, labeled by Amazon Turker annotators for hallucination levels. The evaluation metrics include accuracy for MMBench (1031 multiple-choice questions), F1 score for POPE (3k 'Yes/No' questions in random, adversarial, and popular categories), GPT4 evaluation for LLaVA bench (100 questions), and GPT4 score for MMHalBench (96 questions targeting hallucination levels). Further details are provided in Appendices G and I.", "category": "Experimental_Setup"}, {"question": "How does the evaluation in MMHal-Bench demonstrate alignment between the benchmark dataset and human evaluations, particularly when scores are adjusted for anti-hallucinations?", "answer": "The authors tested LLaVA-13B and IDEFICS-80B on MMHal-Bench, manually inspecting the answers produced by these models and their corresponding GPT-4 reviews. They found a 94% agreement between human inspection and GPT-4 evaluation in judging hallucinations, even though IDEFICS-80B was not involved in designing the question-answer pairs. This high consistency demonstrates that the automated GPT-4 evaluation in MMHal-Bench is well-aligned with human evaluations.", "category": "Experimental_Exposition"}, {"question": "Is detailed information about the labelers involved in RLHF, including their number and recruitment method, provided in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the process of collecting FACT-RLHF annotations involve 28 labelers and cost approximately $5000?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the questions and answers in the MMHAL-BENCH dataset constructed by automatic methods?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the COCO image caption data used for instruction tuning?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the collection of FACT-RLHF annotations take around one week on the MTurk platform?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the COCO image caption data used only as fact enhancement in the reward model during the RL phase?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "pyW37euNXb", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Upgrading VAE Training With Unlimited Data Plans Provided by Diffusion Models", "authors": ["Tim Z. Xiao", "Johannes Zenn", "Robert Bamler"], "abstract": "Variational autoencoders (VAEs) are popular models for representation learning but their encoders are susceptible to overfitting (Cremer et al., 2018) because they are trained on a finite training set instead of the true (continuous) data distribution $p_{\\mathrm{data}}(\\mathbf{x})$. Diffusion models, on the other hand, avoid this issue by keeping the encoder fixed. This makes their representations less interpretable, but it simplifies training, enabling accurate and continuous approximations of $p_{\\mathrm{data}}(\\mathbf{x})$. In this paper, we show that overfitting encoders in VAEs can be effectively mitigated by training on samples from a pre-trained diffusion model. These results are somewhat unexpected as recent findings (Alemohammad et al., 2023; Shumailov et al., 2023) observe a decay in generative performance when models are trained on data generated by another generative model. We analyze generalization performance, amortization gap, and robustness of VAEs trained with our proposed method on three different data sets. We find improvements in all metrics compared to both normal training and conventional data augmentation methods, and we show that a modest amount of samples from the diffusion model suffices to obtain these gains.", "keywords": "VAE;Diffusion Model;Data Augmentation;Distillation;Generalization;Robustness;Amortized Inference;Generative Model", "qa_pairs": [{"question": "How does the distribution of the ELBO on the test-set compare between the VAE and diffusion approaches?", "answer": "The analysis in Appendix G (new) shows that the method improves the ELBO on almost all (99.9%) of the test points, as illustrated in Figure 9 (b). No subset of test data points was identified where the improvement was particularly small or large.", "category": "Method_Comparison"}, {"question": "How do the experiments in Sections 5.3 and 5.4 address the evaluation of representation learning in VAEs, and what additional evaluations were added to support this claim?", "answer": "The experiments in Section 5.3 evaluate the amortization gap, indicating a good encoder with a small gap, and Section 5.4 evaluates the encoder's adversarial robustness by testing if minor input changes affect the output representation. Additionally, Appendix F includes direct evaluations of representation learning tasks, reconstruction tasks, and data generation tasks, showing that the proposed method performs best in all tasks and metrics except sample quality measured by inception score, supporting the claim of improving representation learning and reconstruction.", "category": "Experimental_Analysis"}, {"question": "What is the computational cost of the proposed method, particularly during training, and how does it compare to the test time cost?", "answer": "The computational cost during training is detailed in Appendix C. The method increases training time but does not affect test time, which is more critical in many applications. Additionally, samples can be reused for cross-validation when training VAEs with different configurations.", "category": "Experimental_Analysis"}, {"question": "What is the significance of the generalization gap in the performance evaluation of the proposed DMaaPx method across different datasets?", "answer": "The generalization gap is significant because it indicates that ELBOs evaluated on the training set can reliably predict the final performance of the VAE. The proposed DMaaPx method has a significantly smaller generalization gap than other methods across all datasets, including MNIST, FMNIST, and CIFAR10.", "category": "Experimental_Analysis"}, {"question": "How do the computational costs of DMaaPx compare with aug tuned and aug naive when aiming to achieve similar performance in terms of test ELBO?", "answer": "The computational overhead of DMaaPx primarily arises from training the diffusion model rather than sampling from it. Sampling less than about 10 times the training data size has little benefit, as sampling more yields diminishing returns. Importantly, DMaaPx increases training time but not test time, which is crucial in many applications. Additionally, samples can be reused for training VAEs with different configurations for cross-validation. Further details are provided in Appendix C and item (2) of the overall response.", "category": "Experimental_Analysis"}, {"question": "Does the small magnitude of improvements in downstream applications of VAE (RL, RC, SQ) compared to normal training affect the significance of the paper?", "answer": "The small magnitude of improvements in downstream tasks does not diminish the significance of the paper, as the primary contribution lies in investigating new ways to address the overfitting problem of VAEs, particularly through changes to the training data rather than model architecture or loss function. The generalization gap shrinks significantly (Figure 2), and the paper aims to inspire further research in cross-model-class distillation for VAEs. Authors clarified in Section 4.1 and in the conclusion that authors make the assumptions that (i) cross-model-class distillation requires a generative model that satisfies the two criteria in Table 1, and that (ii) diffusion models satisfy these two criteria. But authors hope that results inspire research that challenges both assumptions, e.g., to develop new generative models that are specifically designed for cross-model-class distillation for VAEs, as this field is much less explored than distillation within discriminative models.", "category": "Experimental_Analysis"}, {"question": "What specific advantages does the proposed method offer in the context of density estimation and representation learning to justify the computational overhead, given that diffusion models already excel at image generation?", "answer": "The proposed method focuses on improving the encoder of the VAE for density estimation and representation learning, areas where VAEs have advantages over diffusion models. The method increases training time but not test time, which is crucial for many applications. Additionally, the generated samples can be reused for training VAEs with different configurations for cross-validation.", "category": "Method_Disambiguation"}, {"question": "How have you addressed the clarity issue in Figure 2 regarding the difficulty in discerning which method performs best, and why is there a zoom-in on the CIFAR-10 plot?", "answer": "Authors have added annotations to Figures 2, 3, 4, and 6 to highlight the gaps and show their numerical values, making it easier to see that proposed method achieves the smallest gaps except for the robustness gap in BinaryMNIST and FashionMNIST. The reason why authors plot ELBO values rather than gaps in Figures 2, 3, 4, and 6 is because both the ELBO values and the gaps are important in practice, and plotting only the gaps would conceal the absolute values. The zoom-in on the CIFAR-10 plot is included to clearly show the small gap between the two green lines, which might be difficult to discern otherwise.", "category": "Method_Mechanics"}, {"question": "How were the hyperparameters selected for each method, and were experiments conducted with different random seeds to ensure randomness in the results?", "answer": "Hyperparameters for BinaryMNIST and FashionMNIST were selected by consulting the literature, while for CIFAR-10, hyperparameters were manually chosen to induce overfitting, as the study focuses on alleviating overfitting in VAEs. Initially, experiments were conducted with one seed, but CIFAR-10 experiments were re-run with three different random seeds, and the results (means and standard deviations) were updated in Figures 2, 3, and 4. Experiments for other datasets are ongoing and will be updated in the paper.", "category": "Experimental_Setup"}, {"question": "What test metrics or samples are available to compare the performance of authors' VAE method against the baselines?", "answer": "Authors added evaluations of downstream applications in Appendix F, including representation learning tasks, reconstruction tasks, and data generation tasks. authors' method improves VAE performance for all tasks and metrics except sample quality (SQ) measured by inception score (IS), as it primarily fixes the encoder, impacting representation learning and reconstruction but not sample quality.", "category": "Experimental_Exposition"}, {"question": "What theoretical explanation or simplified model is provided to understand why the proposed data augmentation method works?", "answer": "In Section 4.2, it is explained that traditional data augmentation generates new data points by sampling conditioned on a single data point, while the proposed method generates new data points conditioned on the entire training set. For a simplified model, consider a dataset with a mixture of Gaussian distribution. Traditional augmentation adds small noise to each data point, but with knowledge of the distribution, discrete translations can be used. A diffusion model trained on the entire dataset can automatically detect such properties.", "category": "Motivation_Analysis"}, {"question": "How does the use of an unconditional diffusion model, as opposed to a conditional one, differentiate the proposed approach from traditional data augmentation methods?", "answer": "The proposed approach uses an unconditional diffusion model, which samples random data without taking an input, unlike conditional diffusion models that are similar to traditional data augmentation. This distinction is clarified in a footnote in Section 4.1 of the paper.", "category": "Method_Mechanics"}, {"question": "How does the performance of diffusion models in the proposed method compare to their performance in settings where data is scarce or in tabular data settings?", "answer": "The performance of diffusion models in the proposed method is shown to be effective even in settings where data is scarce. While the training sets were too small for fitting good VAEs, they were large enough to fit useful diffusion models. This demonstrates that diffusion models can be useful in scenarios where there is insufficient data to train a good VAE, which is a counterintuitive but significant finding. The fact that this observation is counterintuitive highlights why it should be communicated to a broader audience: one might think diffusion models require lots of data, and this is true if the goal is to generate high-quality samples. ", "category": "Experimental_Analysis"}, {"question": "What are 'amortized inference' and 'adversarial robustness', and how are these concepts utilized in the context of VAEs?", "answer": "Amortized inference is a principle in VAEs where a neural network (the 'inference network') maps data points to latent representations, speeding up training and deployment. This is discussed in the 'Amortization Gap' paragraph in Section 2. Adversarial robustness, as analyzed in the paper, is detailed at the end of Section 2. Removing the inference network and obtaining latent representations by iteratively optimizing the ELBO is an alternative but more expensive method.", "category": "Concept_Understanding"}, {"question": "Under what conditions does the statement $\\mathcal G_\text{g} \\geq 0$ hold, and what could lead to $\\mathcal G_\text{g} < 0$?", "answer": "The statement $\\mathcal G_\text{g} \\geq 0$ holds under the assumption that training and test data come from the same distribution. If this assumption is violated and the test data is simpler (i.e., from a distribution with lower entropy) than the training data, it can lead to $\\mathcal G_\text{g} < 0$.", "category": "Experimental_Exposition"}, {"question": "Is the computational cost of the proposed method, including the impact of sampling on training time, evaluated and documented in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was data augmentation omitted when training the diffusion model to prevent leakage into generated samples?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the added evaluation in Appendix F directly demonstrate the enhancement of VAE's performance by the diffusion model sampler in practical downstream applications?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the proposed method in the paper show the best performance for all tasks and metrics related to representation learning and reconstruction, except for sample quality measured by inception score (IS)?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the proposed method improve the VAE’s performance for all tasks and metrics except sample quality measured by inception score (IS)?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the proposed method increase both training and test time computational costs?", "answer": "False", "category": "Claim_Verification"}, {"question": "Can samples generated for training VAEs be reused across different configurations for cross-validation, as mentioned in the rebuttal?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "8dkp41et6U", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression", "authors": ["Huiqiang Jiang", "Qianhui Wu", "Xufang Luo", "Dongsheng Li", "Chin-Yew Lin", "Yuqing Yang", "Lili Qiu"], "abstract": "In long context scenarios, large language models (LLMs) face three main challenges: higher computational/financial cost, longer latency, and inferior performance. Some studies reveal that the performance of LLMs depends on both the density and the position of the key information (question relevant) in the input prompt. Inspired by these findings, we propose LongLLMLingua for prompt compression towards improving LLMs’ perception of the key information to simultaneously address the three challenges. We conduct evaluation on a wide range of long context scenarios including single-/multi-document QA, few-shot learning, summarization, synthetic tasks, and code completion. and experimental results show that LongLLMLingua compressed prompt can derive higher performance with much less cost. The latency of the end-to-end system is also reduced. For example, on NaturalQuestions benchmark, LongLLMLingua gains a performance boost of up to 17.1% over the original prompt with ∼4x fewer tokens as input to GPT-3.5-Turbo. It can derive cost savings of `$`28.5 and `$`27.4 per 1,000 samples from the LongBench and ZeroScrolls benchmark, respectively. Additionally, when compressing prompts of ∼10k tokens at a compression rate of 2x-10x, LongLLMLingua can speed up the end-to-end latency by 1.4x-3.8x.", "keywords": "Prompt Compression;Long Context;LLMs;Black-box LLMs;Efficient Method", "qa_pairs": [{"question": "What is the relevance and effectiveness of subsequence recovery in generative tasks compared to extractive tasks, and how does its removal impact performance across different tasks?", "answer": "While the benefit of subsequence recovery is more limited in generative tasks compared to extractive tasks, it is still relevant as generated content often refers to entities or phrases in the prompt, and subsequence recovery helps mitigate information loss. An ablation study shows that removing subsequence recovery leads to performance drops in Long Bench tasks such as 'SingleDoc (QA)', 'multiDoc (QA)', 'Summ. (summarization)', and 'FewShot', as demonstrated in Table 6.", "category": "Experimental_Analysis"}, {"question": "How does the LongLLMLingua framework address the challenge of evenly distributed key information in tasks like summarization, and what evidence supports its effectiveness?", "answer": "LongLLMLingua is designed as a coarse-to-fine compression framework that allows adjustment of compression granularity based on the task. For tasks with evenly distributed key information, it performs coarse-grained compression at the paragraph or sentence level and adjusts the proportion of content dropped during coarse-grained and fine-grained compression. For summarization, it assigns a smaller ratio to coarse-grained compression and a higher ratio to fine-grained compression to avoid excessive input dropping. Evidence for its effectiveness is provided in Table 2 ('summ.').", "category": "Method_Mechanics"}, {"question": "How does LongLLMLingua manage to improve end-to-end system latency despite the latency introduced by evaluating perplexities and compressing prompts every time?", "answer": "LongLLMLingua improves end-to-end system latency by using a much smaller language model, Llama-7B, for compression. Although evaluating perplexities introduces some latency, the overall system latency is still reduced, as evidenced by the speed-up from 1.4x (Table 1) to 2.3x (Table 2).", "category": "Method_Mechanics"}, {"question": "How is the iteration defined in the dynamic budget scheduler introduced in Section 4.3, particularly regarding the evaluation of token importance?", "answer": "The iteration in the dynamic budget scheduler involves evaluating token importance at the segment level by iteratively conditioning on all previous compressed segments and calculating a document-level threshold based on Eq. 5.", "category": "Concept_Understanding"}, {"question": "How does the LongLLMLingua framework address the concern that removing documents and reordering may not benefit or may even harm tasks such as long-input summarization, question answering, code understanding, and multi-hop question answering?", "answer": "LongLLMLingua is designed as a general coarse-to-fine compression framework that allows adjustment of compression granularity based on the task. For tasks like long-input summarization, coarse-grained compression is performed at the paragraph or sentence level rather than the document level to improve key information density. Reordering helps LLMs better capture important information, benefiting overall performance. Experimental results in long-input summarization (Table 6, LongBench summ.), single-/multi-doc question answering (Table 6, LongBench; Table 8, LooGLE), code understanding (Table 6, LongBench Code), multi-hop question answering (Table 5, MuSiQue), and Timeline reorder (Table 8, LooGLE)  demonstrate the effectiveness of LongLLMLingua.", "category": "Method_Mechanics"}, {"question": "How does the question-aware module impact performance in tasks like summarization and multi-hop QA, and what evidence supports its relevance in these contexts?", "answer": "The question-aware module, designed to adapt to changes in information density within the prompt, is crucial for tasks involving specific instructions, leading to a performance drop of up to 35.1 points on NaturalQuestions. It also benefits summarization and multi-hop QA tasks. For instance, dropping question-awareness results in a performance drop of ~1-2 points in the LongBench summarization task and ~9 points in the MuSiQue multi-hop QA. This indicates that question-awareness, even in prompts like 'please summarize this article,' provides valuable global context and influences perplexity distribution during compression.", "category": "Method_Mechanics"}, {"question": "What ablation studies were conducted to evaluate the usefulness of the contributions, specifically the reordering of the question and document and the relevance of the restricting prompt, and were these studies extended to multi-document scenarios such as MuSiQue?", "answer": "Authors have included ablation studies on reordering and the restrict prompt across different datasets, including Multi-document QA (Table 3), LongBench (Table 6), and MuSiQue (Table 5). These studies demonstrate that both reordering and the restrict prompt yield improvements for LongLLMLingua across various benchmarks.", "category": "Experimental_Exposition"}, {"question": "What are the results of using GPT2-small as the small LM for pre-computation and compression, and how do they compare to the original prompts?", "answer": "The results with GPT2-small (using GPT2-dolly as the small LM) have been added to Table 3, Table 5, and Table 6. These results show that GPT2-small compressed prompts can achieve comparable or even higher performance than the original prompts.", "category": "Experimental_Exposition"}, {"question": "Can you provide a detailed breakdown of the ZeroSCROLLS test set benchmark and compare the results with the baseline results of the same models without token constraints?", "answer": "A detailed breakdown of ZeroSCROLLS with baseline results is provided in Appendix D.2. Compared to the original results, LongLLMLingua achieves slightly higher numbers at 3x compression and maintains the same level of performance at 6x compression.", "category": "Experimental_Exposition"}, {"question": "Was the experiment in Section 4.1, which shows that tokens with high contrastive perplexities concentrate on the left side of the dashed line, conducted only when the ground-truth document is in the first position? If so, could the observed trend of contrastive perplexity approaching zero as token position increases be attributed to the 'short-term memory' of perplexity?", "answer": "The distribution of contrastive perplexity is not caused by the diminishing effect. A theoretical derivation in Appendix B shows that contrastive perplexity is positively correlated with $p(x^{\\text{que}}|x_i, x_{<i})$, indicating the contribution of each token to the question. The larger the contrastive perplexity, the higher the $p(x^{\\text{que}}|x_i, x_{<i})$, the greater the contribution of the token $x_i$, the higher its importance, and the more it should be retained.Additionally, Figure 5 in Appendix D.1 demonstrates the distribution of document-level contrastive perplexity when the ground-truth document is placed at different positions (1st, 5th, and 15th). This shows that contrastive perplexity effectively captures token-level information related to the question, even when key information is not at the beginning of the prompt.", "category": "Experimental_Analysis"}, {"question": "What is the impact of each module (e.g., question-aware coarse-to-fine compression, dynamic compression ratio, document reordering, and subsequence recovery) on the compression rate and performance of LongLLMLingua, as demonstrated in the ablation studies?", "answer": "The ablation studies in the updated version (Multi-document QA: Table 3; LongBench: Table 6; MuSiQue: Table 5) show that removing any module in LongLLMLingua, such as question-aware coarse-to-fine compression, dynamic compression ratio, document reordering, and subsequence recovery, yields performance improvements for LLMLingua across different benchmarks.", "category": "Method_Mechanics"}, {"question": "Does the reordering method to avoid the lost-in-the-middle issue disturb the timeline between documents, particularly in cases where later sections depend on earlier ones, such as in fiction?", "answer": "No, the reordering method does not disturb the timeline between documents. Empirical experiments show that LLMs can understand texts organized in an unnatural order. For example, in the timeline reorder task from LooGLE (Table 8), LongLLMLingua outperformed the original prompt by 15.9 points, indicating that LLMs can effectively process reordered content.", "category": "Experimental_Analysis"}, {"question": "What is the theoretical explanation for the stability of contrastive perplexity compared to the original perplexity metric, as shown in Figure 3b?", "answer": "The theoretical derivation of contrastive perplexity, provided in Appendix B, shows that it is positively correlated with $p(x^{\\text{que}}|x_i, x_{<i})$, indicating the contribution of each token to the question $x^{\\text{que}}$. A larger contrastive perplexity corresponds to a higher $p(x^{\\text{que}}|x_i, x_{<i})$, signifying greater token importance and retention. Additionally, Figure 5 in Appendix D.1 illustrates that contrastive perplexity effectively captures token-level information related to the question, even when key information is not at the beginning of the prompt.", "category": "Experimental_Analysis"}, {"question": "Why should GPT-3.5-Turbo-0613 and LongChat-13B-16k be included in the benchmark comparison, and what do the experimental results show regarding the performance of compressed prompts?", "answer": "GPT-3.5-Turbo-0613 and LongChat-13B-16k were included to understand the cost implications of compression. Experimental results in Table 2 and Appendix D.5 Table 7 show that compressed prompts outperform original prompts by 4.8 points on LongBench.", "category": "Experimental_Exposition"}, {"question": "What additional data has been included in Tables 1 and 2 to provide a more comprehensive understanding of the framework's effectiveness, particularly regarding the performance of retrieval-based methods?", "answer": "The authors have incorporated end-to-end latency results of different methods on Multi-document QA, LongBench, and ZeroSCROLLS in Tables 1 and 2. These results show that LongLLMLingua has a slightly higher latency compared to retrieval-based methods like BM25 but accelerates the end-to-end system latency by 1.4x-2.3x compared to the original prompts.", "category": "Experimental_Exposition"}, {"question": "Was the latency evaluation conducted multiple times and at different hours to account for potential variance in API call latency?", "answer": "Yes, the latency reported in the paper is the result of multiple tests. Additionally, authors have added end-to-end latency results of different methods on Multi-document QA, LongBench, and ZeroSCROLLS in Table 1 and Table 2 to provide a more comprehensive comparison. These results show that LongLLMLingua can speedup the end-to-end system latency from 1.4x to 2.3x compared to the original prompts.", "category": "Experimental_Exposition"}, {"question": "Were the latency experiments conducted multiple times and at different hours to account for variance in API call latency?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the threshold for $s_i$, used to filter out less important tokens, determined based on the current document budget $\\tau_k$?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "7FHrZuKogW", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Contractive Systems Improve Graph Neural Networks Against Adversarial Attacks", "authors": ["Moshe Eliasof", "Davide Murari", "Ferdia Sherry", "Carola-Bibiane Schönlieb"], "abstract": "Graph Neural Networks (GNNs) have established themselves as a key component in addressing diverse graph-based tasks. Despite their notable successes, GNNs remain susceptible to input perturbations in the form of adversarial attacks. This paper introduces an innovative approach to fortify GNNs against adversarial perturbations through the lens of contractive dynamical systems. Our method introduces graph neural layers based on differential equations with contractive properties, which, as we show, improve the robustness of GNNs. A distinctive feature of the proposed approach is the simultaneous learned evolution of both the node features and the adjacency matrix, yielding an intrinsic enhancement of model robustness to perturbations in the input features and the connectivity of the graph. We mathematically derive the underpinnings of our novel architecture and provide theoretical insights to reason about its expected behavior. We demonstrate the efficacy of our method through numerous real-world benchmarks, reading on par or improved performance compared to existing methods.", "keywords": "Graph Neural Networks;Adversarial Defense;Contractive Systems;Dynamical Systems Inspired Neural Networks", "qa_pairs": [{"question": "What additional information has been added to the paper to make it more self-contained for readers unfamiliar with contractive systems and dynamical systems?", "answer": "An Appendix section titled 'Contractive Systems' has been added, which discusses the continuous definition of contractive systems and provides a few of their properties. This section is referenced from the Preliminaries section in the main text. Additionally, the discretized definition of a node contractive system in Theorem 2 (Equation (10)) and the adjacency matrix contractive system definition in Equation (11) were already included in the original submission.", "category": "Method_Mechanics"}, {"question": "What is the rationale behind assuming a piecewise constant function in Equation (6)?", "answer": "The assumption of a piecewise constant function in Equation (6) is a common practice in neural ODEs, allowing the interpretation of learned neural network layers as the discretization of a continuous ODE. This basic assumption is also made in [1] and other cited papers [R1,R2,R3,R4,R5].", "category": "Motivation_Analysis"}, {"question": "What is the motivation for the construction and use of the gradient operator in Equation (8)?", "answer": "The motivation for constructing the node dynamics in Equation (8) is based on building upon a diffusion-based GNN layer known to be stable and contractive under certain assumptions. This allows for theoretical analysis and explanation of the node feature dynamical system of CSGNN. The gradient operator in Equation (8) relies on definitions from (Chamberlain et al., 2021; Eliasof et al.,2021), which are also used in the reviewer's provided paper [5].", "category": "Motivation_Analysis"}, {"question": "Why is the choice to enforce symmetry on $\tilde{\\mathbf{K}}_l$ not arbitrary?", "answer": "The choice to keep $\tilde{\\mathbf K}_l$ symmetric is based on [R6], allowing the GNN layers to be viewed as a discretization of a gradient flow, which is crucial for maintaining the dynamical system view in CSGNN. ", "category": "Experimental_Analysis"}, {"question": "What is the justification for the intricate design of the adjacency matrix update in Equation (14)?", "answer": "The design of Equation (14) is based on results from [R7], aiming to create an equivariant and symmetry-preserving parameterization of the adjacency matrix. This design is a significant contribution, addressing mathematical and technical challenges in implementing an adjacency matrix dynamical system, as detailed in Section 4.3 of the original submission.", "category": "Motivation_Analysis"}, {"question": "What are the memory and time overheads associated with the complete matrix representation in equation (14) of the CSGNN method?", "answer": "The CSGNN method incurs additional computational overhead due to learning to modify the adjacency matrix in an end-to-end fashion. Theoretically, the added runtime complexity for a CSGNN adjacency matrix learning layer is $O(9 \\cdot n^2)$, where $n$ is the number of nodes. The overall complexity of a CSGNN network with L layers is $O(L(n \\cdot c^2 + m\\cdot c + 9\\cdot n^2))$. Experiments on the Cora dataset with an Nvidia RTX-3090 GPU show the following results: CSGNN has a training time of 9.24 ms, an inference time of 5.96 ms, and memory consumption of 1623 MB. For comparison, Pro-GNN has a training time of 1681.17 ms, an inference time of 1.01 ms, and memory consumption of 1989 MB. CSGNN$_{\textrm{noAdj}}$ (without adjacency matrix update) has a training time of 3.29 ms, an inference time of 1.32 ms, and memory consumption of 891 MB. While CSGNN requires more resources than CSGNN$_{\textrm{noAdj}}$, it offers improved performance.\n\n | Method          | Train time [ms] | Inference time [ms] | Memory consumption [MB] | Accuracy (%) @ 25% Metattack | Accuracy (%) @ 5 nodes nettack |\n |-----------------|-----------------|---------------------|-------------------------|------------------------------|--------------------------------|\n | Pro-GNN         | 1681.17         | 1.01                | 1989                    | 69.72                        | 66.21                          | \n| CSGNN$_{\\textrm{noAdj}}$ | 3.29            | 1.32                | 891                     | 70.25                        | 81.90                           | \n| CSGNN           | 9.24            | 5.96                | 1623                    | 74.46                        | 83.29                          | ", "category": "Experimental_Exposition"}, {"question": "What is the rationale behind primarily considering poisoning attacks in the paper, and why does the paper not focus on injection attacks despite their perceived greater threat?", "answer": "The paper focuses on poisoning attacks because it follows the experimental setting of [R8], a popular paper whose settings have been widely adopted. The theoretical assumptions are motivated by attacks where the nodes do not change between clean and attacked graphs, which is stated in the 'Notations' paragraph in Section 3 ('Preliminaries'). While the paper does not focus on injection attacks, there is no inherent limitation in the theoretical results regarding changing the number of nodes, which could extend to injection attacks. ", "category": "Experimental_Analysis"}, {"question": "Does the suboptimal performance of the model on the Pubmed dataset suggest that the efficacy of the models and theorems is dataset-dependent?", "answer": "The authors do not agree that the model achieves suboptimal results on the Pubmed dataset. While CSGNN does not achieve the highest accuracy on all configurations on Pubmed, no model is perfect across all datasets and configurations. The authors highlight that CSGNN achieves the highest results on Pubmed in Figures 5 and 6 and is among the top 3 models in Table 4. The paper seeks to offer a broad comparison with as many models as possible, which motivated the selection of datasets included in the study.", "category": "Experimental_Analysis"}, {"question": "Why was the evaluation of the model limited to smaller datasets like Cora, Citesser, and Polblogs, and what are the challenges in extending it to larger datasets like the ogbn series?", "answer": "The evaluation included results on the larger dataset Pubmed, as detailed in Appendix I (Appendix J in the updated submission). However, extending the evaluation to even larger datasets is challenging due to the memory constraints imposed by the $O(n^2)$ possible edges and the need to retain gradients for backpropagation. This is consistent with other GNN defense methods that incorporate a learnable adjacency matrix mechanism. For instance, the CSGNN model was successfully run on a GPU with 24GB of memory on Pubmed, but Pro-GNN [R8] encountered out-of-memory errors. The lightweight parameterization of the adjacency matrix update in Equation (14) contributes to this efficiency. Nonetheless, the quadratic complexity in the number of nodes makes scaling to larger datasets a significant challenge. ", "category": "Experimental_Analysis"}, {"question": "How do the contractive properties of the feature/adjacency matrix evolution across model layers demonstrate the robustness of the GNN model to input or structure perturbations?", "answer": "Contractivity reduces the sensitivity of network outputs to input perturbations. Since graph attacks involve perturbations to node features or the adjacency matrix, the CSGNN's contractive system fortifies the model against such attacks, improving robustness. Even without knowledge of the clean inputs, a contractive system ensures that outputs from perturbed inputs remain close to those from clean inputs. This is empirically validated in Section 4 and further discussed in Section 3 and Appendix A.", "category": "Experimental_Analysis"}, {"question": "What are the key differences between the current paper and the NeurIPS 2022 paper 'On the robustness of graph neural diffusion to topology perturbations,' and what additional contributions does the current paper provide?", "answer": "The key difference lies in the underlying method and novelty. The NeurIPS 2022 paper uses only node feature update neural ODEs, while the current paper introduces a coupled neural ODE that learns both node features and adjacency matrix updates. ", "category": "Experimental_Exposition"}, {"question": "What type of attack is used in the experiments, and what are its configurations (inductive, modification/injection, whitebox/blackbox)?", "answer": "The experiments follow the settings in Pro-GNN, which involve inductive poisoning attacks using both white box (metattack, nettack) and black box (random) configurations. ", "category": "Experimental_Setup"}, {"question": "Why are the unit tests from Mujkanovic et al. (2022) considered adaptive attacks in the proposed model, and how does this align with the authors' goal of designing a defensive mechanism?", "answer": "The unit tests from Mujkanovic et al. (2022) are considered adaptive because different attacks can be utilized for different attack strengths, as discussed in Section 5.2. The authors' goal is not to propose new attacks but to design a defensive mechanism, and the unit tests were used to compare with recent baselines on challenging test cases, as shown in Figures 4 and 7. ", "category": "Experimental_Analysis"}, {"question": "Why was GNNGuard not included as a baseline in Table 1, and how does CSGNN compare to GNNGuard in terms of accuracy on Metattack and Nettack with specific perturbation rates and targeted nodes?", "answer": "GNNGuard was not included in Table 1 because the original paper did not report results on Metattack using a variable percentage of attack. To address this, a comparison was provided between GNNGuard and CSGNN on Metattack with a 20% perturbation rate and 5 targeted nodes using Nettack, as reported in GNNGuard, both on Cora and Citeseer. On Cora, CSGNN achieved 77.4% accuracy on Metattack and 83.2% on Nettack, outperforming GNNGuard's 72.2% and 77.5%, respectively. On Citeseer, CSGNN achieved 73.0% accuracy on Metattack and 84.6% on Nettack, compared to GNNGuard's 71.1% and 86.5%. These results highlight the significance of CSGNN, which outperforms GNNGuard in most cases.\n| Cora |              |                                             |                                           |\n|------|--------------|---------------------------------------------|-------------------------------------------|\n|      | Method       | Accuracy on Metattack 20% perturbation rate | Accuracy on Nettack with 5 targeted nodes |\n|      | GNNGuard     | 72.2                                        | 77.5                                      |\n|      | CSGNN (Ours) | 77.4                                   | 83.2                                  |\n\n| Citeseer |              |                                             |                                           |\n|----------|--------------|---------------------------------------------|-------------------------------------------|\n|          | Method       | Accuracy on Metattack 20% perturbation rate | Accuracy on Nettack with 5 targeted nodes |\n|          | GNNGuard     | 71.1                                        | 86.5                                      |\n|          | CSGNN (Ours) | 73.0                                        | 84.6                                      |", "category": "Method_Comparison"}, {"question": "What are the settings of the adversarial attacks used in the experiments, and how do they compare to the attacks used in related studies such as Mujkanovic et al. (2022)?", "answer": "The experimental settings follow those presented in Pro-GNN and Mujkanovic et al. (2022), providing an accurate and comprehensive picture of the current state of GNN defense mechanisms. Recent tests from Mujkanovic et al. (2022), which are stronger, were also included. Under these harder attacks, CSGNN achieves improved results compared to other methods, as shown in Figures 4 and 7.", "category": "Experimental_Setup"}, {"question": "How does the individual contractivity of the feature and adjacency matrix evolution imply robustness, given that $\\Psi_{X_l}^{h_l}$ and $\\Psi_{Y_l}^{h_l}$ are different functions and $\\Psi_{X_l}^{h_l}$ is impacted by $\\mathbf{A}*$?", "answer": "The paper provides a theoretical analysis of the expansivity of the whole network, not just the contractivity of the adjacency matrix updates. In Appendix F, a bound on the Lipschitz constant of the layers given by explicit Euler steps (the map $\\mathcal{D}$) is computed. The Lipschitz constant of $\\mathcal{D}$ can be forced close to 1, for example by penalizing larger steps in the loss function. The paper experiments with a fully data-driven adaptation of the expansivity of the layers to maximize robust accuracy. Additionally, the appendices show that imposing contractivity in the feature updates provides improved or on-par results compared to other methods.", "category": "Experimental_Analysis"}]}
{"id": "K7KQkiHanD", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "One-for-All: Generalized LoRA for Parameter-Efficient Fine-tuning", "authors": ["Arnav Chavan", "Zhuang Liu", "Deepak Gupta", "Eric P. Xing", "Zhiqiang Shen"], "abstract": "We present Generalized LoRA (GLoRA), a flexible approach for universal parameter-efficient fine-tuning tasks. Enhancing Low-Rank Adaptation (LoRA), GLoRA employs a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations, providing more flexibility and capability across diverse tasks and datasets. Moreover, GLoRA facilitates efficient parameter adaptation by employing a scalable, modular, layer-wise structure search that learns individual adapter of each layer. \nOriginating from a unified mathematical formulation, GLoRA exhibits strong transfer learning, few-shot learning and domain generalization abilities, as it adapts to new tasks through not only weights but also additional dimensions like activations. Comprehensive experiments demonstrate that GLoRA outperforms all previous methods in natural, specialized, and structured benchmarks in the vision field, achieving superior accuracy with fewer parameters and computations. Our models on LLaMA-1 and 2 also show considerable enhancements compared to the original LoRA in the language domain. Furthermore, our structural re-parameterization design ensures that GLoRA incurs no extra inference cost, rendering it a practical solution for resource-limited applications.", "keywords": "Generalized Low-Rank Adaptation;Parameter-efficient Fine-tuning;Large VIsion Models;Large Language Models", "qa_pairs": [{"question": "How can GLoRA be integrated with convolutional and normalization layers in a neural network?", "answer": "GLoRA, being an enhanced version of LoRA, can be integrated with convolutional layers by using a search space that includes LoRA, Vector, Constant, and None attributes. For CNN layers, specific weights (*A$_d$* and *A$_u$*) are established, and for different ranks, indexed versions of these weights are used for low-rank convolution operations. For a Vector attribute, *A* (*1 $\times$ o $\times$ 1 $\times$ 1*) is indexed from *A$_d$*, facilitating vector multiplication. Similarly, in the case of a constant attribute *A* (*1 $\times$ 1 $\times$ 1 $\times$ 1*), indexing is applied for scalar multiplication with the original weight matrix from *A$_d$* during computation. This approach aligns seamlessly with established conventions in linear layer implementation. For normalization layers, which have fewer parameters, PEFT (including GLoRA) is typically not employed as the parameter count is negligible compared to convolutional layers.", "category": "Method_Mechanics"}, {"question": "What are the details of the evolutionary search process and the final chosen paradigms for GLoRA, as discussed in the paper?", "answer": "The details of the evolutionary search process are described in Appendix B of the paper. There was no uniform structure observed in the final network across datasets, and an average distribution of the network across all datasets is presented in Figures 3 and 4. The details of the evolutionary search have been moved to the main paper upon the reviewer's recommendation.", "category": "Method_Disambiguation"}, {"question": "What additional details have been included in Table 2 to address the concerns about the mean and variance, the number of learnable parameters, and the LoRA baseline for LLaMA-2?", "answer": "The updated Table 2 now includes the mean and variance of 52.66 and 0.0022, respectively, from three searches with different random seeds on LLaMA-7B using GLoRA. It also includes the number of final subnet parameters for LoRA and GLoRA for LLaMA2-7B, showing that GLoRA has fewer final subnet parameters compared to LoRA. Additionally, the LoRA baseline for LLaMA-2-7B has been provided in the main paper and updated in Table 2.\n| ARC (25-s) | HellaSwag (10-s) | MMLU (5-s) | TruthfulQA (0-s) | Average |\n|:---------:|:---------:|:-------:|:-------------:|:------:|\n| 53.2 | 77.4 | 36.2 | 43.9 | 52.7 |\n| 53.1 | 77.6 | 36.8 | 43.1 | 52.6 |\n| 53.1 | 77.4 | 36.7 | 43.8 | 52.7 |\n\n| Method  | Final #Param | Supernet #Param | Inference #Param | Average Perf. |\n|---------|:----------------------:|:-------------------:|:-----------------------:|:--------------:|\n| LoRA    | 3.1M                 | -             | 0M                    | 49.8         |\n| GLoRA   | **1.8M**                 | 8.6M              | 0M                    | **52.7**         |\n\nLLaMA-1 - [anonymous link](https://drive.google.com/drive/folders/124pePmfTG3HW7n5QXeq2MlsfJ2SuN70E?usp=sharing).\n\nLLaMA-2 - [anonymous link](https://drive.google.com/drive/folders/1-3tUPlLiEB2CJnENnViTuGA1fDKkrlx8?usp=sharing).\n", "category": "Experimental_Exposition"}, {"question": "How does GLoRA's use of an evolutionary search procedure and its lack of a fixed design affect its generality and practicality compared to existing modules?", "answer": "GLoRA's generality does not refer to a single set of adapters that can be universally applied. Instead, it is general in the sense that the combination of supernet training and subnet search results in data-specific architectures that perform well across datasets[1]. The core novelty of GLoRA lies in its unified re-parameterizable formulation (Eq. 10), which implicitly mimics the behavior of many existing works without explicitly defining them. Regarding practicality, GLoRA identifies the optimal combination of support tensors/modules/prompts for each dataset, and the optimized network can be absorbed into the base architecture using structural re-parameterization, maintaining or improving practical throughput compared to baseline methods like LoRA and SSF.\n[1] Neural Prompt Search.", "category": "Method_Mechanics"}, {"question": "What are the actual training time and memory costs of GLoRA, and how do they compare to other adaptation methods?", "answer": "Large search space of GLoRA indeed increases the associated training time compared to the other adaptation methods. However, it is important to note that the overall time spent on the adaptation of models using GLoRA is still substantially very small when compared to fine-tuning the full model. Further, when compared to other adaptation methods, the benefits gained with GLoRA in terms of performance improvement as well as generalization across datasets significantly outweigh the additional time spent in the adaptation process. GLoRA requires an additional 5.6 folds of training time compared to a single run of LoRA, amounting to a total of ~142 minutes for each VTAB-1k task. The GPU memory consumption of GLoRA is 13 GB compared to 9 GB for LoRA. This is primarily because GLoRA requires roughly 5 times more epochs than LoRA for appropriate convergence, and the additional time is spent on the evolutionary search process. Compared to other adaptation methods (LoRA, NOAH, FacT, SSF, and RepAdapter), GLoRA's total training time including architecture search is 6.6x, where x is the training time of LoRA. The combined training time of the 5 best existing methods is more than that of GLoRA, yet GLoRA delivers superior performance. The actual training time and memory costs of GLoRA are detailed in the appendix of the main paper.", "category": "Method_Comparison"}, {"question": "Can GLoRA simulate all PEFT modules, including AdaptFormer and LoRA, and in which layers is GLoRA applied in practice?", "answer": "GLoRA can simulate multiple PEFT modules such as SSF, VPT, LoRA, and RepAdapter by setting specific support tensors to mimic their behavior. However, GLoRA cannot simulate AdaptFormer due to its non-linearity in the parallel block, which inhibits structural re-parameterization. For vision tasks, GLoRA is applied to all linear layers (MLP and attention modules). For NLP tasks, GLoRA is applied only in the multi-head self-attention layers for a fair comparison to LoRA, but it can also be applied to MLP layers, potentially improving performance further[2].\n\n|  Method     |   A    |   B   |   C   |   D   |   E   | \n|--------|:-------:|:-------:|:-------:|:-------:|:--------------:|\n| LoRA | LoRA   | None  | None  | None  | None  |\n| VPT | None   | None  | Vector| None  | None  |\n| SSF | Vector | None  | Vector| Vector| None  |\n|RepAdapter | LoRA   | None  | None  | Vector| None  |\n[2] QLoRA: Efficient Finetuning of Quantized LLMs", "category": "Method_Mechanics"}, {"question": "What are the concrete figures for the memory and training time costs of GLoRA compared to other adaptation methods, and how do these costs impact its practical ramifications?", "answer": "GLoRA requires an additional 5.6 folds of training time compared to LoRA, amounting to ~142 minutes per VTAB-1k task. The GPU memory consumption of GLoRA is 13 GB compared to 9 GB for LoRA. This is primarily due to GLoRA requiring roughly 5 times more epochs than LoRA for convergence and the additional time spent on the evolutionary search process. This extra time results in an average increase of 4.5% accuracy across 19 vision tasks compared to LoRA. Additionally, GLoRA's total training time, including architecture search, is 6.6x compared to LoRA's x, and it delivers superior performance over other methods like NOAH, FacT, SSF, and RepAdapter, which have training times of 6x, x, 1.2x, and 0.9x respectively. The combined training time of the 5 best existing methods is more than that of GLoRA, yet GLoRA outperforms them. The actual training time and memory details are provided in the appendix of the main paper.\n| Method   | Training Time | Inference Time | Natural | Specialized | Structured | Average |\n|----------|:---------:|:-------------:|:------------:|:---------:|:-------:|:-------:|\n| NOAH | 6x | &#8593; | 80.3 | 84.9 | 61.3 | 75.5 | \n| Best-5 | 10.1x | &#8593; | 82.5    | 86.7        | 63.3       | 77.5    |\n| GLoRA | 6.6x | - | 83.6    | 87.0        | 63.3       | **78.0**    |", "category": "Method_Comparison"}, {"question": "How does the implementation of structural re-parameterization using GLoRA address the storage overhead issue for multiple downstream tasks?", "answer": "The implementation stores a single base network and multiple GLoRA support tensor weights for each downstream task. GLoRA support tensors require less than 1% of the storage compared to the base model, and re-parameterization takes negligible time (<5s across 100 trials on Llama-7B), allowing for instant runtime adjustments and significant storage and inference time savings.", "category": "Method_Mechanics"}, {"question": "What is the Vapnik-Chervonenkis Dimension (VC Dimension), and how does it relate to the terms H_i and H_uni introduced in the context of GLoRA with Higher Capacity in section 2.6?", "answer": "The Vapnik-Chervonenkis Dimension (VC Dimension) is a measure of the capacity (complexity, expressiveness) of a set of functions that can be learned by a statistical classification algorithm. It is defined as the cardinality of the largest set of points that the algorithm can shatter. In the context of GLoRA, $\\mathcal{H_i}$ denotes the hypothesis space of a randomly sampled subnet, and $\\mathcal{H}_\\mathrm{uni}$ denotes the hypothesis space of the complete supernet. The VC Dimension provides an estimate of how capable the model is of fitting complex datasets, with a very high VC Dimension potentially indicating overfitting and poor generalization.\n**Definition of VC Dimension.** The VC Dimension of a hypothesis class $\\mathcal H$ (a set of functions) is the largest number of points that can be shattered by $\\mathcal H$. A set of points is said to be shattered by $\\mathcal H$ if, for every possible labeling (binary classification) of these points, there exists a hypothesis in $\\mathcal H$ that perfectly classifies the points according to that labeling. \nMathematically, if authors have a set of points $S=\\{x_1, x_2, \\ldots, x_d\\}$, the hypothesis class $\\mathcal H$ shatters $S$ if:\n\\begin{equation}\n    \\forall y \\in\\{0,1\\}^d, \\exists h \\in \\mathcal H: \\forall i \\in\\{1,2, \\ldots, d\\}, h\\left(x_i\\right)=y_i\n\\end{equation}\n\nThe VC Dimension, denoted as $\\mathbf d_\\mathrm{vc}(\\mathcal H)$, is the maximum size of any set $S$ that can be shattered by $\\mathcal H$. If $\\mathcal H$ can shatter a set of size $d$ but cannot shatter any set of size $d+1$, then $\\mathbf d_\\mathrm{vc}(\\mathcal H)=d$. In the context of GLoRA, $\\mathcal{H_i}$ denotes the hypothesis space of a randomly sampled subnet and $\\mathcal{H}_\\mathrm{uni}$ denotes the hypothesis space of the complete supernet.", "category": "Concept_Understanding"}, {"question": "How are existing PEFT methods integrated into the GLoRA framework, and what are the specifications of the A/B/C/D/E support tensors in Eq. (10)?", "answer": "A table has been added to the paper showing how existing methods can be integrated into the GLoRA framework. The support tensors (A/B/C/D/E) are set to choices presented in the search space (Eq. (11)) to approximately mimic the behavior of existing methods. For example, LoRA uses only tensor A, VPT uses tensor C, SSF uses tensors A, C, and D, and RepAdapter uses tensors A and D.\n|  Method     |   A    |   B   |   C   |   D   |   E   | \n|--------|:-------:|:-------:|:-------:|:-------:|:--------------:|\n| LoRA | LoRA   | None  | None  | None  | None  |\n| VPT | None   | None  | Vector| None  | None  |\n| SSF | Vector | None  | Vector| Vector| None  |\n|RepAdapter | LoRA   | None  | None  | Vector| None  |", "category": "Method_Mechanics"}, {"question": "What is the training time required for GLoRA (including supernet training and subnet searching), and how does it compare to other adaptation methods in terms of training efficiency?", "answer": " When compared to other adaptation methods, the benefits gained with GLoRA in terms of performance improvement as well as generalization across datasets significantly outweigh the additional time spent in the adaptation process.\nGLoRA requires an additional 5.6 folds of training time compared to a single run of LoRA, amounting to a total of ~142 minutes for each VTAB-1k task. The GPU memory consumption of GLoRA is 13 GB compared to 9 GB for LoRA. GLoRA requires roughly 5 times more epochs than LoRA for appropriate convergence, and the additional time is spent on the evolutionary search process. Compared to other methods, GLoRA's total training time, including architecture search, is 6.6x (where x is the training time of LoRA). The combined training time of the 5 best-performing existing methods (NOAH, FacT, SSF, RepAdapter, and LoRA) is 10.1x, which is more than GLoRA's 6.6x, yet GLoRA delivers superior performance. Additionally, GLoRA requires minimal hyperparameter search compared to other methods, which often require thorough data-specific hyperparameter search, making GLoRA's relative training time significantly lower when this aspect is considered.\n| Method   | Training Time | Inference Time | Natural | Specialized | Structured | Average |\n|----------|:---------:|:-------------:|:------------:|:---------:|:-------:|:-------:|\n| NOAH | 6x | &#8593; | 80.3 | 84.9 | 61.3 | 75.5 | \n| Best-5 | 10.1x | &#8593; | 82.5    | 86.7        | 63.3       | 77.5    |\n| GLoRA | 6.6x | - | 83.6    | 87.0        | 63.3       | **78.0**    |", "category": "Method_Comparison"}, {"question": "What is the impact on performance and efficiency if the weight/bias scaling terms W_0 x A and D x b_0 in Eq. (10) are omitted, and which of the five tensors (A, B, C, D, E) are essential for optimal performance and efficiency in the GLoRA architecture?", "answer": "Omitting any of the support tensors (A, B, C, D, E) in the GLoRA architecture results in sub-optimal performance, as evidenced by a performance dip on the CIFAR dataset. Each tensor plays a crucial role: A scales the weight, B scales the input and shifts the weight, C serves as a layer-wise prompt, and D and E scale and shift the bias. In terms of efficiency, GLoRA has zero computational overhead during inference. For training efficiency, LoRA is the most computationally expensive due to matrix multiplication, followed by Vector due to dot product operations, and Scalar is the most efficient due to simple scalar multiplication. D and E are the most efficient support tensors, while C is the least efficient due to the absence of scalar multiplication.\n| Settings | Remove A and D | Remove C | Keep All |\n|--------------|:---------:|:------:|:------:|\n| **Accuracy** | 76.0    | 75.9 | 76.5 |", "category": "Experimental_Analysis"}, {"question": "Did the experiments on LLaMA-7B using GLoRA show consistent performance with respect to initialization, as indicated by the mean and variance?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does GLoRA introduce any additional inference cost compared to RepVGG and LoRA?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the number of learnable parameters for LoRA and GLoRA on LLaMA2-7B reported in the updated Table 2?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are anonymous links to the LLaMA-1 and LLaMA-2 models fine-tuned using GLoRA provided for further verification?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does Table 2 in the paper include the mean and variance of the LLaMA-7B experiments using GLoRA on the shareGPT dataset?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does GLoRA's total training time including architecture search amount to 6.6x the training time of LoRA?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "rp5vfyp5Np", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "BATTLE: Towards Behavior-oriented Adversarial Attacks against Deep Reinforcement Learning", "authors": ["Fengshuo Bai", "Runze Liu", "Yali Du", "Ying Wen", "Yaodong Yang"], "abstract": "Evaluating the performance of deep reinforcement learning (DRL) agents under adversarial attacks that aim to induce specific behaviors, i.e., behavior-oriented adversarial attacks, is crucial for understanding the robustness of DRL agents. Prior research primarily focuses on directing agents towards pre-determined states or policies, lacking generality and flexibility. Therefore, it is important to devise universal attacks that target inducing specific behaviors in a victim. In this work, we propose BATTLE, a universal behavior-oriented adversarial attack method. In BATTLE, an intention policy is trained to align with human preferences for flexible behavior orientation, while the adversary is trained to guide the victim policy to imitate the intention policy. To improve the attack performance, we introduce a weighting function that assigns importance weights over each state. Our empirical results over several manipulation tasks of Meta-world show the superiority of BATTLE in behavior-oriented adversarial attack settings, outperforming current adversarial attack algorithms. Furthermore, we also demonstrate that BATTLE can improve the robustness of agents under strong attacks by training with adversary. Lastly, we showcase the strong behavior-inducing capability of BATTLE by guiding Decision Transformer agents to act in line with human preferences across various MuJoCo tasks. Our videos are available in https://sites.google.com/view/jj9uxjgmba5lr3g.", "keywords": "deep reinforcement learning;preference-based reinforcement learning;adversarial reinforcement learning", "qa_pairs": [{"question": "Why does BATTLE train an intention policy with PbRL instead of directly training the adversarial policy with PbRL?", "answer": "Directly training the adversarial policy using PbRL, as shown in Table 2 under 'BATTLE w/o $\\pi_\\theta$', was less effective. The intention policy provides step-by-step guidance that enhances the efficiency and effectiveness of the adversarial policy development process.", "category": "Motivation_Analysis"}, {"question": "What is the process for generating behavior sequences used in human preference labeling?", "answer": "The process involves collecting pairs of victim trajectories under adversary attacks, denoted as $(\\sigma_v^0, \\sigma_v^1)$, which correspond to two adversary-generated trajectories $(\\sigma_a^0, \\sigma_a^1)$. When a human prefers $\\sigma_v^0$ over $\\sigma_v^1$, it implies that the attack in $\\sigma_a^0$ is more effective or aligns better with human intent, leading to $\\sigma_a^0 \\succ \\sigma_a^1$. This preference data is used to learn an adversary that aligns with human intentions.", "category": "Method_Mechanics"}, {"question": "How is the attack success rate defined in the experiments, and how does it address potential biases arising from the approximation of human intention by the intention policy?", "answer": "The attack success rate is defined by evaluating whether the victim performs behaviors as intended by humans, such as opening a window if that is the human intention. The success metric is adopted from Meta-world [1], based on the distance between the task-relevant object and its final goal pose. ", "category": "Experimental_Setup"}, {"question": "Why did the authors modify the PA-AD baseline by replacing the attacker's reward with the learned reward, and how does this modification align with the original PA-AD method?", "answer": "The authors modified the PA-AD baseline by replacing the attacker's reward with a reward model learned from human preferences, which allows for the expression of various intentions and enables universal adversarial attacks. This modification differs from the original PA-AD method, where the adversary's reward model is the negative of the victim's reward function. The authors also provided oracle versions of SA-RL and PA-AD using ground-truth rewards, but PA-AD did not show significant improvement, possibly due to its design, which performs optimally with a fixed reward model and access to the victim's gradients.", "category": "Motivation_Analysis"}, {"question": "Can you clarify the definitions of 'intention policy', 'inner and outer level loss', and 'success rate' in the paper?", "answer": "(a) The intention policy ensures the victim's behavior under adversarial attacks aligns with the policy trained using the PbRL framework. (b) The inner and outer losses are defined in Equations (6) and (7) of the paper. (c) The success rate evaluates whether the victim performs behaviors as intended by humans, such as opening a window if that is the human intention. The success metric is adopted from Meta-world, based on the distance between the task-relevant object and its final goal pose. ", "category": "Concept_Understanding"}, {"question": "What is the rationale for using an intention policy instead of a direct attack in the first paragraph of section 4.1?", "answer": "The results in Table 2, particularly the column 'BATTLE w/o $\\pi_\\theta$', indicate that directly training the adversarial policy using PbRL is less effective. The intention policy provides step-by-step guidance, significantly improving the efficiency and effectiveness of the adversarial policy development process.", "category": "Motivation_Analysis"}, {"question": "What is the key difference between SA-RL and BATTLE in terms of how the adversarial policy is trained, and why does BATTLE perform better?", "answer": "The key difference lies in how the adversarial policy is trained. BATTLE's superior performance is largely due to its bi-level optimization framework, which integrates the intention policy and a weighting function.", "category": "Method_Comparison"}, {"question": "What are the potential real-world applications of the proposed method, and what challenges might arise in these contexts?", "answer": "One of the practical applications of the proposed method is in adversarial training, where it significantly enhances the robustness of RL agents, outperforming methods like SA-RL and PA-AD. This application is particularly relevant in areas where robustness against adversarial attacks is critical.", "category": "Method_Disambiguation"}, {"question": "How does the BATTLE-ATLA method perform against targeted attacks and various defense methods, as detailed in Table 1?", "answer": "The BATTLE-ATLA method demonstrated the most robust performance under three different types of attacks, effectively challenging well-defended agents. It also consistently outperformed other strategies when various defense methods were applied, showcasing its strong capability in executing universal attacks against robust models.", "category": "Method_Comparison"}, {"question": "What is the extent of annotation required for the manipulation and opposite behaviors scenarios, and how does it relate to the complexity of the tasks?", "answer": "In the manipulation scenario, approximately 9000 labels were used across all tasks. For the opposite behaviors scenario, the number of labels varied: 1000 for Window Close, 3000 for Drawer Close, 5000 each for Faucet Open, Faucet Close, and Window Open, and 7000 each for Drawer Open, Door Lock, and Door Unlock. BATTLE represents an advancement over previous work as it does not require knowledge of the victim's reward function or access to the victim's gradients, making labeling more feasible and practical.", "category": "Experimental_Analysis"}, {"question": "What are the details of the training process for the victim policy approximator, including the algorithm used, the neural network architecture, and the amount of data required for effective training?", "answer": "The victim models are trained on the Meta-world tasks using the Soft Actor-Critic (SAC)[1] algorithm. Fully connected neural networks are employed for both the policy and Q function. The detailed hyperparameters are provided in Table 5. Each agent is trained for 1 million time steps, which achieves around a 100% success rate on the original tasks.\n[1] Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. ICML, 2018.", "category": "Experimental_Setup"}, {"question": "How does the BATTLE method address the concern of limited practical applicability in robotics due to its perceived nature as a white-box attack?", "answer": "BATTLE is designed as a black-box attack method, which does not require access to the gradients of the victim model. It operates by observing the external outputs of the system, enhancing its applicability in real-world scenarios where internal access to a robotic platform is not feasible.", "category": "Method_Mechanics"}, {"question": "Is BATTLE designed to operate by observing external outputs, enhancing its applicability in real-world robotics scenarios?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does BATTLE require access to the victim's gradients for its operation?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is BATTLE a white-box attack method that applies perturbations to states, limiting its practical applicability in robotics?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does BATTLE require knowledge of the victim's reward function for annotation?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does BATTLE require access to the gradients of the victim model to operate?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "BWSTBrmRqD", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "DOMINO: A Dual-System for Multi-step Visual Language Reasoning", "authors": ["PeiFeng Wang", "Olga Golovneva", "Armen Aghajanyan", "Xiang Ren", "Muhao Chen", "Asli Celikyilmaz", "Maryam Fazel-Zarandi"], "abstract": "Visual language reasoning requires a system to extract text or numbers from information-dense images like charts or plots and perform logical or arithmetic reasoning to arrive at an answer. To tackle this task, existing work relies on either (1) an end-to-end vision-language model trained on a large amount of data, or (2) a two-stage pipeline where a captioning model converts the image into text that is further read by another large language model to deduce the answer. However, the former approach forces the model to answer a complex question with one single step, and the latter approach is prone to inaccurate or distracting information in the converted text that can confuse the language model. In this work, we propose a dual-system for multi-step multimodal reasoning, which consists of a\n\"System-1\" step for visual information extraction and a \"System-2\" step for deliberate reasoning. Given an input, System-2 breaks down the question into atomic sub-steps, each guiding System-1 to extract the information required for reasoning from the image. Experiments on chart and plot datasets show that our method with a pre-trained System-2 module performs competitively compared to prior work on in- and out-of-distribution data. By fine-tuning the System-2 module (LLaMA-2 70B) on only a small amount of data on multi-step reasoning, the accuracy of our method is further improved and surpasses the best fully-supervised end-to-end approach by 5.7% and a pipeline approach with FlanPaLM (540B) by 7.5% on a challenging dataset with human-authored questions.", "keywords": "multi-modal reasoning;visual language reasoning;chart question answering", "qa_pairs": [{"question": "Why is the correct answer for arithmetic in Table 7 a negative value (-50752953286.0)?", "answer": "The correct answer is from the PlotQA dataset, where question-answer pairs are generated by templates automatically. PlotQA defines the 'difference' between A and B as $A-B$ rather than the absolute difference, which results in the negative value in Table 7.", "category": "Experimental_Analysis"}, {"question": "What is the structure of the evaluation prompt when the reasoning process begins directly with task decomposition, bypassing the `Describe` step?", "answer": "The evaluation prompt starts with a direct task decomposition. For instance, it might ask a specific question like 'In which year is the private health expenditure per person in Oman 210.69?' and then provide a step-by-step reasoning process to arrive at the answer, as demonstrated in the example where the data is extracted and analyzed to conclude that the value 210.69 corresponds to the year 2010. The data is 183.88 in 2008, 233.80 in 2009, 210.69 in 2010, 195.26 in 2011, 196.32 in 2012, 154.21 in 2013, 153.22 in 2014. This prompt is detailed in Appendix A.2.", "category": "Experimental_Exposition"}, {"question": "What are the examples of the model failing to decompose the problem correctly, and what are the consequences of these failures?", "answer": "The model fails to decompose the problem correctly in cases where there is ambiguity in the question. In some instances, the model does not request the necessary information but still provides a correct answer. For example, when asked 'How many people use daily?', the model describes the figure but does not directly request the required data, yet it correctly identifies the value as 18.2. In other cases, incorrect decomposition leads to invalid queries for the vision module, resulting in wrong intermediate results and ultimately incorrect answers. For instance, when asked 'What's the percentage of people who don't believe it can impact us and don't know much about covid?', the model incorrectly identifies the percentage as 1.0 due to an invalid query. ", "category": "Experimental_Analysis"}, {"question": "How does the inference efficiency of DOMINO compare to the baseline method, specifically the few-shot DePlot model and the supervised end-to-end method?", "answer": "DOMINO is more efficient than the few-shot DePlot model because (1) its vision module only generates required information from the image, unlike DePlot, which generates the entire table, and (2) DOMINO avoids consuming a large part of the context window in the language model by not processing the whole table sequence. The supervised end-to-end method is more efficient than both DOMINO and DePlot but lacks interpretability and is prone to errors for complex questions. DOMINO also improves efficiency by caching previous hidden states during intermediate results processing and prevents hallucination by offloading visual information extraction to a vision module.", "category": "Method_Comparison"}, {"question": "When the parameter size of the end-to-end baselines is expanded to the same scale as DOMINO, can DOMINO still outperform these baselines in terms of performance and reasoning quality?", "answer": "Yes, DOMINO can still outperform the baselines because it prevents the LLM from hallucinating visual information by offloading visual information extraction to a vision module, ensuring a grounded reasoning process. Additionally, DOMINO caches previous hidden states (keys and values) to avoid redundant forward passes in the LLM, enhancing efficiency.", "category": "Method_Comparison"}, {"question": "How does DOMINO differ from prior works such as 'Visual Programming', 'Socratic Models', and 'Look, Remember and Reason' in terms of the interaction between the language and vision modules?", "answer": "DOMINO differs from these works in how the interactions between the language and vision modules are decided. Both 'Socratic Models' and 'Look, Remember and Reason' pre-define the interactions with templates, which do not generalize to questions with new reasoning processes. 'Visual Programming' decides the whole interaction process by generating a program, but the LM does not get to react differently based on the intermediate results returned from the vision module, which limits the question types that could be solved. By contrast, DOMINO learns to compose the atomic operations on the fly based on both the question and the intermediate results from the vision module, which is more flexible.", "category": "Method_Comparison"}, {"question": "What is the reason for the significantly worse performance on PlotQA compared to ChartQA, given that both datasets involve template-based reasoning?", "answer": "The charts in PlotQA involve extremely large numbers ranging from 0 to 3:50e+15, which pose a challenge to reasoning for both the vision and language models. The fully-supervised method leverages the sufficiently large training set (with over 20M examples) to learn the data bias, as noted in the DePlot paper (Liu et al., 2023).", "category": "Experimental_Analysis"}, {"question": "Why was SciCap not included as a testbed for evaluating the performance of the proposed approach, despite its real-world data and high-level reasoning requirements?", "answer": "SciCap was not included as a testbed because it is a caption generation task, which does not align with the focus of this work on visual language reasoning. Visual language reasoning requires the language model to rely on visual information to answer questions about charts and conduct multi-step reasoning. In SciCap, some information in the target captions is not grounded by the images, making it less suitable for this specific problem. However, SciCap is considered a potentially good resource for training the vision module to conduct the Describe operation.", "category": "Motivation_Analysis"}, {"question": "How does DOMINO address concerns about efficiency given the multiple calls to different LLMs, and how does it compare to the efficiency of the few-shot DePlot model?", "answer": "DOMINO addresses efficiency concerns by caching the previous hidden states (keys and values) during forward passing in the LLM, eliminating the need to start from scratch for intermediate results. Additionally, DOMINO is more efficient than the few-shot DePlot model because (1) its vision module generates only the required information from the image, unlike DePlot's vision module, which generates the entire table, and (2) DOMINO does not need to input the entire table sequence, saving context window space in the LM.", "category": "Method_Mechanics"}, {"question": "How does the data efficiency of the proposed method compare to fine-tuning Deplot, and what are the performance results on PlotQA-V2?", "answer": "The proposed method is more data-efficient than fine-tuning Deplot, as demonstrated in Figure 3. With 20-100 training examples on multi-step reasoning, fine-tuned DOMINO generally outperforms fine-tuned Deplot. While MATCHA achieves 90.7% accuracy on PlotQA-V2 using over 20M examples, the proposed method achieves the second-best result (80.7%) with only 100 training examples.", "category": "Method_Comparison"}, {"question": "Whether the chart types included in PlotQA are bar charts, line charts, and point line charts", "answer": "True", "category": "Claim_Verification"}, {"question": "Will one qualitative example from PlotQA be moved to the main paper in the updated draft?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the associated charts for the examples in Table 7 available in the updated draft of the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the main paper include qualitative examples from PlotQA in addition to the single example from ChartQA?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the DOMINO version of LLaMA include more information about the chart compared to the few-shot DePlot versions of GPT-4 and LLaMA?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "SXMTK2eltf", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "GPT-Driver: Learning to Drive with GPT", "authors": ["Jiageng Mao", "Hang Zhao", "Yue Wang"], "abstract": "We present a simple yet effective approach that can transform the OpenAI GPT-3.5 model into a reliable motion planner for autonomous vehicles. Motion planning is a core challenge in autonomous driving, aiming to plan a driving trajectory that is safe and comfortable. Existing motion planners predominantly leverage heuristic methods to forecast driving trajectories, yet these approaches demonstrate insufficient generalization capabilities in the face of novel and unseen driving scenarios. In this paper, we propose a novel approach to motion planning that capitalizes on the strong reasoning capabilities and generalization potential inherent to Large Language Models (LLMs). The fundamental insight of our approach is the reformulation of motion planning as a language modeling problem, a perspective not previously explored. Specifically, we represent the planner inputs and outputs as language tokens, and leverage the LLM to generate driving trajectories through a language description of coordinate positions. Furthermore, we propose a novel prompting-reasoning-finetuning strategy to stimulate the numerical reasoning potential of the LLM. With this strategy, the LLM can describe highly precise trajectory coordinates and also its internal decision-making process in natural language. We evaluate our approach on the large-scale nuScenes dataset, and extensive experiments substantiate the effectiveness, generalization ability, and interpretability of our GPT-based motion planner. Code will be released upon acceptance.", "keywords": "Motion Planning;Autonomous Driving;Large Language Models (LLMs);GPT", "qa_pairs": [{"question": "What are the results of the ablation study evaluating the impact of removing chain-of-thought reasoning on the performance of the LLM in terms of L2 and collision metrics?", "answer": "The ablation study results show that removing chain-of-thought reasoning slightly harms the L2 metric (0.48 with reasoning vs. 0.49 without reasoning) but performs slightly better in the collision metric (0.35 with reasoning vs. 0.34 without reasoning). The authors interpret this as an insignificant synergy, suggesting that the LLM does not benefit from joint language modeling of reasoning and trajectory. They propose a cascaded design where the LLM first outputs reasoning and then uses reasoning to output trajectory, which works better.\nMethod | L2   |      |      |      | Collision |      |      |      |\n|------|------|------|------|------|-----------|------|------|------|\nGPT-Driver  | 1s   | 2s   | 3s   | Avg. | 1s        | 2s   | 3s   | Avg. |\nw/ cot. reason | 0.21 | 0.43 | 0.79 | 0.48 | 0.16      | 0.27 | 0.63 | 0.35 |\nw/o cot. reason | 0.22 | 0.45 | 0.82 | 0.49 | 0.16      | 0.25 | 0.63 | 0.34 |", "category": "Experimental_Exposition"}, {"question": "How is the hypothetical ego trajectory generated, and what criteria are used to identify critical objects based on this trajectory?", "answer": "The hypothetical ego trajectory is generated by calculating future waypoint locations using the current velocity and acceleration, assuming the vehicle maintains its speed without interference. Objects within a 5x3 m region of any waypoint are identified as critical, representing those that would affect driving without special actions.", "category": "Method_Mechanics"}, {"question": "What evidence demonstrates the robustness of GPT-Driver to errors in the perception and prediction module?", "answer": "The robustness of GPT-Driver to perception and prediction errors is demonstrated in Table 5, where the planning performance only drops slightly when replacing perfect perception and prediction with learned ones. This is discussed in Section 4.2. Additionally, the authors are working on ablating more perception and prediction approaches.", "category": "Experimental_Exposition"}, {"question": "How does the proposed method differ from other approaches like UniAD in terms of inputs, and why is it less intricate and time-consuming?", "answer": "Other approaches, such as UniAD, require compact maps (H x W grids) and occupancy grids (H x W x T grids) in addition to perception and prediction, which makes their systems intricate and time-consuming. In contrast, the proposed method only uses text-based perception and prediction from UniAD as inputs to GPT-Driver, avoiding the need for compact maps and occupancy grids. This makes the proposed method less intricate and time-consuming.", "category": "Method_Comparison"}, {"question": "How is the motion planning performance of GPT-Driver evaluated, and what metrics are used to assess its stability, reliability, and ability to account for multi-modality in driving behavior?", "answer": "The motion planning performance of GPT-Driver is evaluated using L2 and collision rate metrics. L2 reflects the proximity of a planned trajectory to a human driving trajectory, indicating the imitation learning ability of motion planners. The collision rate serves as a strong indicator of the safety of a planned trajectory. The outputs of LLMs are stable, with very few invalid formats over the validation set. Multi-modality in driving behavior is accounted for by adjusting the temperature of LLMs and performing multiple runs. The approach surpasses state-of-the-art methods in L2 and performs on par with baseline methods in collision rate, without relying on post-optimization tricks.", "category": "Experimental_Setup"}, {"question": "What are the limitations of using nuScenes as a comprehensive benchmark for autonomous driving (AD) planning, and how have the authors addressed these limitations in their paper?", "answer": "The authors acknowledge that nuScenes serves as a mature open-loop planning benchmark but has limitations as a comprehensive benchmark for AD planning due to the open-loop evaluation scheme. They address this by interpreting the open-loop evaluation results, highlighting that their approach surpasses state-of-the-art methods in L2 (indicating stronger imitation learning ability) and performs on par in collision rate (indicating safety). Please note that other baseline methods heavily rely on tricks such as post-optimization to lower the collision rate. They also discuss the limitations of open-loop motion planning in Section 4.6 of the paper.", "category": "Method_Mechanics"}, {"question": "What were the results of the experiment conducted to assess GPT-Driver's planning performance when ego-states were disabled, and how was the dataset prepared for this experiment?", "answer": "The experiment showed that GPT-Driver still exhibited strong planning performance with ego-states disabled. The results were: L2 errors at 1s, 2s, and 3s were 0.32, 0.67, and 1.15 respectively, with an average of 0.71; collision rates at 1s, 2s, and 3s were 0.18, 0.44, and 1.11 respectively, with an average of 0.57. Due to time constraints, GPT-Driver was fine-tuned on only 10% of the data.\n| L2   |      |      |      | Collision |      |      |      |\n|------|------|------|------|------|------|------|------|\n| 1s   | 2s   | 3s   | Avg. | 1s  | 2s   | 3s   | Avg. |\n| 0.32 | 0.67 | 1.15 | 0.71 | 0.18  | 0.44 | 1.11 | 0.57 |\n\nGPT-Driver still has very strong planning performance removing ego-states.", "category": "Experimental_Exposition"}, {"question": "Why is the benchmark comparison between the proposed motion planning approach and end-to-end computer vision approaches from CVPR considered fair, given that the proposed approach assumes object detections are provided and focuses solely on planning?", "answer": "The benchmark comparison is fair because: (1) The end-to-end driving approaches from CVPR are state-of-the-art on the nuScenes motion planning benchmark. (2) The proposed approach uses the same detection outputs as UniAD, ensuring a fair comparison from the perspective of motion planners. (3) Planning is the main focus and ultimate optimization goal of these end-to-end driving methods, as evidenced by the title of UniAD ('Planning-oriented Autonomous Driving') and the optimization of detection and prediction modules for better planning. (4) While conventional planning baselines would be beneficial, there are no published conventional planning methods on the nuScenes benchmark.", "category": "Method_Comparison"}, {"question": "What is the distinction between 'rule-based' and 'optimization-based' planning approaches in the context of autonomous driving, and why are optimization-based control methods considered out of scope for this paper?", "answer": "Learning-based and rule-based planning are two mainstream motion planning approaches in autonomous driving. Optimization-based planning approaches, which the reviewer refers to, are actually control methods that output low-level control signals to trace trajectories and avoid collisions. These control methods are considered a downstream task to motion planning and are out of the scope of this paper, as the focus is on higher-level planning methods such as imitation learning and reinforcement learning, which are already covered in the provided survey paper.", "category": "Method_Comparison"}, {"question": "Do all imitation learning (IL) approaches use absolute value loss, and why is this specific choice significant in the context of motion planning?", "answer": "State-of-the-art methods like UniAD and ST-P3 on nuScenes rely on regression of real-world coordinates. Authors' work highlights the essential difference between regression and language modeling in imitation learning for motion planning. Authors' language modeling approach is simpler to implement and shows promising results compared to regression-based methods, which authors consider a significant contribution.", "category": "Method_Disambiguation"}, {"question": "What datasets were used for fine-tuning and evaluation of the GPT-Driver model, and how does the approach address the multi-modal nature of trajectory prediction in the nuScenes benchmark?", "answer": "The training set of nuScenes was used for fine-tuning the GPT-Driver model, and the validation set was used for evaluation. This protocol ensures a fair comparison with baselines on the nuScenes dataset. While trajectory prediction can be multi-modal, the approach focuses on the most likely mode/trajectory and computes errors relative to human driving trajectories. The results are validated on widely used benchmarks, and the same training and evaluation protocol was followed to ensure fairness.", "category": "Method_Mechanics"}, {"question": "What are the specific prompts used in each stage of the three-stage training approach, and are they all depicted in Figures 2 and 3?", "answer": "The three-stage approach involves designing proper prompts as inputs to the GPT-Driver, enumerating the reasoning process as a learning target, and fine-tuning the LLM using these prompts and reasoning targets. Figures 2 and 3 include all the inputs and outputs to GPT-Driver during this fine-tuning process.", "category": "Method_Mechanics"}, {"question": "How do authors address the concern that the information about obstacles in the simple prompting baseline is incomplete and acausal, given that it does not include obstacle velocities but describes future positions?", "answer": "The information about obstacles is not technically cheating. Authors transform both detection and prediction results into text descriptions, including the current location and future moving trajectory (in 6 waypoint coordinates) of the object. Authors simplify the prediction by describing the location where the object ends up in the 3rd second. This approach aligns with the perception-prediction-planning paradigm used in state-of-the-art driving systems and ensures a fair comparison with the baseline approach by using detection and prediction results from UniAD.", "category": "Method_Mechanics"}, {"question": "How does the proposed method demonstrate sufficient novelty and technical contributions to warrant acceptance at ICLR, given that it primarily utilizes the ChatGPT API for finetuning and appears to lack new modules or methodologies?", "answer": "The proposed method demonstrates sufficient novelty and technical contributions through: (1) Transforming motion planning into a language modeling task, which changes the typical regression problem into a coarse-to-fine region classification problem, overcoming regression weaknesses. (2) Introducing a fundamental architectural change by using a fine-tuned LLM as a motion planner, enhancing generalization and interpretability. (3) Being the first to show that LLMs can perform low-level motion planning and accurate numerical predictions, bringing new insights to the research community. Despite its simplicity, the learning objective, architecture, and observations are unique and different from existing approaches.", "category": "Method_Comparison"}, {"question": "Is Dauner et al. (2023)[1] a rule-based method, as stated in the RELATED WORKS section, or is it a hybrid approach combining rule-based and learning-based methods?\n[1] Dauner, D., Hallgarten, M., Geiger, A., & Chitta, K. (2023). Parting with Misconceptions about Learning-based Vehicle Motion Planning. arXiv preprint arXiv:2306.07962.", "answer": "Dauner et al. (2023)[1] is a hybrid approach that uses a rule-based IDM to generate an initial trajectory and then applies a learning-based neural network to refine the trajectory. authors have moved this literature to the rule-based method ", "category": "Method_Disambiguation"}, {"question": "How does the proposed method handle edge cases or complex driving scenarios, such as errors from the upstream perception module or particularly challenging situations?", "answer": "The method demonstrates robustness in handling edge cases and complex scenarios through a few-shot learning ablation study, which shows strong generalization with limited training data. Additionally, a comparison using UniAD perception and ground-truth perception (Table 1) indicates that the method performs well even with imperfect perception outputs.", "category": "Method_Mechanics"}, {"question": "What are the results of the experiment conducted to evaluate the impact of disabling ego-states in GPT-Driver, and how does GPT-Driver perform without ego-states?", "answer": "The results of the experiment, conducted by fine-tuning GPT-Driver on 10% of the data due to time constraints, are as follows: L2 distances at 1s, 2s, and 3s are 0.32, 0.67, and 1.15 respectively, with an average of 0.71. Collision rates at 1s, 2s, and 3s are 0.18, 0.44, and 1.11 respectively, with an average of 0.57. GPT-Driver demonstrates strong planning performance even without ego-states.\n| L2   |      |      |      | Collision |      |      |      |\n |------|------|------|------|------|------|------|------| \n| 1s   | 2s   | 3s   | Avg. | 1s  | 2s   | 3s   | Avg. |\n | 0.32 | 0.67 | 1.15 | 0.71 | 0.18  | 0.44 | 1.11 | 0.57 |", "category": "Experimental_Exposition"}, {"question": "What is the rationale for using ego-states and historical trajectories as inputs in the GPT-Driver, and how does this differ from UniAD's approach? Additionally, will the authors conduct experiments without ego-states and historical trajectories, as suggested by the reviewer?", "answer": "The rationale for using ego-states and historical trajectories in the GPT-Driver is that these data naturally exist and are easily obtainable in modern autonomous driving systems, making it practical to use them in motion planning. Unlike UniAD, which uses historical trajectories and applies additional tricks like IoU loss and post-optimization, the GPT-Driver does not rely on these methods. The authors appreciate the reviewer's suggestion and will conduct experiments to remove ego-states and historical trajectories, as well as perform closed-loop experiments, to further validate the method. Both in-context learning and finetuning in Table 3 are using the same data. The only difference is that in in-context learning, the input and desired outputs are collectively used as inputs to instruct the LLM, while in fine-tuning the desired outputs are used as targets to fine-tune the LLM. There is no post-optimization stage in fine-tuning.", "category": "Motivation_Analysis"}, {"question": "Did Dauner et al. 2023 independently reproduce the results on nuPlan open loop, and what are their conclusions regarding the use of scene context and ego-states in open-loop planning?", "answer": "Dauner et al. 2023 did not reproduce authors' results. They concluded that completely removing scene context is harmful and that a simple centerline representation is sufficient for strong open-loop performance. The conclusion of Dauner et al. 2023 is actually centerline-based planning outperforms other learning-based approaches, but leveraging only the ego-states didn't perform well on the nuPlan benchmark. Also, on the nuPlan open-loop evaluation benchmark, Dauner et al. 2023 also leveraged the ego-states (see Figure 2) in their approach.Therefore, authors disagree with the suggestion to remove ego-states in the open-loop planning benchmark.", "category": "Experimental_Exposition"}, {"question": "What are the results of the experiment conducted to evaluate the impact of disabling ego-states in GPT-Driver, and how does this affect its planning performance?", "answer": "The experiment results show that GPT-Driver maintains strong planning performance even after disabling ego-states. The L2 and collision metrics at 1s, 2s, and 3s intervals are as follows: L2 (0.32, 0.67, 1.15, Avg. 0.71) and Collision (0.18, 0.44, 1.11, Avg. 0.57).\n| L2   |      |      |      | Collision |      |      |      |\n |------|------|------|------|------|------|------|------|\n | 1s   | 2s   | 3s   | Avg. | 1s  | 2s   | 3s   | Avg. | \n| 0.32 | 0.67 | 1.15 | 0.71 | 0.18  | 0.44 | 1.11 | 0.57 | ", "category": "Experimental_Exposition"}, {"question": "Does the model described in Eq (2) actually require a map as input?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did Dauner et al. 2023 reproduce the results of the revised version on nuPlan open loop, confirming the minor impact of the data leakage issue?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the simplification of predicting the location where the object ends up in the 3rd second a part of the paper's approach to handling obstacle information?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did Dauner et al. 2023 claim that leveraging only ego-states performed well on the nuPlan benchmark?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the prediction of future object locations considered technically cheating in the context of the simple prompting baseline?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is Dauner et al. (2023)[1] described as a purely rule-based method?\n[1] Dauner, D., Hallgarten, M., Geiger, A., & Chitta, K. (2023). Parting with Misconceptions about Learning-based Vehicle Motion Planning. arXiv preprint arXiv:2306.07962.", "answer": "True", "category": "Claim_Verification"}, {"question": "Does Dauner et al. 2023 support the conclusion that removing scene context (as in Zhai et al. 2023) is sufficient for strong open-loop performance?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the proposed method inherently inherit the downsides of existing methods without any mitigation as verified in the experiments?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is Zhai et al. 2023 a peer-reviewed article and free from data leakage issues in their official repository?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper describe the future moving trajectory of objects in terms of 6 waypoint coordinates?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the information about obstacles in the simple prompting baseline describe only the current location of the objects?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the simple prompting baseline include obstacle velocities in its information about obstacles?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "p9pBJv1DTz", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "MindAgent: Emergent Gaming Interaction", "authors": ["Ran Gong", "Qiuyuan Huang", "Xiaojian Ma", "Hoi Vo", "Zane Durante", "Yusuke Noda", "Zilong Zheng", "Demetri Terzopoulos", "Li Fei-Fei", "Jianfeng Gao"], "abstract": "Large Language Models (LLMs) can perform complex scheduling in a multi-agent system and can coordinate agents to complete sophisticated tasks that require extensive collaboration. However, despite the introduction of numerous gaming frameworks, the community lacks adequate benchmarks that support the implementation of a general multi-agent infrastructure encompassing collaboration between LLMs and human-NPCs. We propose a novel infrastructure--- MindAgent---for evaluating planning and coordination-emergent capabilities in the context of gaming interaction. In particular, our infrastructure leverages an existing gaming framework to (i) require understanding of the coordinator for a multi-agent system, (ii) collaborate with human players via instructions, and (iii) enable in-context learning based on few-shot prompting with feedback. Furthermore, we introduce CuisineWorld, a new gaming scenario and its related benchmark that features a multi-agent collaboration efficiency and supervises multiple agents playing the game simultaneously. We have conducted comprehensive evaluations with a new auto-metric collaboration score CoS for assessing the collaboration efficiency. Finally, MindAgent can be deployed in real-world gaming scenarios in a customized VR version of CuisineWorld and adapted in the broader \"Minecraft\" gaming domain. Our work involving LLMs within our new infrastructure for general-purpose scheduling and coordination can elucidate how such skills may be obtained by learning from large language corpora.", "keywords": "Large Language Models;Decision Making;Multi-Agent Systems;Gaming Interaction", "qa_pairs": [{"question": "What are the definitions of the terms $q_{pim}$ and $c_{pim}$ in Equation 2?", "answer": "$q_{pim}$ refers to the utility or rewards the system will generate if the agent performs the specified action, and $c_{pim}$ refers to the cost of the action. The paper has been updated to clarify this.", "category": "Concept_Understanding"}, {"question": "How does the paper address the potential limitations of context length and manual prompt engineering costs in the LLM's ability to adapt to new planning problems across different domains?", "answer": "The paper addresses the potential limitations of context length and manual prompt engineering costs by: 1. Focusing on benchmarking multi-agent gaming interaction planning, where context length is not a major concern given the current collaboration proficiency of LLMs. 2. Highlighting that context length is becoming less of an issue with advancements like GPT-4's ability to handle 128k context tokens, equivalent to 300 pages of text. 3. Utilizing history trimming, a common practice, to mitigate context length limits. 4. Proposing optional memory-based approaches for extremely complicated tasks,  [1] Retrieval augmented generation for knowledge-intensive NLP tasks [2] Generation-augmented retrieval for open-domain question answering [3] JARVIS-1: Open-world Multi-task Agents with Memory-Augmented Multimodal Language Models [4] Compressing Context to Enhance Inference Efficiency of Large Language Models [5] LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models [6] https://openai.com/blog/new-models-and-developer-products-announced-at-devday", "category": "Method_Mechanics"}, {"question": "What does \\tau_int,(1) represent in the environment setting, and how is it determined? Additionally, what is the maximum horizon of an episode?", "answer": "\\tau_int,(1) represents the time step after which a new task is added in the environment. The specific value of \\tau_int(1) depends on the environment complexity and is computed based on the minimum number of steps required to complete the task in a single-agent setting. The process involves constructing a task graph, applying breadth-first search to determine the optimal task sequence, and incorporating activation wait times and edge times into the task intervals. Additionally, for each new connection in the task graph, authors increase the total task interval time by tripling the edge time. authors assume each subgoal requires at least, goto, get, and put 3 actions. The cumulative task interval is adjusted by a scaling factor of 0.3 and the variable \\tau, which ranges from 1.0 to 3.0 to represent different task difficulties. The maximum horizon of an episode is not explicitly stated in the rebuttal.", "category": "Experimental_Setup"}, {"question": "Does the paper discuss the potential drawbacks and failure cases of GPT-4, and if so, what are the common failure modes identified?", "answer": "The paper has been updated to include discussions on GPT-4's potential drawbacks. The common failure modes identified are: 1. Significant performance drop when GPT-4 lacks feedback, as shown in Table 2. 2. Sensitivity to prompt inputs, leading to performance drop without inference knowledge, as indicated in Table 3.", "category": "Experimental_Analysis"}, {"question": "What distinguishes the novelty of MineAgent's approach from prior research that also leverages scratchpad or memory, particularly in terms of centralized coordination planning, combinatorial action space, spatial-temporal reasoning, and benchmark creation?", "answer": "MineAgent distinguishes itself through centralized coordination planning, a complex combinatorial action space (e.g., 2,385,443,281 possible actions with four agents), sophisticated spatial-temporal reasoning (e.g., avoiding space conflicts), and the creation of a new benchmark with extensive experiments, including cross-domain adaptation to Minecraft. These aspects differentiate it from prior work, such as JARVIS-1, which focuses on memory-augmented multimodal language models.\nIn addition, compared with concurrent work focused on planning (e.g., [1]), authors have more agents, more environments, and environments are extensible and adaptable. \n[1] JARVIS-1: Open-world Multi-task Agents with Memory-Augmented Multimodal Language Models", "category": "Method_Comparison"}, {"question": "Why does the paper not provide comparisons between the proposed LLM-based planning approach and heuristic or RL planners, given that the environment features limited state and action spaces?", "answer": "The motivation of the paper is not to compare planning proficiency between LLMs and other canonical planning methods, but to explore whether LLMs can support sophisticated multi-agent collaboration and human-NPC collaboration tasks. The paper focuses on evaluating planning and coordination capabilities in the context of gaming interaction, as outlined in the abstract. However, the authors acknowledge that showing results of canonical planning approaches, including RL, could offer a better understanding of the dynamics of the collaboration tasks. They also argue that the environment's complexity is not trivial, with an action space expanding to 221 in single-agent scenarios and approximately 2.39 billion in four-agent scenarios. The authors implemented a masked version of Proximal Policy Optimization (PPO) and provided heavily engineered dense rewards, but RL strategies showed limited success compared to GPT-4, which achieved a 95% success rate. The RL policy also could not adapt to varying numbers of agents, new tasks, or changes in the number of ingredients, highlighting the complexity and dynamic nature of the environment.", "category": "Motivation_Analysis"}, {"question": "Why does the proposed environment bypass low-level control, and how does this simplification align with the research objectives of focusing on high-level planning in multi-agent gaming scenarios?", "answer": "The proposed environment bypasses low-level control to focus on multi-agent high-level planning in gaming scenarios, a domain that presents substantial challenges[1]. Even the best agent achieved only a limited CoS score below 0.7. This simplification aligns with the research objective of advancing Large Language Model (LLM) capabilities in high-level coordination and planning. The complexity of high-level planning, such as resolving conflicts and task dependencies, is significant, as evidenced by the limited CoS scores achieved. The decision to focus on high-level planning, particularly in the Cuisineworld scenario, was made to deeply understand LLM capabilities in multi-agent coordination before introducing low-level control complexities. This approach is consistent with concurrent research, such as DEPS[2], JARVIS-1[3], and Voyager[4], where low-level control is managed separately from high-level planning, a recognized and effective methodology in the field.\n[1] Korsah et al. A comprehensive taxonomy for multi-robot task allocation. [2] Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents [3] JARVIS-1: Open-world Multi-task Agents with Memory-Augmented Multimodal Language Models [4] Voyager: An Open-Ended Embodied Agent with Large Language Models", "category": "Motivation_Analysis"}, {"question": "What is the primary motivation behind the proposed multi-agent infrastructure, and how does it address the limitations of existing gaming frameworks?", "answer": "The primary motivation of this paper is to explore whether LLMs can support sophisticated multi-agent collaboration and human-NPC collaboration tasks, addressing the lack of adequate benchmarks for such interactions. The proposed infrastructure, MindAgent, evaluates planning and coordination capabilities in gaming contexts, focusing on the collaboration between LLMs and human-NPCs.", "category": "Motivation_Analysis"}, {"question": "How does the proposed method compare to canonical planning approaches like RL in terms of success rates and adaptability in multi-agent settings?", "answer": "The proposed method using GPT-4 achieves a 95% success rate in both level 1 and level 4 settings, whereas the canonical RL approach (PPO) achieves 45.14% and 59.8% success rates respectively. Additionally, the trained RL policy cannot adapt to varying numbers of agents, new tasks, or changes in the number of ingredients, highlighting the complexity and dynamic nature of the proposed environment.", "category": "Method_Comparison"}, {"question": "Why does the current study not include inter-agent communication in the multi-agent collaboration model, and how does this limitation align with the focus of the work on gaming scenarios?", "answer": "The current study does not include inter-agent communication because the focus is on a centralized planning scheme tailored for gaming scenarios, where faster response times are prioritized. While communication between agents is acknowledged as a critical aspect of collaboration, its inclusion would add communication costs and complexity, which may not be ideal for gaming applications. This limitation is aligned with the study's focus on collaboration between LLMs and human-NPCs in gaming, where a centralized approach is more suitable. ", "category": "Experimental_Analysis"}, {"question": "How is the term 'emergent interaction' defined and justified in the context of the paper, given the lack of direct communication between agents?", "answer": "The term 'emergent interaction' is defined twofold: (1) Interaction with the Environment, emphasizing agents' dynamic interaction without specialized training, showcasing emergent behaviors, and (2) Interaction with Human Gamers, focusing on agents' adaptation to human strategies and efficient collaboration with human players without needing specialized training to understand human gameplay.", "category": "Concept_Understanding"}, {"question": "Why does the work not provide a direct comparison between the proposed method and existing LLM-based Minecraft agents, despite evaluating in the same Minecraft environment?", "answer": "The work does not provide a direct comparison because it operates within a distinctly different framework, focusing on Multi-Agent Systems (MAS) rather than single-agent environments. The challenges and dynamics of MAS are fundamentally different from those of single-agent systems, which changes the nature of the problem and the applicable solutions. However, the authors have incorporated techniques from existing single-agent LLM-based works, such as leveraging feedback and incorporating demonstrations, which are also used in seminal and concurrent work on LLM + gaming.", "category": "Experimental_Analysis"}, {"question": "What are the common failure cases of the LLM agent allocation, and under what environmental or optimization constraints would the agent fail to allocate tasks correctly?", "answer": "Common failure cases include scenarios where the task complexity exceeds the LLM's reasoning capabilities, such as when there are too many dependencies between tasks or when tasks require highly specialized domain knowledge not present in the LLM's training data. Additionally, the agent may fail in dynamic environments where tasks or constraints change rapidly, as the centralized planning scheme may not adapt quickly enough. These limitations are discussed in detail in the updated paper.", "category": "Experimental_Analysis"}, {"question": "What approach did the authors adopt to facilitate human-agent communication in MindAgent, and how does it enhance collaboration?", "answer": "The authors adopted a straightforward method for facilitating communication between the human user and the LLM-based agent. Participants in the experiment were instructed to verbally express their intents or comments, which were then transcribed and integrated into the LLM’s context as 'current human instruction: <instruction>'. This approach allowed for real-time adjustments in the behavior of the LLM-based agents based on human inputs. For instance, in authors' demo videos can be found in the supplementary videos, when a user suggested a specific strategy or highlighted a priority task, the agent was able to promptly adapt its actions accordingly. This led to a more dynamic and responsive collaboration between humans and agents, showcasing the potential of natural language communication in enhancing the collaborative experience. In summary, while authors' initial trials in human-agent communication are relatively basic, they clearly demonstrate the effectiveness of integrating natural language communication in LLM-based collaborative environments. ", "category": "Method_Mechanics"}, {"question": "How can the design of MindAgent be adapted to support collaborative games requiring domain-specific knowledge, such as MOBA, FPS, and Diplomacy, and what modifications to the infrastructure are necessary to enhance its generalizability?", "answer": "The design of MindAgent can be adapted to support collaborative games requiring domain-specific knowledge through techniques like In-Context Learning (ICL), which injects domain-specific knowledge (e.g., recipes in CuisineWorld and Minecraft) into the context for decision-making and collaboration. Additionally, techniques like Retrieval-Augmented Generation (RAG) and Fine Tuning can further enhance MindAgent's capabilities to handle domain-specific complexities. These adaptations demonstrate the feasibility of extending MindAgent to other domains, and the paper has been updated to reflect these insights.", "category": "Method_Mechanics"}, {"question": "How are the actions of human players obtained from an inverse dynamic model in the real-time VR version of CuisineWorld, and how are the time steps and 'GoTo' action for the LLM agent implemented in this context?", "answer": "In the real-time VR version of CuisineWorld, the LLM checks user actions at regular 0.1-second intervals during the game loop. Human players control their agents through keyboard inputs, resulting in low-level actions like moving or picking up items. The LLM-agent operates on a higher level, inferring temporally extended actions by assessing changes in the high-level game state. This is achieved using a simple inverse dynamics model to translate low-level actions into high-level atomic actions. The 'GoTo' action is implemented using the A* pathfinding algorithm integrated into the Unreal Engine.", "category": "Method_Mechanics"}, {"question": "How does the collaboration score (CoS), defined as the average task success rate across different time intervals, effectively measure collaboration efficiency in CuisineWorld, and how can it be generalized to other domains?", "answer": "In CuisineWorld, CoS measures collaboration efficiency by evaluating how well teams maintain task success rates as the interval between dish orders decreases. Teams with good collaboration sustain high success rates even with shorter intervals, while poorly collaborating teams experience significant drops. This metric aligns with the dynamics of the Overcooked! game, where tasks flood in at varying intervals. CoS has been generalized to other domains, such as Minecraft, by adjusting task intervals to measure collaboration efficiency. In general multi-agent gaming scenarios, CoS can be applied wherever the game involves completing tasks, as it assesses a team's ability to handle varying task loads effectively.", "category": "Method_Mechanics"}, {"question": "Why was a traditional train/test split not implemented in the CuisineWorld dataset, and how does this align with the dataset's objectives?", "answer": "The CuisineWorld dataset does not include a traditional train/test split because its primary objective is to serve as a grand challenge for LLM-based multi-agent collaboration and emergent gaming interaction. The dataset is designed as a probing tool to explore the capabilities and limitations of LLMs in complex, dynamic environments. Omitting the train/test split allows for broader exploration and experimentation, aligning with recent research efforts in probing datasets, such as the study referenced in [1]. While specific prompt engineering for tweaking the benchmark is discouraged, the authors believe that open exploration of all approaches will lead to more interesting ideas and findings.\n[1] Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality", "category": "Motivation_Analysis"}, {"question": "How does the CuisineWorld benchmark differ in its approach to enhancing task complexity and diversity compared to existing Overcooked environments[1][2], particularly in terms of kitchen layouts, task design, and collaborative strategy assessment?\n[1] Voyager: An Open-Ended Embodied Agent with Large Language Models\n\n[2] Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents\n", "answer": "CuisineWorld differs from existing Overcooked environments by focusing on high-level multi-agent planning and coordination using Large Language Model-based Multi-agent Systems (MAS). Instead of diversifying kitchen maps, it enhances complexity through: 1) Enhanced Kitchen Layouts with a wider range of tools and ingredients (27 unique ingredients and 10 tools). For example, in [1], there are only two types of ingredients and two types of tools. In [2], there are two types of ingredients, and two types of tools; 2) Complex Task Design with a broader spectrum of recipes and varying difficulty levels (33 unique dishes).In [1,2] there are very limited number of dishes, 1 and 3 respectively; 3) Multi-Dish Episodes and Collaborative Strategy Assessment, requiring agents to complete multiple dishes within a single episode, using metrics like 'Collaborative Score' (CoS) to evaluate collaboration proficiency, especially considering task density and its effect on coordination. In [1] the goal is to finish as many dishes as possible in a limited amount of time. In [2], the goal is to finish one dish in the least amount of time. Both of them do not consider the density of the tasks (interval between dish orders coming to the kitchen) and its effect on coordination. ", "category": "Method_Comparison"}]}
{"id": "6M5G5hNiAU", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Composition", "authors": ["Guanting Dong", "Hongyi Yuan", "Keming Lu", "Chengpeng Li", "Mingfeng Xue", "Dayiheng Liu", "Wei Wang", "Zheng Yuan", "Chang Zhou", "Jingren Zhou"], "abstract": "Large language models (LLMs) with enormous pre-training tokens and parameter amounts emerge abilities including math reasoning, code generation, and instruction following. These abilities are further enhanced by supervised fine-tuning (SFT). The open-source community has studied on ad-hoc SFT for each ability, while proprietary LLMs are versatile for all abilities. It is important to investigate how to unlock them with multiple abilities via SFT. In this study, we specifically focus on the data composition between mathematical reasoning, code generation, and general human-aligning abilities during SFT. From a scaling perspective, we investigate the relationship between model abilities and various factors including data amounts, data composition ratio, model parameters, and SFT strategies. Our experiments reveal that different abilities exhibit different scaling patterns, and larger models generally show superior performance with the same amount of data. Mathematical reasoning and code generation improve as data amounts increase consistently, while the general ability is enhanced with about a thousand samples and improves slowly. We find data composition results in various abilities improvements with low data amounts, while conflicts of abilities with high data amounts. Our experiments further show that composition data amount impacts performance, while the influence of composition ratio is insignificant. Regarding the SFT strategies, we evaluate sequential learning multiple abilities are prone to catastrophic forgetting. Our proposed Dual-stage Mixed Fine-tuning (DMT) strategy learns specialized abilities first and then learns general abilities with a small amount of specialized data to prevent forgetting, offering a promising solution to learn multiple abilities with different scaling patterns.", "keywords": "Data Composition;Large Language Model;Scaling Analysis;Supervised Fine-tuning", "qa_pairs": [{"question": "How does the performance of training sequences math -> code -> general abilities compare to code -> math -> general abilities in terms of SFT abilities?", "answer": "The experiments conducted with six different training orders show that the SFT ability trained in the final stage retains relatively good performance. Specifically, if general and code abilities are trained in the first two stages, there is a noticeable performance decrease in code capability, while math capability does not show significant impact. Detailed results are provided in the Appendix section titled 'COMPARISON EXPERIMENT OF DIFFERENT TRAINING SEQUENCES'.", "category": "Experimental_Exposition"}, {"question": "What is the significance of the t-SNE graphs in understanding the task distribution, and why did the authors not derive a scaling law equation?", "answer": "The t-SNE graphs are used to understand the task distribution, revealing that the math reasoning dataset is distinct from others and that ShareGPT overlaps with coding tasks, which helps explain performance changes under multi-task learning. The authors did not derive a scaling law equation because they do not use test losses to evaluate task performance. Metrics like MT-Bench, which are not simple functions of test losses, make it unclear whether the power law still applies in this context. However, MT-Bench and Alpaca-Eval are valuable for quickly assessing human alignment performance.", "category": "Experimental_Analysis"}, {"question": "What is the novelty of the proposed method, given that it has been described as a mixture of different datasets?", "answer": "The novelty of the proposed method lies in its training routine, which involves training on specialized tasks first and then multi-task training on human alignment data with a small amount of specialized tasks. This approach relieves the decline in high-resource settings during multi-task learning and is empirically useful for building human-aligned chatbots with specialized abilities.", "category": "Method_Disambiguation"}, {"question": "What additional experimental results have been added to Table 1 to address the lack of comparisons between datasets and to demonstrate the generalization of the DMT strategy?", "answer": "The authors have added zero-shot results of single domain SFT models on other datasets to Table 1. Additionally, they evaluated the DMT strategy on more benchmarks, including GSM8K and Code Alpaca, and assessed performance on individual domain, mixed domain, and different training strategies on specialized ability benchmarks such as MATH and MBPP. The findings include: 1. Positive correlation between data volume and Llama model performance in MATH and MBPP. 2. Consistent patterns of conflicting high-resource and low-resource performance gains in MATH and MBPP for the llama-7B model. 3. Competitive results of DMT in MATH and MBPP while prioritizing general abilities, validating its effectiveness.", "category": "Experimental_Exposition"}, {"question": "What is the intuition behind selecting the value of k in the DMT experiments, and why does the performance trend appear inconsistent as k varies in Table 5?", "answer": "The trend of k is reliable. The inconsistency arises from the behavior of k in different ranges. When k increases from 0 to 1/256, a small amount of specialized data is mixed into the general ability data, leading to mutual benefits between general and low-resource specialized abilities, as concluded in RQ2. However, as k increases further (from 1/256 to 1/1), more specialized data is introduced, improving specialized abilities but causing conflicts in high-resource performance among different ability items, which results in a decline in general abilities.", "category": "Experimental_Analysis"}, {"question": "Why do the results on code mixing having negative effects with more data differ from prior work that showed adding code could help with NL tasks by adding structure?", "answer": "The discrepancy in results arises due to the significant difference in the amount of code data used for SFT. Prior studies used code data with token counts exceeding 10 billion, enabling models to capture transferable structured knowledge. In contrast, the paper's experiments used code data with token counts less than 0.01 billion, making it difficult for large models to learn transferable knowledge. Additionally, the paper's findings indicate that factors influencing code mixing are more related to the style, format, and distribution of inputs from different domains.", "category": "Experimental_Analysis"}, {"question": "Can the observed trends and scaling laws in the paper generalize beyond the three abilities (reasoning, coding, and aligning general human intentions) and the single dataset used for each ability?", "answer": "Yes, the DMT strategy demonstrates a certain degree of generalization. Results from additional benchmarks (MATH for mathematics and MBPP for code) show that the strategy performs well in both in-distribution (IND) and out-of-distribution (OOD) scenarios for mathematics and code. Detailed results are provided in the Appendix section titled 'RESULTS ON MORE BENCHMARKS IN MATH AND CODE'.", "category": "Experimental_Exposition"}, {"question": "What is the process for selecting data in each training dataset, and how does the use of multiple random seeds affect the comparative results of different subsets?", "answer": "The data selection for each dataset is performed using random selection with three distinct seeds. A comparative analysis of the results from different subsets across three benchmark tests shows that DMT consistently outperforms other methods regardless of the random seed used. The differences in subsets do not significantly impact the overall trend of the results.", "category": "Experimental_Setup"}, {"question": "How does the model's performance compare when fine-tuned with the same number of data points from each training dataset versus using the same subset ratio, and does the size disparity between datasets affect learning other abilities?", "answer": "The total number of samples for each dataset is provided in the Appendix section titled 'SFT Dataset,' along with a comparative table showcasing the data size and proportions. The results indicate no significant difference in performance between the two settings (equal data amount vs. equal data proportion) across three benchmark tests, and these findings do not significantly impact the main experimental conclusions. Detailed results are available in the Appendix section titled 'EQUAL DATA AMOUNT VS. EQUAL DATA PROPORTION.'", "category": "Experimental_Exposition"}, {"question": "What is the rationale for selecting the 15th layer of the model in Figure 5, and how do the visualization results compare between the start, middle, and end layers?", "answer": "The visualization results of the start layer are relatively chaotic, while the middle and end layers provide clearer and more stable visualizations. The middle layer and the last layer both show consistent separation in mathematical data representations, though there is some overlap between code and general samples. The middle layer is particularly stable compared to the early layer. Additional comparison results are provided in the Appendix section titled 'VISUALIZATION OF DIFFERENT LAYERS'.", "category": "Motivation_Analysis"}, {"question": "Does the improved test-time accuracy of the model, as discussed in Section 4.2, necessarily indicate better generalization ability, especially when the model might overfit to the training distribution after removing code and math data from ShareGPT?", "answer": "The authors acknowledge that better test-time accuracy does not necessarily equate to better generalization ability. They have conducted additional experiments on more diverse datasets, including out-of-distribution benchmarks, to evaluate the model's generalization capability. The results, included in the Appendix section titled 'RESULTS ON MORE BENCHMARKS IN MATH AND CODE,' demonstrate that the DMT strategy maintains competitive performance across both in-distribution and out-of-distribution benchmarks, supporting its generalization ability. ", "category": "Experimental_Analysis"}]}
{"id": "nKvGCUoiuW", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "MiniGPT-v2: Large Language Model as a Unified Interface for Vision-Language Multi-task Learning", "authors": ["Jun Chen", "Deyao Zhu", "Xiaoqian Shen", "Xiang Li", "Zechun Liu", "Pengchuan Zhang", "Raghuraman Krishnamoorthi", "Vikas Chandra", "Yunyang Xiong", "Mohamed Elhoseiny"], "abstract": "Large language models have shown their remarkable capabilities as a general interface for various language-related applications. Motivated by this, we target to build a unified interface for completing many vision-language tasks including image description, visual question answering, and visual grounding, among others. The challenge for achieving this is to use a single model for performing diverse vision-language tasks effectively with simple multi-modal instructions. To address this issue, we introduce MiniGPT-v2, a model can be treated a unified interface for better handling various vision-language tasks. We propose using unique identifiers for different tasks when training the model. These identifiers enable our model to distinguish each task instruction effortlessly and also improve the model learning efficiency for each task. After our three-stage training, our experiments show that MiniGPT-v2 achieves strong performance on many visual question answering and visual grounding benchmarks compared to other vision-language generalist models. Our trained models and codes will be made available.", "keywords": "vision-language foundation model", "qa_pairs": [{"question": "Have the authors evaluated the model's performance in a zero-shot manner by removing all supervised data, and if so, what were the results on various VQA benchmarks?", "answer": "Yes, the authors evaluated the model in a zero-shot manner by removing all supervised data and using only training caption data (LAION, CC3M, SBU) and weakly labeled visual grounding data (GRIT-20M). The results on various VQA benchmarks showed a performance drop: VizWiz (24.1 vs. 29.4), VSR (51.1 vs. 59.9), IconVQA (43.7 vs. 45.6), RefCOCO (34.2 vs. 74.7), RefCOCO+ (34.6 vs. 60.9), and RefCOCOg (52.2 vs. 74.2) when compared to using supervised data.\n|                         \t| vizwiz |  VSR  | IconVQA | RefCOCO (testB) | RefCOCO+ (testB) | RefCOCOg (test) |\n|-----------------------------|:------:|:-----:|:-------:|:---------------:|:----------------:|:---------------:|\n| Without supervised data \t|  24.1  | 51.1  |  43.7   |  \t34.2   \t|   \t34.6   \t|   \t52.2  \t|\n| With training supervised data|  29.4  | 59.9  |  45.6   |  \t74.7   \t|   \t60.9   \t|   \t74.2  \t|\n", "category": "Experimental_Exposition"}, {"question": "What is the impact on performance when freezing the image encoder versus not freezing it, and when using LoRA versus not using it?", "answer": "Fine-tuning the vision encoder results in a similar performance to not fine-tuning, with an average performance drop of 1.1 acc across VQA benchmarks. The slight drop might be due to the challenges in optimizing the coupled vision encoder and large language model.\n| Model                              \t| OKVQA |  GQA  | vizwiz |  VSR  | IconVQA |  HM   | average |\n|----------------------------------------|:-----:|:-----:|:------:|:-----:|:-------:|:-----:|:-------:|\n| Finetuning the vision encoder      \t| 51.6  | 53.8  |  27.6  | 56.3  |  45.3   | 57.6  |  48.7   |\n| Without finetuning the vision encoder  | 52.1  | 54.6  |  29.4  | 59.9  |  45.6   | 57.4  |  49.8   |", "category": "Experimental_Exposition"}, {"question": "What is the rationale behind concatenating four adjacent visual output tokens, and have the authors experimented with concatenating more tokens (e.g., 8 or 16) to evaluate the impact on performance?", "answer": "The authors conducted ablation studies by concatenating 8 and 16 adjacent visual output tokens. The results, provided in the tables, show that concatenating more tokens leads to a drop in performance for both visual question answering and visual grounding tasks. The performance drop is more pronounced in visual grounding, likely due to the loss of spatial understanding when merging more tokens.\n| Model               \t| OKVQA |  GQA  | vizwiz |  VSR  | IconVQA |  HM   | average | \n|-------------------------|:-----:|:-----:|:------:|:-----:|:-------:|:-----:|:-------:|\n| Concatenate 4 tokens\t| 52.1  | 54.6  |  29.4  | 59.9  |  45.6   | 57.4  |  49.8   | \n| Concatenate 8 tokens\t| 50.3  | 53.9  |  29.1  | 55.6  |  45.9   | 57.1  |  48.6   |  \t\n| Concatenate 16 tokens   | 50.5  | 52.5  |  26.7  | 55.4  |  45.8   | 57.2  |  48.0   | \n\n| Model               \t| RefCOCO (testB) | RefCOCO+ (testB) | RefCOCOg (test) | average |\n|-------------------------|:---------------:|:----------------:|:---------------:|:-------:|\n| Concatenate 4 tokens\t|  \t74.7   \t|   \t60.9   \t|   \t74.2  \t|  69.9   |\n| Concatenate 8 tokens\t|  \t70.3   \t|   \t57.4   \t|   \t71.1  \t|  66.3   |\n| Concatenate 16 tokens   |  \t66.3   \t|   \t52.4   \t|   \t64.4  \t|  61.0   |", "category": "Experimental_Exposition"}, {"question": "How does the performance of using 448x448 images compare to using 224x224 images in terms of visual question answering and visual grounding tasks?", "answer": "Training on 448x448 images yields similar performance to 224x224 images for visual question answering tasks, as shown by the average scores of 49.8 and 49.6, respectively. However, for visual grounding tasks, 448x448 images significantly outperform 224x224 images, with an average accuracy increase of +10.2 (69.9 vs. 59.7).\n| Image Size  | OKVQA |  GQA  | vizwiz |  VSR  | IconVQA |  HM   | average |\n|--------|:-----:|:-----:|:------:|:-----:|:-------:|:-----:|:-------:|\n| 224x224| 52.0  | 53.9  |  30.2  | 58.6  |  46.5   | 56.2  |  49.6   |\n| 448x448| 52.1  | 54.6  |  29.4  | 59.9  |  45.6   | 57.4  |  49.8   |\n\n| Image Size | RefCOCO (testB) | RefCOCO+ (testB) | RefCOCOg (test) | average |\n|------------|:---------------:|:----------------:|:---------------:|:-------:|\n| 224x224\t|  \t63.3   \t|   \t52.6   \t|   \t63.1  \t|  59.7   |\n| 448x448\t|  \t74.7   \t|   \t60.9   \t|   \t74.2  \t|  69.9   |", "category": "Method_Comparison"}, {"question": "What criteria were used to select the baseline models for each task, and why were the vision-and-language foundation models listed in Table 6 not compared in Table 3 and Table 4, specifically addressing the inclusion of Minigpt4 in Table 3 and the exclusion of mPLUG-Owl?", "answer": "The baseline models for Table 4 were selected to align with evaluation benchmarks from the work [1] for direct and consistent comparison with existing works. The results in Table 6 were included to maintain consistency with previous evaluation benchmarks from the work [2]. The mPLUG-Owl baseline was added in the updated paper (Table 3) to address its importance in comparison.\n[1] Chen, etc. Shikra: Unleashing multi-modal llm’s referential dialogue magic. [2] Li, etc. Evaluating Object Hallucination in Large Vision-Language Models.", "category": "Experimental_Analysis"}, {"question": "Why was 'MIMIC-IT: Multi-Modal In-Context Instruction Tuning' not included as a baseline in the original study, and how does the proposed model compare to the Otter model fine-tuned on MIMIC-IT?", "answer": "The authors acknowledge the importance of MIMIC-IT, comparison with the Otter model fine-tuned on MIMIC-IT in Table 3. The results show that the proposed model consistently outperforms the Otter model in VQA performance.", "category": "Method_Comparison"}, {"question": "Why were the tasks of visual question answering and referring expression comprehension chosen for evaluation, and how does this address the concern of potential overfitting from using test sets in training?", "answer": "The tasks of visual question answering and referring expression comprehension were chosen for evaluation due to their representativeness in assessing the foundational capabilities of vision-language models in visual knowledge querying and visual grounding. Additionally, these tasks have well-established benchmarks, unlike other tasks such as detailed image captioning and object parsing, which lack standardized evaluation metrics. This focus on well-established benchmarks helps mitigate the risk of overfitting.", "category": "Method_Mechanics"}, {"question": "What is the performance improvement observed when using the [refer] task identifier compared to not using it, as shown in the ablation results for RefCOCO/+/g?", "answer": "The ablation results show that using the [refer] task identifier improves the model performance on RefCOCO/+/g by 1.3 accuracy points on average. Specifically, the model with the [refer] task identifier achieves 74.7, 60.9, and 74.2 accuracy on RefCOCO (test-B), RefCOCO+ (test-B), and RefCOCOg (test), respectively, compared to 73.2, 59.7, and 72.8 without the [refer] task identifier.\n| Model                               | RefCOCO (test-B) | RefCOCO+ (test-B) | RefCOCOg (test) | average | \n|----------------------------------------|:----------------:|:-----------------:|:---------------:|:-------:| \n| authors' w/o [refer] task identifier    |    73.2    |     59.7    |    72.8   |  68.6   | \n| authors' w/ [refer]                     |    74.7    |     60.9    |    74.2   |  69.9   |", "category": "Experimental_Exposition"}, {"question": "Why did the authors choose to evaluate the model only on visual question answering and referring expression comprehension tasks, despite training it on various tasks?", "answer": "The authors focused their evaluation on visual question answering and referring expression comprehension tasks due to two key reasons: 1. These tasks are representative and essential for evaluating the foundational capabilities of vision-language models, particularly in visual knowledge querying and visual grounding. 2. Other tasks, such as detailed image captioning and object parsing and grounding, lack well-established benchmarks and standardized evaluation metrics, making it difficult to assess performance reliably. Therefore, the authors prioritized well-established benchmarks like VQA and RefCOCO.", "category": "Motivation_Analysis"}, {"question": "What evaluation metrics were used to benchmark the model, and how do they contribute to a comprehensive evaluation?", "answer": "The study employed three evaluation metrics: open-ended VQA, visual grounding, and hallucination evaluation. These metrics collectively provide a comprehensive evaluation of the model, as detailed in the Appendix.", "category": "Experimental_Setup"}, {"question": "Why does the model perform less effectively on RefCOCO+ compared to RefCOCO and RefCOCO(g), as shown in Table 4?", "answer": "The model performs less effectively on RefCOCO+ because RefCOCO+ excludes location-based expressions and focuses on more complex descriptions of object appearances, which contrasts with RefCOCO and RefCOCO(g) that include location-based and longer descriptions. Performance of MiniGPT-v2 on RefCOCO+ also performs well and can outperform Shikra on Test-B split. The results in Table 4 indicate that the general performance of MiniGPT-v2 on RefCOCO+ does not perform as well as RefCOCO and RefCOCO(g). The model is stronger in processing location-based referring expressions and longer descriptions, which aligns better with the features of RefCOCO and RefCOCO(g). This difference in dataset features and the model's design and training strategy contribute to the observed performance gap.", "category": "Experimental_Analysis"}, {"question": "How are the CHAIR_i, CHAIR_s, and Len metrics calculated, and how do they evaluate hallucination in MiniGPT-v2? Additionally, how do the results for MiniGPT-v2 differ across three different prompts (without task identifier, with [grounding] identifier, and with [caption] identifier)?", "answer": "In authors' study, authors employ the evaluation metrics for hallucination as defined in [1] to assess the performance of MiniGPT-v2. The CHAIR_i (Instance-level Hallucination) metric evaluates hallucination at the object instance level and is calculated as the ratio of hallucinated objects to the total number of objects mentioned in the caption: $$ CHAIR_i = \\frac{| \\{\\text{hallucinated objects}\\} |}{| \\{\\text{all mentioned objects}\\} |} $$ CHAIR_s (Sentence-level Hallucination). The CHAIR_s (Sentence-level Hallucination) metric assesses hallucination at the sentence level and is calculated as the ratio of captions containing hallucinated objects to the total number of captions generated: $$ CHAIR_s = \\frac{| \\{\\text{captions with hallucinated objects}\\} |}{| \\{\\text{all captions}\\} |} $$ Len (Length of Caption) refers to the number of words in the generated caption, measuring caption verbosity or conciseness. For MiniGPT-v2, the results vary across prompts: 1. Without Task Identifier: The model generates captions with rich details. 2. With [grounding] Identifier: The captions are more grounded to image objects but describe fewer visual attributes and relationships. 3. With [caption] Identifier: The captions are very short.\n [1] Li, etc. Evaluating Object Hallucination in Large Vision-Language Models.", "category": "Experimental_Exposition"}, {"question": "How does MiniGPT-V2's ability to handle compositional tasks and its unified platform compare to the performance and capabilities of specialized models like BEIT-3, X-VLM, and Co-DETR?", "answer": "MiniGPT-V2 offers a unified and versatile platform capable of handling various tasks without needing multiple specialized systems. It can merge different abilities to handle compositional tasks, as demonstrated in Fig. 4, where it detects a specific cow in a photo and answers subsequent questions about it, a task beyond the scope of VQA expert models. Additionally, MiniGPT-V2 has the potential to activate more advanced vision-language abilities (advanced reasoning and in-contextual learning, etc)  in the future, which specialized methods lack.", "category": "Method_Comparison"}, {"question": "What are the key differences between the three-stage training strategy proposed in this paper and the strategies used in Lynx and Qwen-VL, particularly in terms of image resolution consistency, data sampling strategy, and vision encoder training?", "answer": "The key differences are: 1. Consistency in image resolution: Unlike Lynx and Qwen-VL, which progressively increase image resolution during training, the model maintains a constant resolution of 448x448 throughout the whole training process. 2. Data sampling strategy: Qwen-VL focuses on image caption data in the first stage, diverse multi-task datasets in the second stage, and instruction finetuning and dialogue data in the third stage. In contrast, the model samples noisy and fine-grained data (e.g., image captions, visual grounding, and visual question answering) in the first stage, fine-grained multi-task data in the second stage, and both fine-grained multi-task data and visual instruction/dialogue data in the third stage. 3. Vision encoder training: Qwen-VL trains the vision encoder in the first two stages, while the model keeps the vision encoder frozen throughout the entire training process.", "category": "Method_Comparison"}, {"question": "How does the proposed method of concatenating every four adjacent visual tokens into a single token in Visual Transformer (ViT) outputs compare in terms of performance and computational efficiency when trained on 224x224 images without token merging and 448x448 images with token merging?", "answer": "The method of concatenating every four adjacent visual tokens into a single token was tested by training the model on 224x224 images without token merging and 448x448 images with token merging, both resulting in an equivalent number of image patches. Training on 448x448 images without merging tokens significantly slowed down the training speed and increased GPU memory usage. The comparative analysis showed that the 448x448 configuration with token merging yielded superior results in both Visual Question Answering (VQA) and visual grounding tasks, achieving a 10.2% higher accuracy in the RefCOCO dataset compared to the 224x224 configuration without token merging.", "category": "Method_Comparison"}, {"question": "Why does MiniGPT-V2 use task identifiers such as '[vqa]' and '[grounding]', and how do they impact the model's ability to handle diverse visual-language tasks?", "answer": "MiniGPT-V2 uses task identifiers  such as '[vqa]' and '[grounding]' to enhance learning efficiency and reduce ambiguity in diverse visual-language tasks. While MiniGPT-V2 utilizes task identifiers for improved performance across different tasks, these identifiers act as regularization for different task inputs and do not hinder the model's conversational capabilities. Users can engage in a wide range of vision-language tasks without the need to adding task identifiers.The model can also handle identifier-free interactions effectively. An ablation study shows that a variant of MiniGPT-V2 without task identifiers performs nearly as well as the original, but overall, task identifiers still lead to better performance.", "category": "Method_Disambiguation"}, {"question": "Were task identifier tokens and spatial location tokens like <0>, <1>, …<100> added to the vocabulary as special tokens and their embeddings initialized?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the paper use only a single evaluation metric to benchmark the model's performance?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "FAO4VS9QRV", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Delta-LoRA: Fine-Tuning High-Rank Parameters with the Delta of Low-Rank Matrices", "authors": ["Bojia Zi", "Xianbiao Qi", "Lingzhi Wang", "Jianan Wang", "Kam-Fai Wong", "Lei Zhang"], "abstract": "In this paper, we present \\textbf{Delta-LoRA}, which is a novel parameter-efficient approach to fine-tune large language models (LLMs). In contrast to LoRA and other low-rank adaptation methods such as AdaLoRA, Delta-LoRA not only updates the low-rank matrices $A$ and $B$, but also propagate the learning to the pre-trained weights $W$ via updates utilizing the delta of the product of two low-rank matrices ($A^{(t+1)}B^{(t+1)} - A^{(t)}B^{(t)}$). Such a strategy effectively addresses the limitation that the incremental update of low-rank matrices is inadequate for learning representations capable for downstream tasks. Moreover, as the update of $W$ does not need to compute the gradients of $W$ and store their momentums, Delta-LoRA shares comparable memory requirements and computational costs with LoRA. Extensive experiments show that Delta-LoRA significantly outperforms existing low-rank adaptation methods. We further support these results with comprehensive analyses that underscore the effectiveness of Delta-LoRA.", "keywords": "Parameter-Efficient Fine-Tuning", "qa_pairs": [{"question": "What are the fundamental differences between Delta-LoRA and LoRA, particularly in terms of weight updates, rank of incremental weight matrices, and gradient flow?", "answer": "Delta-LoRA differs from LoRA in three key aspects: (1) In Delta-LoRA, the pretrained weight matrix $\\boldsymbol{W}$ is updated, whereas it remains fixed in LoRA. This enhances the model's representative ability. (2) The rank of the learned incremental weight matrix in Delta-LoRA, given by $\\text{Rank}(\\Delta \\boldsymbol{W_{Delta-LoRA}}) = \\text{Rank}(\\boldsymbol{W}^{(T)} - \\boldsymbol{W}^{(0)} + \\boldsymbol{AB})$, is larger than that in LoRA, which is $\\text{Rank}(\\Delta \\boldsymbol{W}_{LoRA}) = \\text{Rank}(\\boldsymbol{AB})$. (3) The gradient flow differs between the two methods, as evidenced by the distinct back-propagation processes for Delta-LoRA and LoRA.$\\frac{\\partial \\mathcal{L}}{\\partial A} = B^{\\top}\\cdot\\frac{\\partial \\mathcal{L}}{\\partial W}$ and $\\frac{\\partial \\mathcal{L}}{\\partial B} = \\frac{\\partial \\mathcal{L}}{\\partial W}\\cdot A^{\\top}$.\n\nHere is the **back-propagation process of LoRA**: $\\frac{\\partial \\mathcal{L}}{\\partial A^{(t+1)}}= \\frac{\\partial \\mathcal{L}}{\\partial W^{(t+1)}}\\cdot (B^{(t+1)})^{\\top}=\\frac{\\partial \\mathcal{L}}{\\partial W^{(0)}+(A^{(t)}+\\Delta A^{(t)})(B^{(t)} + \\Delta B^{(t)})}\\cdot (B^{(t)}+\\Delta (B^{(t)}))^{\\top}$\n$\\frac{\\partial \\mathcal{L}}{\\partial B^{(t+1)}} = (A^{(t+1)})^{\\top} \\cdot \\frac{\\partial \\mathcal{L}}{\\partial W^{(t+1)}}=(A^{(t)}+\\Delta (A^{(t)}))^{\\top} \\cdot \\frac{\\partial \\mathcal{L}}{\\partial W^{(0)}+(A^{(t)}+\\Delta A^{(t)})(B^{(t)} + \\Delta B^{(t)})}$\n\nHere is the **back-propagation process of Delta-LoRA**:\n$\\frac{\\partial \\mathcal{L}}{\\partial A^{(t+1)}} = \\frac{\\partial \\mathcal{L}}{\\partial W^{(t+1)}}\\cdot (B^{(t+1)})^{\\top}=\\frac{\\partial \\mathcal{L}}{\\partial ((W^{(t)}+ \\lambda \\Delta A^{(t)}B^{(t)}) + (A^{(t)}+\\Delta A^{(t)})(B^{(t)}+\\Delta B^{(t)})) }\\cdot (B^{(t)}+\\Delta (B^{(t)}))^{\\top}$\n$\\frac{\\partial \\mathcal{L}}{\\partial B^{(t+1)}} = (A^{(t+1)})^{\\top} \\cdot \\frac{\\partial \\mathcal{L}}{\\partial W^{(t+1)}} = (A^{(t)}+\\Delta (A^{(t)}))^{\\top} \\cdot \\frac{\\partial \\mathcal{L}}{\\partial ((W^{(t)}+ \\lambda\\Delta A^{(t)}B^{(t)}) + (A^{(t)} + \\Delta A^{(t)})(B^{(t)} + \\Delta B^{(t)})  )}$.\n", "category": "Method_Comparison"}, {"question": "What experiments were conducted to evaluate the performance of Delta-LoRA on larger models, and what were the results?", "answer": "The authors evaluated the performance of the LLaMA-7B model on the Alpaca dataset. Using GPT-4 evaluation, Delta-LoRA improved LoRA by around 10%, demonstrating its effectiveness.", "category": "Experimental_Exposition"}, {"question": "Under what conditions does the performance loss compared to fine-tuning not occur, and why?", "answer": "The performance loss compared to fine-tuning does not occur under two conditions: 1. When the gap between the finetuned data and the pretrained data is large, as the full-finetuned method can outperform LoRA and other PEFT methods in such cases. 2. When the data scale of the finetuned data is small, as full finetuning mode tends to overfit quickly, making Delta-LoRA and LoRA show better results. These conditions are consistent with the observed results.", "category": "Experimental_Analysis"}, {"question": "How does Delta-LoRA's performance compare to full finetuning with stateless optimizers, specifically SGD, on the CoLA dataset under the specified experimental conditions?", "answer": "Under the specified experimental conditions (rank=8, α=16, learning rate=0.01, and momentum=0.9), Delta-LoRA achieves a performance of 64.32, while LoRA achieves 63.67 on the CoLA dataset. Additionally, the training process for LoRA with SGD crashed after 57 epochs, highlighting the instability of the SGD optimizer for NLP tasks.", "category": "Method_Comparison"}, {"question": "What evidence supports the claim that Delta-LoRA can nearly match the performance of full finetuning with large models?", "answer": "The effectiveness of Delta-LoRA was tested with LLaMA-7B on the Alpaca dataset. The results show that Delta-LoRA has a memory cost comparable to LoRA, while outperforming LoRA by approximately 10% under LLM evaluation.", "category": "Experimental_Exposition"}, {"question": "How do you address the concern that the differences in empirical results are not statistically significant and may be due to noise rather than real performance gains?", "answer": "The paper maintained consistent training settings with LoRA, including learning rate, weight decay, rank number, training epochs, and a fixed random seed (0), ensuring no parameter search. This method showed significant improvement over AdaLoRA and DyLoRA under the same settings and demonstrated consistent gains across 8 NLU and 4 NLG datasets, validating the effectiveness of the approach.", "category": "Experimental_Setup"}, {"question": "What are the standard errors for the empirical results on the GLUE benchmark, and how do the choices of learning rate and start steps affect performance?", "answer": "The standard errors for the GLUE benchmark are as follows: MNLI (87.62 $\\pm$ 0.21), SST-2 (95.29 $\\pm$ 0.23), MRPC (90.60 $\\pm$ 0.14), CoLA (64.64 $\\pm$ 0.86), QNLI (93.09 $\\pm$ 0.15), QQP (91.01 $\\pm$ 0.06), RTE (87.00 $\\pm$ 0.36), STS-B (91.61 $\\pm$ 0.04), AVG (87.60). A better choice of learning rate and start steps, as detailed in Section A.5, can further improve the performance of Delta-LoRA.\n\n| MNLI| SST-2 | MRPC | CoLA | QNLI | QQP | RTE | STS-B | AVG| \n|----------|----------|----------|----------|----------|----------|----------|----------|----------|\n | 87.62 $\\pm$ 0.21 | 95.29 $\\pm$ 0.23 | 90.60 $\\pm$ 0.14 | 64.64 $\\pm$ 0.86 | 93.09 $\\pm$ 0.15 | 91.01 $\\pm$ 0.06 | 87.00 $\\pm$ 0.36 | 91.61 $\\pm$ 0.04 | 87.60| ", "category": "Experimental_Exposition"}, {"question": "Why did the authors choose different training settings for AdaLoRA and DyLoRA compared to LoRA, and how does this affect the performance comparison reported in the paper?", "answer": "For a fair comparison, the authors used the same training setups as LoRA. DyLoRA aims to improve the performance of lower rank during inference rather than the rank selected at the training stage, resulting in memory overhead if higher ranks were used. Therefore, the authors reported DyLoRA's results based on the low-rank setting. AdaLoRA uses different settings from LoRA, including fine-tuning each layer of the pretrained matrices in Transformer, which causes extra activation memory consumption (19.79 GB extra GPU memory for LLaMA-7B). Additionally, AdaLoRA has 7 extra hyperparameters (i.e. target_rank, reg_orth_coef, init_warmup, final_warmup, mask_interval, beta1, beta2) that are difficult to optimize for different datasets and models. Thus, the authors chose not to use similar training setups with AdaLoRA but instead used the same setups as LoRA.", "category": "Experimental_Analysis"}, {"question": "What is the rationale for initially limiting the experiments to rank = 8, and what are the results of experiments using different ranks on the CoLA dataset?", "answer": "The initial experiments were conducted with rank = 8 to maintain consistency with the training setups used in LoRA. Authors use the rank=8 on 8 datasets in GLUE benchmark, while rank=4/8 for NLG tasks.Experiments using different ranks on the CoLA dataset showed that higher ranks, such as 16, 32, and 64, further improve the performance of Delta-LoRA, as evidenced by the results in the provided table.\n| Rank | LoRA | Delta-LoRA |\n|----------|----------|----------|\n| 8 | 63.17 | **63.82** |\n| 16 | 63.06 | **64.30** | \n| 32 | 63.81 | **64.79** | \n| 64 | 63.39 | **65.59** |", "category": "Experimental_Exposition"}, {"question": "Why was a fixed learning rate used for Delta-LoRA, and how does this ensure a fair comparison with LoRA?", "answer": "A fixed learning rate was used for Delta-LoRA to ensure a fair comparison, as it matches the learning rate reported in the original paper and code (e.g., 2e-4 on the E2E Challenge dataset). Additionally, a linear learning rate scheduler was applied during training, which is the same setting used for LoRA. This consistency in hyperparameters ensures that the reported performance of Delta-LoRA is directly comparable to LoRA.\n| MNLI| SST-2 | MRPC | CoLA | QNLI | QQP | RTE | STS-B | AVG|\n|----------|----------|----------|----------|----------|----------|----------|----------|----------|\n| 87.62 $\\pm$ 0.21 | 95.29 $\\pm$ 0.23 | 90.60 $\\pm$ 0.14 | 64.64 $\\pm$ 0.86 | 93.09 $\\pm$ 0.15 | 91.01 $\\pm$ 0.06 | 87.00 $\\pm$ 0.36 | 91.61 $\\pm$ 0.04 | 87.60 |", "category": "Experimental_Setup"}, {"question": "What is the reason for setting the trainable parameters of Delta-LoRA higher than LoRA in multiple experiments, and how does this relate to the improvement in performance?", "answer": "The higher trainable parameters in Delta-LoRA are due to the inclusion of adjustable pretrained parameters, updated using the delta of low rank matrices without additional GPU cost. LoRA cannot update the pretrained $\\boldsymbol{W}$, resulting in zero extra updatable trainable parameters. The improvement in performance is attributed to the Delta-LoRA update strategy, which is further enhanced by removing the Dropout layer.", "category": "Experimental_Setup"}, {"question": "Can Delta-LoRA be applied to more large language models, and what are the key strengths of Delta-LoRA that explain its effectiveness?", "answer": "Delta-LoRA has been applied to a variety of language models, including GPT-2, Bart-Large, and LLaMA-7B, and the experimental results suggest it can generalize well to other large language models. The key strengths of Delta-LoRA are: 1) Mathematically, $g_W = g_{AB}$, making it reasonable to update the pretrained $W$ by $\\Delta AB$. 2) Delta-LoRA has the lowest cosine similarity with the pretrained model among all baselines, indicating it induces the most significant modifications to the final matrix $\\boldsymbol{W}^{*}$. 3) Delta-LoRA shows better performance compared to LoRA on LLaMA-7B, yielding higher evaluation results with the Alpaca dataset.", "category": "Method_Disambiguation"}, {"question": "Have the authors applied Delta-LoRA to LLaMA-7B and provided the experimental results for comparison with LoRA?", "answer": "Yes, authors have tested Delta-LoRA on LLaMA-7B, adhering to standard settings. The results, which show approximately a 10% improvement over LoRA.", "category": "Experimental_Exposition"}, {"question": "Are standard errors provided for the empirical results on the GLUE benchmark in the Section A.5?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was a parameter search or random seed selection used to optimize the performance of Delta-LoRA?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the differences in empirical results statistically significant with error margins provided?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "89l6VLPrin", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Graph layouts and graph contrastive learning via neighbour embeddings", "authors": ["Marius Keute", "Alica Guzmán", "Dmitry Kobak"], "abstract": "In node-level graph representation learning, there are two distinct paradigms. One is known as graph layouts, where nodes are embedded into 2D space for visualization purposes. Another is graph contrastive learning, where nodes are parametrically embedded into a high-dimensional vector space based on node features. In this work, we show that these two paradigms are intimately related, and that both can be successfully approached via neighbour embedding methods. First, we introduce graph t-SNE for two-dimensional graph drawing, and show that the resulting layouts outperform all existing algorithms in terms of local structure preservation, as measured by kNN classification accuracy. Second, we introduce graph contrastive neighbor embedding (graph CNE)}, which uses a fully-connected neural network to transform graph node features into an embedding space by optimizing the contrastive InfoNCE objective. We show that graph CNE, while being conceptually simpler than most existing graph contrastive learning methods, produces competitive node representations, with state-of-the-art linear classification accuracy.", "keywords": "Graph Layout;Contrastive Learning;t-SNE", "qa_pairs": [{"question": "Why was the GNN architecture not used for comparing the proposed graph CNE with other GCL methods?", "answer": "GNN architecture was not used because all existing GCL papers already use GCNs, and it was deemed sufficient to refer to them. Additionally, Local-GCL, the most similar method to the paper, uses GCN but also reports results using MLP architecture, which the paper' Graph CNE outperforms on all datasets. Furthermore, GCNs are not very appropriate for node-level graph learning as they cannot process one node at a time, though they can be suitable for graph-level learning.", "category": "Motivation_Analysis"}, {"question": "Why did the authors choose to evaluate the models using Euclidean distance-based $k$NN recall/accuracy instead of cosine distance, given that they used cosine distance and Gaussian similarity kernel for $d$=128 embeddings?", "answer": "The authors also tried using cosine-metric kNN for evaluation but found it did not significantly affect the results. For simplicity, they reported only the Euclidean kNN evaluation and have added a clarification about this in the text.", "category": "Motivation_Analysis"}, {"question": "What evidence supports the claim that Graph CNE performs comparably to state-of-the-art graph contrastive learning algorithms, given that Local-GCL appears better in Table 2?", "answer": "While some GCL methods outperform Graph CNE by a small margin in Table 2, Graph CNE is the best on one dataset (PUB). The claim 'comparably' means it performs close to the best, not necessarily the best. Additionally, when comparing to GCL methods using MLP architecture, Graph CNE outperforms MLP-based GCL results on all datasets, as demonstrated by results from Guo et al., NeurIPS 2023, which have been added to Table 2.\nhttps://arxiv.org/abs/2311.02687", "category": "Method_Comparison"}, {"question": "What process was used to determine the hyperparameters for graph t-SNE and graph CNE?", "answer": "For graph t-SNE, all default hyperparameters of openTSNE were used. For graph CNE, default hyperparameters of CNE were mostly used, but the number of epochs and the number of negative samples were increased to ensure convergence and improve the InfoNCE approximation (as per Damrich et al. 2023). Additionally, the batch size was adapted based on pilot experiments to ensure convergence.", "category": "Experimental_Setup"}, {"question": "Why did the authors choose to use MLP instead of GCN for transforming node features, and what evidence supports this decision?", "answer": "The authors chose MLP over GCN because all existing GCL papers use GCNs, making it sufficient to refer to them. For instance, Local-GCL, the most similar method to theirs, uses GCN but also reports results using MLP architecture, which the authors include in Table 2. Their Graph CNE outperforms Local-GCL on all datasets. Additionally, they argue that GCNs are not suitable for node-level graph learning as they cannot process one node at a time, whereas MLPs are more appropriate. They also cite another paper (Guo et al., NeurIPS 2023：https://arxiv.org/abs/2311.02687) that reports MLP-based results, which they added to Table 2, showing that Graph CNE outperforms both MLP-based GCL results on all datasets.", "category": "Motivation_Analysis"}, {"question": "Does the rebuttal provide additional evidence from another paper showing that Graph CNE outperforms MLP-based GCL results on all datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the sentence suggesting that running CNE for more epochs and/or with higher $m$ could yield better results retained in the paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the performance difference between Graph CNE and some GCL methods in Table 2 significant, according to the rebuttal?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Table 2 show that Graph CNE consistently outperforms all state-of-the-art graph contrastive learning algorithms?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the baseline methods in Table 2 use the same train/val/test split settings as the proposed method?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the number of epochs (100) found to be sufficient for model convergence in the updated experiments?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the statements regarding potential improvements with more epochs or higher $m$ value still present in the paper?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "rlloVZoKrX", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "SQL-PaLM: Improved large language model adaptation for Text-to-SQL", "authors": ["Ruoxi Sun", "Sercan O Arik", "Alexandre Muzio", "Lesly Miculicich", "Satya Kesav Gundabathula", "Pengcheng Yin", "Hanjun Dai", "Hootan Nakhost", "Rajarishi Sinha", "Zifeng Wang", "Tomas Pfister"], "abstract": "Text-to-SQL, the process of translating natural language into Structured Query Language (SQL), represents a transformative application of large language models (LLMs), potentially revolutionizing how humans interact with data. This paper introduces the SQL-PaLM framework, a comprehensive solution for understanding and enhancing Text-to-SQL using LLMs, using in the learning regimes of few-shot prompting and instruction fine-tuning. With few-shot prompting, we explore the effectiveness of consistency decoding with execution-based error filtering. With instruction fine-tuning, we delve deep in understanding the critical paradigms that influence the performance of tuned LLMs. In particular, we investigate how performance can be improved through expanded training data coverage and diversity, synthetic data augmentation, and integrating query-specific database content. We propose a test-time selection method to further refine accuracy by integrating SQL outputs from multiple paradigms with execution feedback as guidance. Additionally, we tackle the practical challenge of navigating intricate databases with a significant number of tables and columns, proposing efficient techniques for accurately selecting relevant database elements to enhance Text-to-SQL performance. Our holistic approach yields substantial advancements in Text-to-SQL, as demonstrated on two key public benchmarks, Spider and BIRD. Through comprehensive ablations and error analyses, we shed light on the strengths and weaknesses of our framework, offering valuable insights into Text-to-SQL's future work.", "keywords": "", "qa_pairs": [{"question": "What are the specific details of the LoRA fine-tuning experiments conducted in the study?", "answer": "The LoRA fine-tuning experiments involve incorporating trainable linear low-rank modules into the query and value projections of each self-attention layer, following Hu et al. (2021). The rank of LoRA is set to 32, the learning rate is 1e-4, and the model architecture used is the Gecko PaLM model.", "category": "Experimental_Setup"}, {"question": "What are the details of LoRA fine-tuning, including the rank, learning rate, and model architecture used?", "answer": "The details of LoRA fine-tuning include incorporating trainable linear low-rank modules into the query and value projections of each self-attention layer, following Hu et al. (2021). The rank of LoRA is set to 32, the learning rate is 1e-4, and the model architecture is the Gecko PaLM model.", "category": "Experimental_Setup"}, {"question": "Why did the authors choose to focus on tuning large language models (LLMs) instead of smaller, more efficient models like CodeS for Text2SQL systems?", "answer": "The authors chose to focus on tuning large language models (LLMs) because their paper aims to push the best achievable performance in Text2SQL systems, filling a gap in the field where most current research focuses on smaller models due to computational constraints. They highlight that larger LLMs face exacerbated generalization and computational challenges, making this an underexplored area. Additionally, they believe that achieving human-level accuracy is crucial for real-world impact, and tuning larger LLMs is currently closer to this goal compared to smaller models. While larger LLMs are more expensive, the authors argue that their performance improvements (e.g., Unicorn yields 55% vs. Gecko's 33%) make them lucrative for most applications, considering the time value of humans.", "category": "Motivation_Analysis"}, {"question": "Why was the Valid Efficiency Score (VES) not initially evaluated for BIRD, and what are the results of its subsequent evaluation?", "answer": "The initial focus was on execution accuracy (EX) as it is a critical measurement for comparison. However, recognizing the value of VES for combining efficiency and accuracy, authors evaluated SQL-PaLM for VES and found it performed 15% better, indicating superior performance in both efficacy and accuracy.", "category": "Experimental_Exposition"}, {"question": "How does the prompt design used in the study generalize across different datasets like Spider and BIRD, and across various LLMs?", "answer": "The same prompt design is used for both BIRD and SPIDER, as clarified in Section 8. This approach allows combining BIRD and SPIDER to train a LLM, resulting in better performance when tested on either dataset compared to individual training. The prompt design is dataset agnostic and robust to distribution shifts.", "category": "Method_Mechanics"}, {"question": "How does the paper address synthetic data augmentation in cases without manual annotations?", "answer": "The paper focuses on a 'conservative' synthetic data approach where the schema and questions remain unchanged, and only the SQL is modified. This allows for the evaluation of synthetic data quality by re-using the current evaluation framework, thus eliminating confounding factors related to unknown data quality. ", "category": "Method_Mechanics"}, {"question": "Why does the comparison against state-of-the-art baselines use both Spider and BIRD data for fine-tuning, while the baselines do not?", "answer": "The paper focuses on investigating different factors influencing Text-to-SQL performance, and combining BIRD and SPIDER is one investigation of diverse training data that helps with fine-tuning performance. While performance comparisons are provided for readers to understand the results in Table 24 and 25, the paper does not aim to achieve or claim state-of-the-art performance.", "category": "Motivation_Analysis"}, {"question": "What baselines or comparisons are used in Table 17 and Table 18, and how are they defined?", "answer": "Table 17 compares soft and hard column selection in a supervised fine-tuning setup. Hard column selection involves including only selected columns and removing unselected ones, while soft column selection emphasizes selected column descriptions in the prompt without removing unselected columns. Table 18 focuses on ablation studies for soft column selection, including comparisons of program-aided and retrieval-based methods for identifying selected columns, ground truth column selection (oracle), and the inclusion of full column descriptions. The manuscript has been updated to clarify these descriptions.", "category": "Method_Comparison"}, {"question": "How is the similarity score in equation 14 computed, and is it determined by the LLM?", "answer": "The similarity score in equation 14 is computed by the LLM, which generates different SQL queries and outputs the similarity scores. Detailed information on the prompt design can be found in section 'A.5.1 Synthetic data prompt design'. The manuscript has been updated to clarify this process.", "category": "Method_Disambiguation"}, {"question": "What updates have been made to the manuscript to address the discussion about LLM-agents, specifically MAC-SQL, and how does the performance of SQL-PaLM compare to MAC-SQL on BIRD and SPIDER datasets?", "answer": "The manuscript has been updated to include a discussion of the 'LLM agent line' of work, specifically MAC-SQL, in the 'related work' section. Additionally, performance evaluations comparing SQL-PaLM and MAC-SQL are provided in Table 24 BIRD and Table 25 SPIDER. The results show that SQL-PaLM outperforms MAC-SQL on both BIRD and SPIDER datasets, with SQL-PaLM achieving an execution accuracy of 61.7% on BIRD compared to MAC-SQL's 59.39%, and an execution accuracy of 87.3% on SPIDER compared to MAC-SQL's 86.75%.\n**Table 1: BIRD**: |                     |                             | |:-------------------:|:---------------------------:| |                     | Execution Accuracy (EX) | | SQL-PaLM (Ours) |             **61.7**            | |   [1] MAC-SQL   |            59.39            | |                     |                             | **Table 2: SPIDER**: |                 |                         |                          | |-----------------|:-------------------------:|:--------------------------:| |                 | Execution Accuracy (EX) | Test Suite Accuracy (TS) | | SQL-PaLM (Ours) |           **87.3**          |           83.5           | | [1] MAC-SQL     |          86.75          |               -           | |                 |                         |                          |", "category": "Method_Comparison"}, {"question": "What previous work on synthesizing data for Text-to-SQL tasks, and how does SQL-PaLM compare to these methods?", "answer": "Ddiscussion of previous work on synthesizing data for Text-to-SQL tasks in the 'Related Work' section (Sec. 3.2). Specifically, [2] GRAPPA and [3] TaBERT propose pretraining frameworks to jointly train structured schema of DB tables and questions. [2] uses a synchronous context-free grammar to extract question-SQL templates, while [4] emphasizes 'strong-typing' and 'key relations' in its synthetic framework. [5][6] synthesize SQL using context-free grammar rules and generate corresponding natural questions. SQL-PaLM outperforms these methods on both BIRD and SPIDER datasets, as shown in Table 3 and Table 4. For example, SQL-PaLM achieves an Execution Accuracy (EX) of 61.7 on BIRD, compared to 58.47 for [7] CodeS, and an EX of 87.3 on SPIDER, compared to 85.4 for [7] CodeS.\n\nTable 3: BIRD \n| Method                     | Execution Accuracy (EX) |\n| -------------------------- | :---------------------: |\n| **SQL-PaLM (Ours)**        |         **61.7**        |\n| \\[2] GraPPa                |            -            |\n| \\[3] TaBERT                |            -            |\n| \\[4] Importance ...        |            -            |\n| \\[5] Learning to ...       |            -            |\n| \\[6] Data Augmentation ... |            -            |\n| \\[7] CodeS                 |          58.47          |\n\nTable 4 SPIDER\n| Method              | Execution Accuracy (EX) | Test Suite Accuracy (TS) | Exact Match (EM) |\n| ------------------- | :---------------------: | :----------------------: | :--------------: |\n| **SQL-PaLM (Ours)** |         **87.3**        |         **83.5**         |         -        |\n| \\[2] GraPPa         |            -            |             -            |       73.4       |\n| \\[3] TaBERT         |            -            |             -            |       64.5       |\n| \\[4]                |           81.4          |             -            |       76.1       |\n| \\[5]                |           72.5          |             -            |       71.8       |\n| \\[6]                |            -            |             -            |       68.2       |\n| \\[7] CodeS          |           85.4          |           80.3           |         -        |\n ", "category": "Method_Comparison"}, {"question": "Did earlier works [2,3] include both EM and EX evaluation metrics?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper show that SQL-PaLM outperforms MAC-SQL in Test Suite Accuracy on the SPIDER dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the Valid Efficiency Score (VES) evaluated for BIRD in the initial manuscript?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the execution accuracy of SQL-PaLM higher than that of MAC-SQL on both BIRD and SPIDER datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the references [2-7] in the 'Related Work' section regarding synthetic data for the text-to-SQL task?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the evaluation metrics Exact Match (EM), Execution Accuracy (EX), and Test Suite Accuracy (TS) explained?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the same prompt design used for both BIRD and SPIDER datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Has the Valid Efficiency Score (VES) been included in the updated manuscript for SQL-PaLM, showing a 15% improvement?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper provide a performance evaluation comparing MAC-SQL with SQL-PaLM on BIRD and SPIDER datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper claim that the prompt design is robust to distribution shifts and agnostic to specific datasets?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "rAnB7JSMXL", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Patches Are All You Need?", "authors": ["Asher Trockman", "J Zico Kolter"], "abstract": "Although convolutional neural networks have been the dominant architecture for computer vision for many years, Vision Transformers (ViTs) have recently shown promise as an alternative. Subsequently, many new models have been proposed which replace the self-attention layer within the ViT architecture with novel operations (such as MLPs), all of which have also been relatively performant. We note that these architectures all share a common component--the patch embedding layer--which enables the use of a simple isotropic template with alternating steps of channel- and spatial-dimension mixing. This raises a question: is the success of ViT-style models due to novel, highly-expressive operations like self-attention, or is it at least in part due to using patches? In this paper, we present some evidence for the latter: specifically, we propose the ConvMixer, an extremely simple and parameter-efficient fully-convolutional model in which we replace the self-attention and MLP layers within the ViT with less-expressive depthwise and pointwise convolutional layers, respectively. Despite its unusual simplicity, ConvMixer outperforms the ViT, MLP-Mixer, and their variants for similar data set sizes and parameter counts, in addition to outperforming classical vision models like ResNet. We argue that this contributes to the evidence that patches are sufficient for designing simple and effective vision models. Our code is available at https://github.com/locuslab/convmixer.", "keywords": "", "qa_pairs": [{"question": "What insights can be gained from the finding that patches are sufficient for representation, and how does this contribute to the development of more efficient architectures?", "answer": "The key insight is that a wide variety of operations, including convolution, work well within patch-based architectures, rather than concluding that convolution is universally better than self-attention or MLPs. However, determining the best universal operation would require a more extensive and computationally expensive study. While not all modalities can be decomposed into patches, the focus of this work is on computer vision. Convolution can be applied to data with a notion of order in time or space, and it is possible that convolutional models may perform as well as attention-based models in some language modeling tasks, though this remains uncertain. [1, 2]\n[1] https://arxiv.org/abs/2210.09298\n\n[2] https://arxiv.org/abs/2008.02496", "category": "Experimental_Analysis"}, {"question": "Why were the 14 training runs in Table 2 chosen, and what is the significance of the rows without colored dots?", "answer": "The 14 training runs in Table 2 were chosen as part of an ad-hoc exploration to determine suitable configurations for width, depth, kernel size, patch size, training time, and activation function for the architecture. Due to computational constraints, the exploration was not systematic. Authors started out with some smaller ConvMixers trained for fewer epochs, and worked the way up to larger ones. For example, authors were curious if authors could surpass 80% top-1 accuracy on ImageNet using 14x14 patches rather than 7x7, and ConvMixer-1536/24 was best guess at a configuration that might work, based on previous runs (and it did…). The rows without colored dots represent additional training runs that were part of the exploration, and all results in the table are from actual training runs, with no trial runs excluded. The table is included in the appendix because the exploration was unstructured, but it still provides valuable insights.", "category": "Experimental_Setup"}, {"question": "To what extent was hyperparameter optimization (HPO) performed on ImageNet, and how does the lack of HPO affect the reported accuracies and comparison with other models?", "answer": "No hyperparameter optimization (HPO) was performed on ImageNet. The parameters used were inspired by work on CIFAR-10, defaults from the timm library, and initial guesses, later improved by parameters from 'ResNet Strikes Back[1]'. The lack of HPO establishes a lower bound on ConvMixer’s advantage, but a more realistic comparison with other models using HPO was not feasible due to computational constraints.\n[1] https://arxiv.org/pdf/2110.00476.pdf", "category": "Experimental_Analysis"}, {"question": "Was hyperparameter optimization (HPO) performed on ImageNet as part of the ConvMixer experiments?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was a plot showing the inference speed and accuracy trends of ConvMixer at varying image resolutions added to the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper originally include a study on the inference speed and accuracy trends of ConvMixer at varying image resolutions?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "4nPswr1KcP", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers", "authors": ["Andreas Peter Steiner", "Alexander Kolesnikov", "Xiaohua Zhai", "Ross Wightman", "Jakob Uszkoreit", "Lucas Beyer"], "abstract": "Vision Transformers (ViT) have been shown to attain highly competitive performance for a wide range of vision applications, such as image classification, object detection and semantic image segmentation. In comparison to convolutional neural networks, the Vision Transformer's weaker inductive bias is generally found to cause an increased reliance on model regularization or data augmentation (``AugReg'' for short) when training on smaller training datasets. We conduct a systematic empirical study in order to better understand the interplay between the amount of training data, AugReg, model size and compute budget. As one result of this study we find that the combination of increased compute and AugReg can yield models with the same performance as models trained on an order of magnitude more training data: we train ViT models of various sizes on the public ImageNet-21k dataset which either match or outperform their counterparts trained on the larger, but not publicly available JFT-300M dataset.\n", "keywords": "", "qa_pairs": [{"question": "Is it feasible to conduct toy experiments with fewer training epochs to obtain consistent conclusions about the optimal AugReg settings, given the differences observed in model behavior across training durations?", "answer": "No, it is not feasible to conduct toy experiments with fewer training epochs to obtain consistent conclusions about the optimal AugReg settings. Insights from large sweeps indicate that the optimal strength of AugReg varies across diverse settings. For example, smaller models like Ti/16 show similar patterns when trained for 30 or 300 epochs. But already modestly sized models (in terms of compute, not parameters) like B/32 behave completely differently when trained for 30 epochs (best upstream validation accuracy without AugReg) vs. 300 epochs (worst upstream validation accuracy without AugReg).", "category": "Experimental_Exposition"}, {"question": "How does the effect of strong AugReg vary depending on model capacity, dataset size, and number of epochs, and what is the value of the study in reporting these differences?", "answer": "The effect of strong AugReg varies depending on model capacity, dataset size, and number of epochs. For example, strong AugReg increases validation accuracy from 43 to 51 for the L/16 model when training for 300 epochs on ImageNet-21k but decreases it from 46 to 40  for the S/16 model on the same dataset. Additionally, it does not change validation accuracy for L/16 would increase the validation accuracy from 45 to 48 when training for only 30 epochs, where 'light2' augmentation is more effective. The value of the study lies in reporting these differences, reminding the community that the exact effect of AugReg depends on these parameters, and providing the best vanilla ViT checkpoints pre-trained on ImageNet-21k for various applications.", "category": "Experimental_Exposition"}, {"question": "Why does Figure 5 show both positive and negative values in the left figure, even though adapting all pre-trained models should theoretically achieve the best performance?", "answer": "Figure 5 shows that selecting the best model by validation accuracy does not always result in the best test accuracy due to random fluctuations in the validation/test set scores, particularly evident in the Resisc45 dataset. Additionally, for ImageNet-1k, there is a data leak because ImageNet-21k contains the ImageNet-1k minival subset, which biases model selection toward those that memorize upstream data rather than generalize well. To address this, a different set of ImageNet-like images (ImageNet-v2) is used for model selection.", "category": "Experimental_Analysis"}, {"question": "Why did the authors not conduct experiments on object detection and segmentation, and what additional experiments were performed instead?", "answer": "The authors did not conduct experiments on object detection and segmentation because no clear best setup has emerged yet for these tasks with vanilla ViTs, and adopting different approaches would significantly extend the scope of the study. Instead, they performed new experiments on multi-modal image-text retrieval, which are detailed in a new appendix section titled 'Using recommended checkpoints for other computer vision tasks.' This section includes contrastive training of some checkpoints and reports zero-shot classification numbers and image-text retrieval performance.", "category": "Experimental_Setup"}, {"question": "How do the authors address the unexpected finding regarding AugReg's effect on overfitting?", "answer": "The authors acknowledge that AugReg's role in mitigating overfitting is already known and not a novel contribution. They emphasize their contribution lies in quantifying AugReg's effects across various settings, revealing a nuanced landscape of optimal AugReg strength, which they assert is a significant and previously unknown insight for ViT research.", "category": "Experimental_Analysis"}, {"question": "Are toy experiments with fewer training epochs expected to yield consistent conclusions about ideal AugReg settings based on the insights from large sweeps?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "ZPQhzTSWA7", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "A Simple Convergence Proof of Adam and Adagrad", "authors": ["Alexandre Défossez", "Leon Bottou", "Francis Bach", "Nicolas Usunier"], "abstract": "We provide a simple proof of convergence covering both the Adam and Adagrad adaptive optimization algorithms when applied to smooth (possibly non-convex) objective functions with bounded gradients. We show that in expectation, the squared norm of the objective gradient averaged over the trajectory has an upper-bound which is explicit in the constants of the problem, parameters of the optimizer, the dimension $d$, and the total number of iterations $N$. \nThis bound can be made arbitrarily small, and with the right hyper-parameters, Adam can be shown to converge with the same rate of convergence $O(d\\ln(N)/\\sqrt{N})$. When used with the default parameters, Adam doesn't converge, however, and just like constant step-size SGD, it moves away from the initialization point faster than\nAdagrad, which might explain its practical success.\nFinally, we obtain the tightest dependency on the heavy ball momentum decay rate $\\beta_1$ among all previous convergence bounds for non-convex Adam and Adagrad,\nimproving from $O((1-\\beta_1)^{-3})$ to $O((1-\\beta_1)^{-1})$. ", "keywords": "", "qa_pairs": [{"question": "What is the purpose of including the toy experiment, and how does it relate to real-life problems and the use of $\\beta_2$ values?", "answer": "The toy experiment is designed to illustrate the worst-case dependency in $1/\\sqrt(1 - \\beta_2)$. In most real-life problems, this part of the bound does not dominate, as demonstrated by the CIFAR-10 experiment and the typical use of $\\beta_2$ values ranging from 0.9 to 0.999.", "category": "Motivation_Analysis"}, {"question": "Can the proposed approach be extended to avoid the bounded diameter assumption in a convergence proof for convex functions?", "answer": "The proof could be extended to obtain a bound on the primal sub optimality $F(x_n) - F_*$ for the convex case instead of a bound on the squared norm of the gradients. However, it still requires a uniform bound on the gradients, limiting its scope. This approach would apply to problems like smoothed L1 regression or logistic regression without L2 penalty, where the original Adagrad proof would require the extra bounded diameter assumption.", "category": "Method_Mechanics"}, {"question": "Why does the sampling distribution for iterates in Equation 8 heavily discount the most recent iterations?", "answer": "The heavy discounting of the most recent iterations in the sampling distribution can be understood by considering the limit $eta_1 \\rightarrow 1$. In this scenario, the last gradient $\\nabla F(x_n)$ has minimal influence on the next iterate $x_{n+1}$ due to the slow updates caused by momentum. Consequently, it has little impact on $F(x_{n+1})$. Non-convex convergence proofs rely on the principle that each iteration either results in a small gradient or a decrease in the primal objective $F(x)$. With large momentum values, the immediate effect of the latest gradient is delayed, causing the primal value to decrease gradually over several iterations. Thus, the impact of sampling the last iterations is limited over a sufficient number of iterations. This explanation has been added to Section 4.2.", "category": "Motivation_Analysis"}, {"question": "What is the definition of 'optimality' used in Section 4.3, particularly in the context of the choices $\\alpha = N^{-a}$ and $\\beta_2 = 1 - N^{-b}$, and how does it relate to minimizing the upper bound from Equation 10?", "answer": "Optimality in Section 4.3 is defined as finding the largest number $q$ such that the bound obtained is $O(\\ln(N)/ N^q)$ for hyper-parameters $\\alpha(N)$ and $\\beta_2(N)$. Convergence is achievable only if $\\alpha(N) \\rightarrow 0$ and $\\beta_2 \\rightarrow 1$ as $N \\rightarrow \\infty$, motivating the assumption of an asymptotic development in $N$. This analysis is standard in optimization, as referenced in [Bach and Moulines 2011].", "category": "Concept_Understanding"}, {"question": "Why does the removal of the momentum corrective term in the Adam algorithm not significantly impact the optimization process, and how does this compare to removing the corrective term for the average of the squared gradients?", "answer": "The momentum corrective term decays in approximately 10 iterations, and its removal is equivalent to having a learning rate warmup, which is harmless for optimization. In contrast, removing the corrective term for the average of the squared gradients would multiply the effective learning rate by a factor of around 30 during the first ~1000 epochs, which could destabilize early training. Experimental results on CIFAR-10 confirm that dropping the momentum corrective term has no impact on optimization dynamics, while dropping the denominator corrective term strongly impacts the dynamics.", "category": "Experimental_Analysis"}, {"question": "How does the paper address the issue of smooth interpolation between Adagrad and Adam when a constant $\\alpha$ is used, and why does the analysis seem to handle the edge case of $\beta_2 = 1$ differently?", "answer": "The paper shows in Section 4.3 that using $\\alpha = 1/\\sqrt{N}$ and $\\beta_2 = 1 - 1/N$ results in the same optimal convergence rate as Adam, suggesting equivalence to Adagrad under these hyperparameters. The apparent blow-up when $\\beta_2 = 1$ and $\\alpha = 1/\\sqrt{N}$ is due to neglected terms in the proof, which have minimal impact for $\\beta_2 < 1$. n particular, p.10: 'Given that $\\beta_2 < 1$, $\\alpha_n \\leq \\frac{\\alpha}{\\sqrt{1 - \\beta_2}}$', while taking the limit as you do would require the paper to keep a $\\sqrt{1 - \\beta_2^n}$ term. The paper addresses $\\beta_2 = 1$ separately, maintaining the desired optimal rate by setting $\\beta_2 = 1 - 1/N$.", "category": "Method_Mechanics"}, {"question": "What are the differences between the analyses of momentum SGD rates in the current paper and those presented in [Defazio, 2020] and [Liu et al., 2020]?", "answer": "The paper primarily comments on the work by Liu et al. (2020), which has a similar dependency in $β1$ as the step size $α∼ (1 - β1)$, with their Theorem 1 including a term in $O(1 / α)$. Liu et al.'s work requires less strict assumptions. Defazio (2020) does not provide a clear bound. The initial release of the paper did not include this comparison as the results at the time were much worse, and the claim in the abstract has been removed with an update to related work to acknowledge more recent results with weaker assumptions.", "category": "Experimental_Analysis"}, {"question": "Have the authors conducted experiments comparing Adam with theoretical optimal parameters to Adam with practical default parameters, and what insights do they provide regarding the practical effectiveness of Adam?", "answer": "The authors have provided an experiment on Figure 2 to show the difference due to the corrective term change in Section 2.2. They explain that comparing Adam with optimal rate parameters to Adam in practice is unnecessary because the former is equivalent to Adagrad, which is suboptimal for deep learning. In Section 4.3, they offer insights that Adam with default parameters behaves like constant step size SGD, which doesn't converge but often yields the best practical results by rapidly reducing the term in $F(x_0) - F_*$. They also note that certain hyperparameters can make Adam converge, but these are not recommended for practical training.", "category": "Experimental_Setup"}, {"question": "Why was the scalar version of Adagrad and Adam not included in the analysis?", "answer": "The scalar version of Adagrad was previously covered by Ward et al. 2019 and is primarily used as a simple milestone for adaptive proofs. The diagonal version is more commonly used in practice, making further analysis of the scalar version redundant.", "category": "Motivation_Analysis"}, {"question": "How does the difference in step-size formulation between the proposed method and Adam, particularly regarding the scaling of $m_{n-1}$, impact the algorithm and its analysis?", "answer": "The proposed recurrent formula is not equivalent to the one in Adam, as it would introduce a $(1 - \\beta_1)^n$ term不同于Adam中的单一$(1 - \\beta_1)$因子。我们在第2.2节中澄清了推导过程，消除了对公式的疑虑。经过几次$1 / (1 - \\beta_1)$迭代后，两者无差异。在非凸问题中，前几次迭代可能导致不同解，但缺少偏置修正项等同于学习率预热，这是训练DNN的常见做法。我们在第2.2节中详细介绍了带修正的原始Adam步长，并逐项研究。通过在CIFAR-10上比较去除任一修正项的训练轨迹，显示去除动量修正项对常用值无影响，而去除$v_n$修正项则显著改变训练损失和梯度范数（见图2）。", "category": "Experimental_Analysis"}, {"question": "Does the paper claim to offer the proof with the weakest assumptions among recent works?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the authors provide a proof for at least one momentum method under the assumption of uniformly bounded stochastic gradients, confirming the dependency on $β1$ remains unchanged?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "ehfRiF0R3a", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Voyager: An Open-Ended Embodied Agent with Large Language Models", "authors": ["Guanzhi Wang", "Yuqi Xie", "Yunfan Jiang", "Ajay Mandlekar", "Chaowei Xiao", "Yuke Zhu", "Linxi Fan", "Anima Anandkumar"], "abstract": "We introduce Voyager, the first LLM-powered embodied lifelong learning agent in an open-ended world that continuously explores, acquires diverse skills, and makes novel discoveries without human intervention in Minecraft. Voyager consists of three key components: 1) an automatic curriculum that maximizes exploration, 2) an ever-growing skill library of executable code for storing and retrieving complex behaviors, and 3) a new iterative prompting mechanism that incorporates environment feedback, execution errors, and self-verification for program improvement. Voyager interacts with GPT-4 via blackbox queries, which bypasses the need for model parameter fine-tuning. The skills developed by Voyager are temporally extended, interpretable, and compositional, which compounds the agent’s capability rapidly and alleviates catastrophic forgetting. Empirically, Voyager demonstrates strong in-context lifelong learning capabilities. It outperforms prior SOTA by obtaining 3.1x more unique items, unlocking tech tree milestones up to 15.3x faster, and traveling 2.3x longer distances. Voyager is able to utilize the learned skill library in a new Minecraft world to solve novel tasks from scratch, while other techniques struggle to generalize.", "keywords": "", "qa_pairs": [{"question": "What are 'self-ask questions' and how are they formulated in Section 2.1?", "answer": "Self-ask questions are formulated by informing GPT-3.5 about the agent's current state and task history, then instructing it to create questions directly related to specific Minecraft concepts. Examples include 'How do you make a wooden pickaxe?' and 'What are the blocks that can be discovered in a sparse jungle?' These questions are self-explanatory and their responses provide context for brainstorming new tasks.", "category": "Concept_Understanding"}, {"question": "To what extent can the performance of the high-level action API demonstrated in Minecraft be replicated in other domains?", "answer": "Voyager's method is general-purpose and does not make strong assumptions about the environment. It can be adapted to other domains where code serves as the action space and comes with a well-explained API, such as Webshop and WebArena. Authors choose to focus on Minecraft because it is a highly representative environment that requires long-horizon planning, compositional reasoning, and extensive exploration. ", "category": "Experimental_Analysis"}, {"question": "What do the shaded regions in Figure 1 represent, and how are the items shown related to the multiple experimental runs, particularly regarding the acquisition of the diamond tool?", "answer": "The shaded regions in Figure 1 represent the standard deviation of items acquired by each agent over three experimental runs. The items shown highlight significant and noteworthy acquisitions, with Voyager being the only method that obtained a diamond tool in one of the three runs, as clarified in the figure's caption.", "category": "Experimental_Exposition"}, {"question": "How does the proposed approach generalize to completely new games that are not mentioned on the web and have equally complex tech trees?", "answer": "The proposed approach uses GPT-4 to extract game mechanics from wiki pages or handbooks, providing insights into unfamiliar environments. Additionally, Voyager requires environment observation APIs and low-level control primitives to be explained or demonstrated within the LLM’s prompt, as seen in Inner Monologue[1] and Code as Policies[2].\n[1] Huang et al., Inner Monologue: Embodied Reasoning through Planning with Language Models, CoRL 2022.\n\n[2] Liang et al., Code as Policies: Language Model Programs for Embodied Control, ICRA 2023.\n", "category": "Method_Mechanics"}, {"question": "How does the automatic curriculum differ from a manually constructed curriculum derived from the tech tree in terms of scalability and task recommendation?", "answer": "The automatic curriculum proposes tasks based on the current skill level and world state, making it scalable for open-ended exploration. In contrast, a manual curriculum requires access to the full tech tree, is not scalable, and lacks awareness of the current skill level and world state, limiting its ability to recommend the most pertinent tasks consistently.", "category": "Method_Comparison"}, {"question": "How does the `craftItem` function in Fig. 5 Left handle the creation of multiple sticks, and does Voyager engage in explicit code refactoring for repetitive tasks?", "answer": "The third parameter in the `craftItem` function specifies the number of iterations for the recipe execution. In Minecraft, executing the stick recipe once consumes 2 logs and yields 4 sticks, which exceeds the required amount of 2. Voyager does not engage in explicit code refactoring, but repetitive tasks in varied world states might be suggested by the automated curriculum, potentially overwriting old skills with new ones.", "category": "Method_Mechanics"}, {"question": "In the zero-shot generalization experiment, is the skill library fixed, and does the evaluation allow for open-ended exploration?", "answer": "The skill library is fixed to demonstrate the effectiveness of the pre-trained skill library for unseen tasks. Open-ended exploration is not conducted, as GPT-4 is used to break down unseen tasks into subgoals.", "category": "Experimental_Setup"}, {"question": "In Figure 10, is the human feedback classified as type (1), where humans are used as critics?", "answer": "In the Nether portal example, humans are used as critics, providing feedback on the correctness of the portal construction. For instance, they point out that the portal size is incorrect and that a valid 10-obsidian portal should have 2 blocks on the top and bottom edges and 3 blocks on the sides. They also suggest using a different material such as cobblestone for the four corners, offering guidance not only on structure but also on material aesthetics.\n\nIn contrast, in the house-building example, humans act as curricula, guiding Voyager step-by-step through the construction process. The instructions include starting from Y = 0, building a 5×5 oak wood structure 3 blocks high, and continuing the task at (615, 79, 1206) by adding 9 wooden stairs as decorations (placed at each corner, the center of each edge, and the building center). In the final stage, humans instruct the addition of 8 windows (one on each wall, placed 1 block above the ground and 1 block apart) and a door centered on one wall. These structured and detailed instructions reflect a highly organized and actionable form of human teaching.", "category": "Method_Disambiguation"}, {"question": "How can the limitations of priors encoded in pretrained language models for new games be addressed to improve the transferability of the exploration phase and automatic curriculum?", "answer": "To address these limitations, GPT-4 can be used to extract game mechanics from wiki pages or handbooks, providing insights into the domain. Additionally, Voyager requires environment observation APIs and low-level control primitives to be explained or demonstrated within the LLM's prompt, similar to Inner Monologue[1] and Code as Policies[2].\n[1] Huang et al., Inner Monologue: Embodied Reasoning through Planning with Language Models, CoRL 2022.\n\n[2] Liang et al., Code as Policies: Language Model Programs for Embodied Control, ICRA 2023.\n", "category": "Method_Mechanics"}, {"question": "How does the automatic curriculum differ from a manual curriculum that uses the full tech tree, and what are the limitations of the manual curriculum?", "answer": "The automatic curriculum proposes tasks based on the current skill level and world state, making it scalable for open-ended exploration. In contrast, a manual curriculum requires access to the full tech tree and is not scalable. The manual curriculum has been expanded with additional tasks from the tech tree, but it lacks awareness of the current skill level and world state, limiting its ability to consistently recommend the most pertinent tasks.", "category": "Method_Comparison"}, {"question": "How does Voyager's learning process differ from being directly instructed by Chat-GPT, and what novel aspects of learning does it introduce?", "answer": "Instead of being a 'pro player', GPT-4 is an LLM that does not have open-ended exploration and Minecraft-playing skills out of the box. Voyager is an advanced LLM-based agent algorithm with blackbox access to GPT-4, which serves as a base reasoning engine. Unlike being directly instructed by Chat-GPT, Voyager iteratively constructs a codebase of skills through open-ended exploration. This codebase acts as explicit 'learned parameters' acquired by coding, allowing Voyager to continuously improve without modifying the underlying GPT-4, demonstrating 'in-context lifelong learning capability'.", "category": "Method_Comparison"}, {"question": "What are the key differences between Voyager and the baseline methods ReAct, Reflexion, and AutoGPT, and how does Voyager incorporate all components to deliver the best performance?", "answer": "The key differences are outlined in Table A.3, which shows that Voyager is the only method that incorporates all components such as Chain-of-Thought, Self Verification, Environment Feedback, Execution Errors, Agent State, Skill Library, and Automatic Curriculum. These components work in tandem to deliver the best performance.\n|                           | **ReAct \\[1]** | **Reflexion \\[2]** | **AutoGPT \\[3]** | **Voyager** |\n| ------------------------- | :------------: | :----------------: | :--------------: | :---------: |\n| **Chain-of-Thought \\[4]** |        ✓       |          ✓         |         ✓        |      ✓      |\n| **Self Verification**     |                |          ✓         |                  |      ✓      |\n| **Environment Feedback**  |        ✓       |          ✓         |         ✓        |      ✓      |\n| **Execution Errors**      |                |          ✓         |         ✓        |      ✓      |\n| **Agent State**           |        ✓       |          ✓         |         ✓        |      ✓      |\n| **Skill Library**         |                |                    |                  |      ✓      |\n| **Automatic Curriculum**  |                |                    |                  |      ✓      |\n[1] Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models, ICLR 2023.\n\n[2] Shinn et al., Reflexion: Language Agents with Verbal Reinforcement Learning, NeurIPS 2023.\n\n[3] Toran Bruce Richards. Significant-gravitas/auto-gpt: An experimental open-source attempt to make gpt-4 fully autonomous., 2023. URL https://github.com/Significant-Gravitas/Auto-GPT/tree/master.\n\n[4] Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, NeurIPS 2022.", "category": "Method_Comparison"}, {"question": "How can Voyager be applied to other domains, particularly in robotics, and what potential impact could it have on research in these areas?", "answer": "Voyager's method is general-purpose and can be adapted to domains where code serves as the action space and comes with a well-explained API, such as Webshop[5] and WebArena[6]. In robotics, prior works like Code as Policies[7] and ProgPrompt[8] assume the existence of a skill library or control primitives  (pick and place an object, open a container, switch on/off an object, etc.) and use LLMs to generate code for policy execution. LLMs can expand the skill library based on given control primitives and previously learned skills to solve real-world robotic manipulation tasks. Let’s say authors tell LLMs: 'LLMs are a cooking expert. Here are the skills that authors already acquired: find a microwave, put an object into a microwave, open and close a microwave. Please propose a novel task for me to pursue.' It would make sense for LLMs to output something like 'microwave some food' as the next task and write a program to compose the previously learned skills to synthesize a new one through iterative prompting.  In less familiar domains, LLMs can reference knowledge from wiki pages or handbooks (e.g., MineDojo [9] compiles a wiki dataset for Minecraft) to guide decision-making and identify next steps in uncharted environments.\n[5] Yao et al., WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents, NeurIPS 2022.\n\n[6] Zhou et al., WebArena: A Realistic Web Environment for Building Autonomous Agents, 2023.\n\n[7] Liang et al., Code as Policies: Language Model Programs for Embodied Control, ICRA 2023.\n\n[8] Singh et al., ProgPrompt: Generating Situated Robot Task Plans using Large Language Models, ICRA 2023.\n\n[9] Fan et al., MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge, NeurIPS 2022.\n\n\n", "category": "Method_Mechanics"}]}
{"id": "tzW948kU6x", "venue": "TMLR 2024", "paper_decision": "Withdraw", "title": "Diffusion on Graph: Augmentation of Graph Structure for Node Classification", "authors": ["Hari Sowrirajan", "Jingbo Yang", "Andrew Y. Ng", "Pranav Rajpurkar"], "abstract": "Graph diffusion models have recently been proposed to synthesize entire graphs, such as molecule graphs. Although existing methods have shown great performance in generating entire graphs for graph-level learning tasks, no graph diffusion models have been developed to generate synthetic graph structures, that is, synthetic nodes and associated edges within a given graph, for node-level learning tasks. Inspired by the research in the computer vision literature using synthetic data for enhanced performance, we propose Diffusion on Graph (DoG), which generates synthetic graph structures to boost the performance of GNNs. The synthetic graph structures generated by DoG are combined with the original graph to form an augmented graph for the training of node-level learning tasks, such as node classification and graph contrastive learning (GCL). To improve the efficiency of the generation process, a Bi-Level Neighbor Map Decoder (BLND) is introduced in DoG. To mitigate the adverse effect of the noise introduced by the synthetic graph structures, a low-rank regularization method is proposed for the training of graph neural networks (GNNs) on the augmented graphs. Extensive experiments on various graph datasets for semi-supervised node classification and graph contrastive learning have been conducted to demonstrate the effectiveness of DoG with low-rank regularization. The code of DoG is available at \\url{https://anonymous.4open.science/r/DoG/}.", "keywords": "", "qa_pairs": [{"question": "Why does the DoG model not apply to non-attributed graphs, and what is the role of node attributes in its training and generation processes?", "answer": "The DoG model focuses on augmenting attributed graphs for node-level learning tasks. It generates synthetic node attributes and edges using a Graph Autoencoder (GAE) that encodes node attributes into latent node features, and a latent diffusion model trained on these features. Node attributes are essential for both the training and generation processes of DoG, making it inapplicable to non-attributed graphs.", "category": "Method_Disambiguation"}, {"question": "How does the proposed method address the challenges of labeling and imputing features and connectivity when adding new nodes to the original graph?", "answer": "The proposed method employs a conditional Latent Diffusion Model (LDM)[1] based on Classifier-Free Guidance (CFG)[2] to generate labeled synthetic nodes. The class labels are converted to conditional embeddings and incorporated into the latent features of the synthetic nodes. To connect synthetic nodes to real nodes in the original graph, a novel Graph Autoencoder (GAE) is used. The GAE encodes both node attributes and neighbor maps into latent features. The decoder of the GAE then decodes these latent features into node attributes and neighbor maps, indicating the real nodes in the original graph that each synthetic node should connect to. Detailed descriptions of these processes are provided in Section 4.2 and Algorithm 2 in the appendix.\n[1] Rombach, Robin, et al. \"High-resolution image synthesis with latent diffusion models.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.\n\n[2] Ho, Jonathan, and Tim Salimans. \"Classifier-free diffusion guidance.\" arXiv preprint arXiv:2207.12598 (2022).", "category": "Method_Mechanics"}, {"question": "What are the recently added node-level data augmentation methods used as baselines for DoG, and how do they compare to DoG in performance?", "answer": "The recently added node-level data augmentation methods are TADA [5], LGGD [6], and GeoMix [7], which serve as baselines for DoG. DoG performs better than these methods individually, and combining DoG with these methods further improves their performance across all datasets. For example, GeoMix combined with DoG enhances GeoMix's node classification accuracy by 1.1% on Pubmed.\n| Methods |Cora|Citeseer|Pubmed |Coauthor CS|ogbn-arxiv |\n| :-- | :--: | :--: | :--: | :--: | :--: |\n| EXPHORMER Exphormer| 84.1| 72.9 | 83.2| 95.7| 72.4 |\n| DoG (EXPHORMER)| 85.7 ($\\uparrow$ 1.6)| 74.8 ($\\uparrow$ 1.9) | 84.6 ($\\uparrow$ 1.4)| 96.9 ($\\uparrow$ 1.2)| 73.4 ($\\uparrow$ 1.0)|\n| TADA (EXPHORMER) | 85.4 ($\\uparrow$ 1.3)| 74.4 ($\\uparrow$ 1.5) | 84.0 ($\\uparrow$ 0.8)| 96.3 ($\\uparrow$ 0.6)| 72.6 ($\\uparrow$ 0.2) |\n| TADA + DoG (EXPHORMER) | **86.5** ($\\uparrow$ 2.4)| **75.5** ($\\uparrow$ 2.6) | 85.1 ($\\uparrow$ 1.9)| 97.0 ($\\uparrow$ 1.3)| 73.5 ($\\uparrow$ 1.1) |\n| LGGD (EXPHORMER) | 85.5 ($\\uparrow$ 1.4)| 74.2 ($\\uparrow$ 1.3) | 83.9 ($\\uparrow$ 0.7)| 96.2 ($\\uparrow$ 0.5)| 72.7 ($\\uparrow$ 0.3) |\n| LGGD + DoG (EXPHORMER) | 86.2 ($\\uparrow$ 2.1)| 75.1 ($\\uparrow$ 2.2) | 84.8 ($\\uparrow$ 1.6)| 96.8 ($\\uparrow$ 1.1)| 73.7 ($\\uparrow$ 1.3) |\n| GeoMix (EXPHORMER) | 85.0 ($\\uparrow$ 0.9)| 74.2 ($\\uparrow$ 1.3) | 84.1 ($\\uparrow$ 0.9)| 96.4 ($\\uparrow$ 0.7)| 72.7 ($\\uparrow$ 0.3) |\n| GeoMix + DoG (EXPHORMER) | 86.3 ($\\uparrow$ 2.2)| 75.3 ($\\uparrow$ 2.4) | **85.2** ($\\uparrow$ 2.0)| **97.2** ($\\uparrow$ 1.5)| **73.8** ($\\uparrow$ 1.4) |\n[5] Lai, Yurui, et al. \"Efficient Topology-aware Data Augmentation for High-Degree Graph Neural Networks.\" KDD 2024.\n\n[6] Azad, Amitoz, and Yuan Fang. \"A Learned Generalized Geodesic Distance Function-Based Approach for Node Feature Augmentation on Graphs.\" KDD 2024.\n\n[7] Zhao, Wentao, et al. \"GeoMix: Towards Geometry-Aware Data Augmentation.\" KDD 2024.", "category": "Method_Comparison"}, {"question": "What is the motivation for using the latent diffusion model (LDM) to process latent representations instead of directly processing high-dimensional input data?", "answer": "The motivation for using the latent diffusion model (LDM) is to enhance the generative capabilities of the diffusion model while ensuring computational efficiency. The LDM trains the diffusion model in the compressed latent features generated by the graph autoencoder (GAE), reducing the dimensionality of the data processed. This decreases computational resources required for training and inference. For instance, the dimension of the input feature of Coauthor CS is reduced from 6805 to 512, significantly improving computational efficiency. Additionally, the LDM has shown remarkable performance in generating high-quality synthetic data across various domains[1, 2], allowing the diffusion model to capture high-level patterns more effectively and improve performance on generative tasks[1, 2].\n[1] Rombach, Robin, et al. \"High-resolution image synthesis with latent diffusion models.\" CVPR 2022.\n\n[2] Lovelace, Justin, et al. \"Latent diffusion for language generation.\" NeurIPS 2023.", "category": "Motivation_Analysis"}, {"question": "Why is it necessary to find a position within the entire graph for placing the generated synthetic graph structure, rather than placing it directly at the position of the given node?", "answer": "The DoG generates the synthetic graph structure from a Gaussian noise through a two-step process, rather than based on a given node. The training process of DoG comprises two main steps, as described in Algorithm 1 of Appendix D. In the first step, a Graph Autoencoder (GAE) is trained on the original graph. For each node, the GAE encoder takes as input the node’s attributes, its neighbors’ attributes, and the binary edge map of its local neighborhood, encoding these into a latent representation. To incorporate global positional information, positional embeddings of the node and its neighbors are added to their respective attributes. The GAE decoder then reconstructs both the node attributes and the binary edge map from the latent representation. The training is supervised using a mean squared error (MSE) loss. In the second step, a conditional diffusion model is trained on the latent representations obtained from the GAE, following the standard Latent Diffusion Model (LDM) training procedure [1]. The generation process of DoG also involves two steps, as outlined in Algorithm 2 of Appendix D. First, the conditional diffusion model generates a synthetic latent representation from Gaussian noise, conditioned on a given node label. Then, in the second step, the decoder of the GAE reconstructs the corresponding node attributes and the binary edge map that defines how the synthetic node connects to existing nodes in the original graph, based on the generated latent representation. The synthetic graph structure, including the synthetic node and its connecting edges, is reconstructed using the decoder of the GAE, which integrates the synthetic latent representation with the original graph's nodes, necessitating a search for an appropriate position within the entire graph.\n[1] Rombach, Robin, et al. \"High-resolution image synthesis with latent diffusion models.\" CVPR 2022.", "category": "Motivation_Analysis"}, {"question": "Which edges are included when feeding a given node with its edges into the model, and how does the model incorporate both local and global information?", "answer": "In the training process of the GAE in the DoG, only the edges directly connecting to a given node are included as the inputs to the encoder of the GAE. Additionally, the positional embeddings of both the given node and its neighboring nodes are added to their node attributes to incorporate global positional information. Existing works [3, 4] have shown that the positional embeddings of the nodes in the given graph can encode global positional information of nodes within the graph.This approach allows the model to utilize both local and global information of the nodes in the given graph instead of only processing the information in a very local manner.\n[3] Ma, Liheng, Reihaneh Rabbany, and Adriana Romero-Soriano. \"Graph attention networks with positional embeddings.\" PAKDD 2021.\n\n[4] You, Jiaxuan, Rex Ying, and Jure Leskovec. \"Position-aware graph neural networks.\" ICML 2019.", "category": "Experimental_Setup"}, {"question": "How does the bi-level neighborhood decoder connect the generated nodes and edges to the original graph?", "answer": "The bi-level neighborhood decoder first divides the nodes in the original graph into clusters using a balanced clustering method. It then generates a binary inter-cluster neighbor map for each synthetic node to identify clusters to connect to. For each identified cluster, it generates a binary intra-cluster neighbor map to identify individual nodes within the cluster to connect to the synthetic node.", "category": "Method_Mechanics"}, {"question": "Why did the authors choose a graph diffusion model approach instead of methods like GNNs or Graph Variational Autoencoders for synthetic graph generation?", "answer": "The authors chose a graph diffusion model approach because existing methods like GNNs and Graph Variational Autoencoders focus on graph-level synthetic graph generation, which generates entire synthetic graphs. In contrast, the authors' work [3, 4, 5, 6, 7]  focuses on node-level synthetic graph structure generation, aiming to create synthetic nodes and edges within an existing graph. This node-level approach is necessary for generating synthetic labeled nodes to enlarge the training set for node classification, which graph-level methods cannot achieve.\n[3] Mitton, Joshua, et al. \"A graph vae and graph transformer approach to generating molecular graphs.\" ICML 2020.\n\n[4] Bongini, P., M. Bianchini, and F. Scarselli. \"Molecular graph generation with graph neural networks.\" stat (2021).\n\n[5] Zhu, Yanqiao, et al. \"A survey on deep graph generation: Methods and applications.\" Learning on Graphs Conference. PMLR, 2022.\n\n[6] Vignac, Clement, et al. \"Digress: Discrete denoising diffusion for graph generation.\" ICLR 2023.\n\n[7] Cai, Zhou, Xiyuan Wang, and Muhan Zhang. \"Latent Graph Diffusion: A Unified Framework for Generation and Prediction on Graphs.\" arXiv preprint arXiv:2402.02518 (2024).", "category": "Motivation_Analysis"}, {"question": "Have the authors conducted experiments on large-scale graphs to evaluate the performance of DoG, and if so, what were the results?", "answer": "Yes, the authors conducted an experiment on AMiner-CS, a large-scale graph with 593,486 nodes and 6,217,004 edges. They used DoG to augment the original graph for training GRAND+ [9], resulting in a $1.4\\%$ improvement in accuracy, demonstrating DoG's scalability for large-scale graph structure generation.\n|   Methods    | Accuracy |\n| :----------: | :------: |\n|  GRAND+ [9]  | $54.2\\%$ |\n| DoG (GRAND+) | $55.6\\%$ |\n[9] Feng, Wenzheng, et al. \"Grand+: Scalable graph random neural networks.\" Proceedings of the ACM Web Conference 2022.", "category": "Experimental_Exposition"}, {"question": "What evidence supports that the improvement in the proposed method (DoG) is primarily due to the synthetic graph structures rather than the increase in the number of nodes?", "answer": "The ablation study compares DoG with two models: DoG-Duplicate, which adds synthetic nodes without connectivity, and DoG-KNN, which is referred to as DoG-KNN, connects the synthetic nodes generated by the DoG to the real nodes in the original graph using the K-Nearest Neighbor (KNN) algorithm on the node attributes. The value of $K$ in the KNN algorithm is set to {$1, \\lceil d_{avg}/4 \\rceil, \\lceil d_{avg}/2 \\rceil,  \\lceil d_{avg} \\rceil,  \\lceil 2\\times d_{avg} \\rceil,  \\lceil 4 \\times d_{avg} \\rceil$}, where $d_{avg}$ is the average node degree of the original real graph. DoG outperforms DoG-Duplicate by 1.4% and DoG-KNN by up to 1.0% in node classification accuracy on the Citeseer dataset. Additionally, DoG significantly outperforms NODEDUP, which duplicates real nodes, by 1.1%. These results demonstrate that the synthetic graph structures, rather than just the increase in the number of nodes, are the primary driver of the performance improvement.", "category": "Experimental_Analysis"}, {"question": "Was the performance of DoG on large-scale graphs demonstrated through an experiment on a graph with a million nodes and billion edges?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the use of DoG in the experiment on AMiner-CS improve the performance of GRAND+ by 1.4%?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "dSbbZwCTQI", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "You Only Debias Once: Towards Flexible Accuracy-Fairness Trade-offs at Inference Time", "authors": ["Hari Sowrirajan", "Jingbo Yang", "Andrew Y. Ng", "Pranav Rajpurkar"], "abstract": "Deep neural networks are prone to various bias issues, jeopardizing their applications for high-stake decision-making. Existing fairness methods typically offer a fixed accuracy-fairness trade-off at the inference time, since the weight of the well-trained model is a fixed point (fairness-optimum) in the weight space. Nevertheless, more flexible accuracy-fairness trade-offs at the inference time are practically desired since: 1) stakes of the same downstream task can vary for different individuals, and 2) different regions have diverse laws or regularization for fairness. If using the previous fairness methods, we have to train multiple models, each offering a specific level of accuracy-fairness trade-off. This is often computationally expensive, time-consuming, and difficult to deploy, making it less practical for real-world applications. To address this problem, we propose \\textit{You Only Debias Once} (YODO) to achieve in-situ flexible accuracy-fairness trade-offs at the inference time, using \\textit{a single model} that trained only once.  Instead of pursuing one individual fixed point (fairness-optimum) in the weight space, we aim to find a ``line'' in the weight space that connects the accuracy-optimum and fairness-optimum points using a single model. Points (models) on this line implement varying levels of accuracy-fairness trade-offs. At the inference time, by manually selecting the specific position of the learned ``line'', our proposed method can achieve arbitrary accuracy-fairness trade-offs for different end-users and scenarios. Experimental results on tabular and image datasets show that YODO achieves flexible trade-offs between model accuracy and fairness, at ultra-low overheads. Our codes are anonymously available at https://anonymous.4open.science/r/yodo-BB81 .", "keywords": "", "qa_pairs": [{"question": "How does the proposed method compare with other widely-adopted methods for ensuring demographic parity?", "answer": "To confirm the efficacy of the proposed method, additional baselines for demographic parity, including Prejudice Remover and Adversarial Debiasing, were incorporated and tested. The results, presented in Figure 3 on Page 9, such as: 1) Prejudice Remover [1], 2) Adversarial Debiasing [2][3], continue to demonstrate the effectiveness of the proposed method.\n\n[1] Fairness-aware classifier with prejudice remover regularizer, ECML-PKDD 2012. [2] Learning to pivot with adversarial networks, NeurIPS2017. [3] Learning adversarially fair and transferable representations, ICML2018.", "category": "Method_Comparison"}, {"question": "What is the rationale behind sampling alpha randomly instead of training two separate models with alpha = 0 and large alpha and then interpolating between them?", "answer": "Sampling alpha randomly is necessary because training two separate models and interpolating their weights at inference time does not achieve the goal of flexible accuracy-fairness trade-offs. Previous studies[1][2][3][4] have shown that interpolating weights of two separate well-trained models results in no better accuracy than an untrained model. The paper's experimental results in Appendix C.1 confirm that interpolation leads to worse accuracy, particularly when alpha = 0.5. To ensure that any point on the line between accuracy-optimum and fairness-optimum corresponds to a specific level of fairness, the paper sample alpha randomly during each epoch, In other words, each $\\alpha$ corresponds to one model, and with a wide range of different $\\alpha$ values, effectively training numerous models with a wide range of alpha values.\n [1] Linear mode connectivity and the lottery ticket hypothesis. ICML2020. [2] Learning neural network subspaces. ICML2021. [3] Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the neural tangent kernel. NeurIPS2020. [4] Loss surface simplexes for mode connecting volumes and fast ensembling. ICML2021.", "category": "Method_Disambiguation"}, {"question": "Why was alpha = 1 chosen as the 'fair classifier,' and did the authors explore values of alpha greater than 1 to evaluate the accuracy-fairness tradeoff?", "answer": "The choice of alpha = 1 was based on addressing the trade-off between model accuracy and fairness. The authors conducted experiments using YODO with alpha values ranging from 1 to 5. The results showed that while larger values of alpha improve fairness, they also significantly decrease the accuracy of the downstream task. Alpha = 1 was found to achieve an optimal balance between accuracy and fairness.", "category": "Motivation_Analysis"}, {"question": "What does the claim 'Our model can be regarded as training infinite models with different fairness constraints' mean in the context of the method described in the paper?", "answer": "The claim means that the method aims to find a 'line' between accuracy-optimum and fairness-optimum points in the weight space. By randomly sampling an $\\alpha$ during each epoch, each $\\alpha$ corresponds to one model. With a wide range of different $\\alpha$ values, numerous models (infinite) are trained throughout the process, each representing a specific level of fairness.", "category": "Method_Disambiguation"}, {"question": "How is the prediction procedure carried out for a given value of $\\alpha$ using the learned weights?", "answer": "The prediction procedure involves three steps: 1. Choose the desired trade-off parameter $\\alpha$ to balance accuracy and fairness. 2. Compute the weighted combination of the two sets of trained weights, $(1-\\alpha)\\omega_1 + \\alpha\\omega_2$, to obtain the model parameters for the desired trade-off. 3. Apply the prediction function to the test sample $\\mathbf{x}$ as $f(\\mathbf{x}; (1-\\alpha)\\omega_1 + \\alpha\\omega_2)$ to obtain the predicted output. This allows users to select the desired level of accuracy and fairness by choosing the appropriate $\\alpha$.", "category": "Experimental_Setup"}, {"question": "What additional fairness models were incorporated as baselines to confirm the efficacy of the proposed method, and where are the results presented?", "answer": "The authors incorporated additional widely-adopted baselines for demographic parity, including 1) Prejudice Remover [1] and 2) Adversarial Debiasing [2][3]. The results are presented in Figure 3 on Page 9, demonstrating the effectiveness of the proposed method.\n[1] Fairness-aware classifier with prejudice remover regularizer, ECML-PKDD 2012. [2] Learning to pivot with adversarial networks, NeurIPS 2017. [3] Learning adversarially fair and transferable representations, ICML 2018.", "category": "Method_Mechanics"}, {"question": "How does the proposed method's performance compare to previous work, and what specific improvements are attributed to the double trainable weights compared to fixed training?", "answer": "The proposed method's better performance is attributed to the double trainable weights, which improve the expressive power of the neural network, leading to a better accuracy-fairness trade-off than fixed training in most cases. For example, in the CelebA-Smiling-Gender dataset (shown in the 3rd subfigure of Figure 4), the method demonstrates its advantage. Additionally, the proposed method requires only one model, whereas fixed training needs 21 models trained from scratch, as shown in Figure 3.", "category": "Method_Comparison"}, {"question": "What is the utility of the regularization loss $\\mathcal{L}_{reg}$ in ensuring diversity between the endpoints $\\omega_1$ and $\\omega_2$, and why is it important despite the absence of ensembling networks with different values of $\\alpha$?", "answer": "The regularization loss $\\mathcal{L}_{reg}$ minimizes the cosine similarity between $\\omega_1$ and $\\omega_2$, promoting diversity between the two endpoints. Since the cosine similarity of two random high-dimensional vectors is typically close to zero, and the accuracy-optimum and fairness-optimum points in the weight space are generally distinct, $\\mathcal{L}_{reg} = \\frac{ \\langle \\omega_1, \\omega_2 \\rangle^2 }{ | \\omega_1 |_{2}^2 | \\omega_2 |_{2}^2 } = 0$ ensures that the two points do not become identical. While it may not happen in experiments, $\\mathcal{L}_{reg}$ is used as a safeguard to prevent this outcome. A detailed discussion on its functionality has been added to Section 4.2 on Page 7.", "category": "Experimental_Analysis"}, {"question": "Did the experiments include larger models beyond ResNet-18 on the CelebA dataset, and what were the observed outcomes regarding accuracy-fairness trade-off and stability?", "answer": "Yes, the experiments included larger models such as ResNet18, ResNet34, ResNet50, ResNet101, WideResNet50, and WideResNet101 on the CelebA dataset. The results indicated that all models achieved a flexible accuracy-fairness trade-off during inference and that larger models demonstrated greater stability.", "category": "Experimental_Exposition"}, {"question": "What is the relationship between fixed training with $A=0$ and ERM, and are they the same?", "answer": "Fixed training with $A=0$ and ERM are the same. The reported result of ERM was conducted using fixed training with $A=0$. Different values of $A$, which is set to $[0,1]$ with an interval of $0.05$* to *with different values of $A$, which is set to $(0,1]$ with an interval of $0.05$*.", "category": "Experimental_Exposition"}, {"question": "Why is there a need for model weights to be orthogonal in the proposed fairness-accuracy trade-off framework, and what ensures that the model cannot simultaneously achieve the best accuracy and fairness?", "answer": "Minimizing $\\mathcal{L}_{reg}$ (cosine similarity between $\\omega_1$ and $\\omega_2$) ensures orthogonality between the accuracy-optimum and fairness-optimum points in the weight space. Since the cosine similarity of two random high-dimensional vectors is typically close to zero, and the accuracy-optimum and fairness-optimum points are generally distinct, this regularization term $\\mathcal{L}_{reg} = \\frac{ \\langle \\omega_1, \\omega_2 \\rangle^2 }{ | \\omega_1 |_{2}^2 | \\omega_2 |_{2}^2 } = 0$ prevents the two points from becoming identical. While it is theoretically possible for these points to coincide, the regularization term is used to prevent this scenario.", "category": "Motivation_Analysis"}, {"question": "What comparisons with existing methods and enhancements to the literature review have been included to address the lack of informative content and method comparisons?", "answer": "The literature review now includes a paper on hypernetworks for loss conditional training and a discussion on multi-objective optimization in the context of Pareto frontiers, focusing on accuracy and fairness. For model comparisons, widely-adopted baselines for demographic parity, such as Prejudice Remover and Adversarial Debiasing, have been integrated, with updated results shown in Figure 3. The primary contribution of the method is its ability to achieve flexible accuracy-fairness trade-offs at inference time, complementing rather than competing with existing fairness methods, and it can be seamlessly integrated into existing techniques.", "category": "Method_Disambiguation"}, {"question": "How does the proposed method substantiate its effectiveness in terms of quantitative performance and how does it relate to existing fairness techniques?", "answer": "The effectiveness of the proposed method is substantiated by integrating widely-adopted baselines for demographic parity, such as Prejudice Remover and Adversarial Debiasing, and showcasing the updated results in Figure 3, which demonstrate the efficacy of the approach. The primary contribution of the method is its ability to achieve flexible accuracy-fairness trade-offs at inference time, complementing existing fairness methods rather than competing with them. Additionally, the method can be seamlessly integrated into existing fairness techniques.", "category": "Method_Mechanics"}, {"question": "What hidden representation is measured in Figure 6, and how does it relate to demographic parity and sensitive attributes?", "answer": "The hidden representation measured in Figure 6 is from the penultimate layer of the model. Demographic parity requires that model predictions be independent of sensitive attributes (e.g., demographic groups). The representation is examined to determine the independence between the hidden representation and the sensitive attribute. When $\\alpha=0$ (unfair model), the representations of the male and female groups contain more sensitive information, as shown in the first figure on the bottom row of Figure 6. Conversely, when $\\alpha=1$ (fair model), the representations display minimal sensitive information, as shown in the last figure on the bottom row of Figure 6. Probing representations is a common approach to examining fairness[1][2], as supported by prior research. [1] The variational fair autoencoder, ICLR2016 [2] Fairness via Representation Neutralization, NeurIPS2021. ", "category": "Experimental_Analysis"}, {"question": "Why was the value of $A$ set to 1 in the regularized objective for training the 'fair weights' instead of using a larger value to achieve greater fairness?", "answer": "While larger values of $A$ are expected to increase fairness, they also significantly reduce the accuracy of the downstream task. Experiments with varying values of $A$ (ranging from 1 to 5) showed that increasing $A$ improves fairness but deteriorates task accuracy. Setting $A=1$ achieves an optimal balance between fairness and accuracy.", "category": "Motivation_Analysis"}, {"question": "How is the accuracy-fairness trade-off depicted in Figures 2 and 3, and why does the Pareto front in these figures differ from the concave curve typically seen in multi-objective optimization (MOO) literature?", "answer": "Figures 2 and 3 correctly depict the accuracy-fairness trade-off, with the optimal trade-off in the upper-right corner, unlike the MOO paper where it is in the lower-right. The y-axis represents the error rate (lower is better), and the x-axis represents $\\Delta DP$ (lower is better), making the upper-right corner the best accuracy-fairness trade-off. In the case, the Pareto front would appear as a concave curve if the optimal trade-off were in the upper-right corner. Consequently, the upper-right corner signifies the best accuracy-fairness trade-off (while in Figure 1 of the MOO paper[1], the best trade-off is in the lower-right corner). [1] Scalable pareto front approximation for deep multi-objective learning. ICDM2021", "category": "Experimental_Analysis"}, {"question": "Did the paper originally claim that the examined fairness notion was general without loss of generality?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the undefined citation in the paragraph after Definition 2.1 in Section 2.1 a result of mistakenly citing the same paper twice?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were experiments conducted on larger models beyond ResNet-18 on the CelebA dataset?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "kLZsLlIpDU", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Escaping the Big Data Paradigm with Compact Transformers", "authors": ["Zeyu Han", "Chao Gao", "Jinyang Liu", "Jeff Zhang", "Sai Qian Zhang"], "abstract": "With the rise of Transformers as the standard for language processing, and their advancements in computer vision, there has been a corresponding growth in parameter size and amounts of training data. Many have come to believe that because of this, transformers are not suitable for small sets of data.\nThis trend leads to concerns such as: limited availability of data in certain scientific domains and the exclusion of those with limited resource from research in the field.\nIn this paper, we aim to present an approach for small-scale learning by introducing Compact Transformers. \nWe show for the first time that with the right size, convolutional tokenization, transformers can avoid overfitting and outperform state-of-the-art CNNs on small datasets. \nOur models are flexible in terms of model size, and can have as little as 0.28M parameters while achieving competitive results.\nOur best model can reach 98% accuracy when training from scratch on CIFAR-10 with only 3.7M parameters, which is a significant improvement in data-efficiency over previous Transformer based models being over 10x smaller than other transformers and is 15% the size of ResNet50 while achieving similar performance. \nCCT also outperforms many modern CNN based approaches, and even some recent NAS-based approaches.\nAdditionally, we obtain a new SOTA result on Flowers-102 with 99.76% top-1 accuracy, and improve upon the existing baseline on ImageNet (82.71% accuracy with 29% as many parameters as ViT), as well as NLP tasks.\nOur simple and compact design for transformers makes them more feasible to study for those with limited computing resources and/or dealing with small datasets, while extending existing research efforts in data efficient transformers.", "keywords": "", "qa_pairs": [{"question": "What are the theoretical insights behind the three changes introduced in the Vision Transformer (Change A: reducing sizes, Change B: adding a softmax function to sequence outputs, and Change C: applying max pooling + conv2d as a tokenizer), and are all these changes necessary?", "answer": "The theoretical insights behind the changes are as follows: (1) Reducing sizes inside Vision Transformer (Change A) addresses the lack of locality and inductive biases in the Transformer Encoder, which limits generalizability in small-scale datasets. Smaller models converge more easily without additional augmentations. (2) Adding a softmax function to sequence outputs as weights (Change B) improves generalization in depth-limited Transformer Encoders by pooling over output tokens and aggregating context immediately before classification. (3) Applying max pooling + conv2d as a tokenizer (Change C) eases convergence for limited data by using a more relaxed tokenizer with overlapping convolutions. The necessity of these changes is supported by the ablation study, which includes experiments testing various combinations of these changes: None: ViT-12/16, A: ViT-Lite, A+B: CVT, A+B+C: CCT, B: ViT-12/16 + SP, C: ViT-12/7x1 and ViT-12/3x2 + CT, B+C: ViT-12/7x1 and ViT-12/3x2 + SP, and A+C: CCT-7/7x1 and CCT-7/3x2 + CT.", "category": "Experimental_Exposition"}, {"question": "How are the MACs (Multiply-Accumulate Operations) computed in the tables, and what methodology was used to calculate them?", "answer": "MACs, which are used interchangeably with FLOPs in many papers, represent the total number of floating point multiplies and additions. They were computed for all experiments using a standard PyTorch extension, and these details will be open-sourced with the code.", "category": "Method_Mechanics"}, {"question": "How do the chosen datasets (CIFAR, MNIST, Flowers) support the motivation for developing pre-train-free and data-efficient training for medical and scientific domains, given their differences from typical medical and scientific datasets?", "answer": "The chosen datasets (CIFAR, MNIST, Flowers) are used to simulate data scarcity and domain-specific challenges common in medical and scientific studies. CIFAR and MNIST represent small-resolution datasets, while Flowers represents a medium-resolution dataset with limited and highly diverse samples, similar to medical data. Training from scratch on these datasets demonstrates the feasibility of pre-train-free and data-efficient training, which is less reliant on large-scale pretraining and more applicable to domains like medical studies where pretraining is less useful. Additionally, the work has been cited over 100 times in scientific and medical studies, indicating its practical relevance.", "category": "Experimental_Setup"}, {"question": "Why does the paper not include comparisons with other Vision Transformer (ViT) architectures that introduce convolutional inductive biases, such as Swin Transformer and PiT?", "answer": "The paper does not include comparisons with other ViT architectures like Swin Transformer and PiT because the focus of this study is on making minimal changes to the original ViT model to improve its applicability to low-data/low-dimensionality tasks. Unlike these referenced works, which introduce convolutional inductive biases in every layer and aim to create new hierarchical vision transformers, this paper investigates the claim from the original ViT paper about the lack of inductive biases in Transformers. It does so by maintaining the same Transformer architecture and improving classification, tokenization, and training techniques, demonstrating that these improvements enable the model to compete with CNNs in data-limited settings.", "category": "Experimental_Analysis"}, {"question": "What is the rationale for not comparing the proposed approach to hierarchical vision transformers and hybrid models in the study?", "answer": "The rationale for not comparing to hierarchical vision transformers and hybrid models is twofold: 1) The study focuses on investigating the limitations of the original Vision Transformer (ViT) by maintaining its architecture and making minimal changes to improve its applicability to low-data tasks. Hybrid and hierarchical models introduce convolutions or inductive biases at the layer level, which diverges from the study's goal of evaluating the original ViT architecture. 2) The paper is not aimed at proposing a new architecture but rather at testing the claim from the original ViT paper about the lack of inductive biases in Transformers. By improving pre- and post-processing modules while keeping the Transformer encoder unchanged, the study demonstrates that these enhancements enable the model to compete with CNNs in data-limited settings.", "category": "Motivation_Analysis"}, {"question": "Do the other methods mentioned in the paper use the same augmentation strategy as the one adopted by the authors?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "v8i3Meu2Zu", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Dreamix: Video Diffusion Models are General Video Editors", "authors": ["Zeyu Han", "Chao Gao", "Jinyang Liu", "Jeff Zhang", "Sai Qian Zhang"], "abstract": "Text-driven image and video diffusion models have recently achieved unprecedented generation realism. While diffusion models have been successfully applied for image editing, few can edit motion in video. We present a diffusion-based method that is able to perform text-based motion and appearance editing of general, real-world videos. Our approach uses a video diffusion model to combine, at inference time, the low-resolution spatio-temporal information from the original video with new, high resolution information that it synthesized to align with the guiding text prompt. As maintaining high-fidelity to the original video requires retaining some of its high-resolution information, we add a preliminary stage of finetuning the model on the original video, significantly boosting fidelity. We propose to improve motion editability by using a mixed objective that jointly finetunes with full temporal attention and with temporal attention masking. We extend our method for animating images, bringing them to life by adding motion to existing or new objects, and camera movements. Extensive experiments showcase our method's remarkable ability to edit motion in videos.", "keywords": "", "qa_pairs": [{"question": "How does the proposed approach in Dreamix differ from Dreambooth in terms of technical novelty and application to video editing?", "answer": "The proposed approach in Dreamix transfers image-based techniques to video editing and demonstrates strong performance compared to more complex methods, showing that a VDM backbone with a simple method can achieve good video editing results.", "category": "Method_Comparison"}, {"question": "How does the proposed method with the joint objective address the issue of preserving high-resolution details such as fine textures or object identity in the input video?", "answer": "Fine-tuning enhances the fidelity to the original details, and the mixed image-video component enables better separation of motion and appearance information.", "category": "Method_Mechanics"}, {"question": "What mechanisms are in place to ensure spatial-temporal consistency in the edited video sequences using the proposed approach?", "answer": "For video-to-video editing, temporal consistency is ensured by the input video and the video prior of the VDM backbone. For other tasks, the VDM prior alone enforces temporal consistency. Mixed image-video fine-tuning is used to leverage this prior without overfitting.", "category": "Method_Disambiguation"}, {"question": "How does the choice of VDM backbone affect the generation of fine details in the proposed framework?", "answer": "The choice of VDM backbone significantly impacts the amount of details and the visual appearance of the video. Using a ModelScope backbone improves the visual appearance, and with better VDM backbones, the method should generate better fine details. These results are discussed in the 'Model Availability' paragraph of the limitations section, appendix C.5.", "category": "Experimental_Analysis"}, {"question": "Why were the comparisons with existing methods like Tune-A-Video and other advanced video editing tools considered not comprehensive or fair, and how does Dreamix address this limitation?", "answer": "The comparisons were considered lacking because methods like Tune-A-Video do not edit the motion in video. Dreamix addresses this limitation by proposing multiple new baselines that demonstrate its superiority and by comparing against multiple strong baselines and commercial products for other use cases, such as image-to-video comparisons with Gen2 and PikaLabs.", "category": "Method_Comparison"}, {"question": "How does Dreamix address concerns about maintaining high fidelity and detail, particularly in transitions and fine features during video editing?", "answer": "The choice of VDM backbone significantly impacts the amount of detail and visual appearance of the video. Using a ModelScope backbone improves visual appearance, and with better VDM backbones, Dreamix can generate finer details. These results are discussed in the 'Model Availability' paragraph of the limitations section, appendix C.5.", "category": "Method_Mechanics"}, {"question": "How does the computational cost of the method compare to other current video editing methods, particularly with newer VDMs like ModelScope?", "answer": "With newer VDMs like ModelScope, the computational cost of the method is reduced to a standard 24GB GPU and a few minutes, which is on par with other current video editing methods.", "category": "Method_Comparison"}, {"question": "Is the rare string  t^{*} randomly selected or not?", "answer": "True", "category": "Claim_Verification"}, {"question": "The experiments show whether the choice of specific strings significantly affects the results？", "answer": "False", "category": "Claim_Verification"}]}
{"id": "dtvJF1Vy2i", "venue": "NEURIPS 2024", "paper_decision": "Accept (poster)", "title": "What matters when building vision-language models?", "authors": ["Hugo Laurençon", "Leo Tronchon", "Matthieu Cord", "Victor Sanh"], "abstract": "The growing interest in vision-language models (VLMs) has been driven by improvements in large language models and vision transformers. Despite the abundance of literature on this subject, we observe that critical decisions regarding the design of VLMs are often not justified. We argue that these unsupported decisions impede progress in the field by making it difficult to identify which choices improve model performance. To address this issue, we conduct extensive experiments around pre-trained models, architecture choice, data, and training methods. Our consolidation of findings includes the development of Idefics2, an efficient foundational VLM of 8 billion parameters. Idefics2 achieves state-of-the-art performance within its size category across various multimodal benchmarks, and is often on par with models four times its size. We release the model (base, instructed, and chat) along with the datasets created for its training.", "keywords": "vision-language model;multimodal", "qa_pairs": [{"question": "What are the specific ways in which the paper' work expands upon or differs from previous studies on the impact of vision encoders, language models, and input resolution, such as those in papers [a, b, c, d]?", "answer": "The paper' work expands upon previous studies in several ways: For [a] Prismatic VLMs, the paper demonstrated that a LoRA strategy can achieve stable training and improved performance when adding trainable weights to the vision encoder, unlike their findings of significant performance degradation. For [d] VILA, the paper extended the scope of image resolution studies to 980 pixels and focused more on image splitting strategies rather than fixed input resolutions. For [b] 'Eyes Wide Shut?', the paper' study is broader, addressing more aspects beyond the impact of pre-trained vision encoders. Additionally, [c] BRAVE is considered concurrent work due to its recent publication date. The paper' unique contributions are detailed in sections 3.2, 3.3, 3.4, and 4, providing new insights and comprehensive comparisons in the field.", "category": "Experimental_Exposition"}, {"question": "How does the proposed method for improving VLMs compare to other existing methods such as [b,c,d], and what unique contributions does it make?", "answer": "The proposed method complements existing methods [b,c,d] by introducing novel approaches: 1) An efficient pooling strategy using a small number of visual tokens, compared favorably against MM1, SPHINX-2K, DeepSeek-VL, and DePALM. 2) An image aspect ratio preserving strategy, unique among VLMs. 3) Fine-tuning with variable visual token counts, offering flexibility not present in models like Llava, MM1, and Monkey, which only fine-tune with image splitting.", "category": "Method_Comparison"}, {"question": "How do the observations and findings of this work align with or differ from recent advancements in the field, including the dataset Cambrian-10M, which was released after NeurIPS?", "answer": "The findings of this work remain relevant in the context of recent advancements. Recent works have shifted towards unfreezing language models when feasible or using LoRA for improved stability, which aligns with the proposed methods in this work. Additionally, recent works use image splitting with variable sub-image counts during training to enable compute-performance trade-offs at inference. The dataset YYYYYY demonstrates the potential of leveraging diverse academic datasets for instruction fine-tuning, previously implemented at a smaller scale. The dataset Cambrian-10M is inspired by this work, containing most of the selected public academic datasets and the rendered table images created by the authors. Furthermore, the XXXXXX-8B model outperforms Cambrian-1-8B on challenging benchmarks like MMMU and MathVista while being significantly faster at inference due to fewer visual tokens.", "category": "Experimental_Exposition"}, {"question": "How do the ablation results in Tables 9 and 15 support the claim that reducing the number of visual tokens is beneficial for non-OCR tasks, and why does Finding 4 need to be stated more cautiously?", "answer": "The ablation results indicate that for most non-OCR tasks, the number of tokens to encode the image is not critically important, which is significant for applications like robotics that need efficient inference. Tables 9 and 15 show that the image splitting strategy's utility is primarily for OCR tasks. This aligns with subsequent works on Arxiv. Finding 4 has been rephrased to emphasize that reducing visual tokens with learned pooling improves compute efficiency and performance on non-OCR tasks.", "category": "Experimental_Exposition"}, {"question": "How does the dataset YYYYYY differ from existing VLM instruction tuning datasets, and what measures were taken to prevent test dataset leaks?", "answer": "VLM instruction tuning datasets were primarily of two types: synthetically generated Q/A pairs, mostly using ChatGPT (like Llava), and compilations of existing training sets from different benchmarks (like InstructBLIP). YYYYYY differs from existing VLM instruction tuning datasets by compiling a diverse mixture of 50 high-quality datasets, transforming them into a consistent format, creating images for benchmarks without them, merging Q/A pairs for multi-turn dialogues, and preparing non-prompted datasets with prompts. To prevent test dataset leaks, authors removed images present in test sets of evaluated benchmarks and conducted a contamination analysis to ensure no overlap.", "category": "Experimental_Setup"}, {"question": "How do the authors demonstrate that the superior performance of YYYYYY is not solely due to the dataset but also influenced by their findings?", "answer": "The dataset YYYYYY is introduced during the fine-tuning stage and plays a crucial role in the final performance. However, the findings, particularly those derived before the fine-tuning stage, informed better model and training method choices. Finding 6, which is independent of the dataset, further supports that the superior performance is not solely driven by the dataset.", "category": "Experimental_Setup"}, {"question": "Why were state-of-the-art commercial models not extensively included in the baseline comparisons, and were any non-open-source models considered?", "answer": "State-of-the-art commercial models were not extensively included because they tend to be significantly larger, making direct comparisons challenging. However, MM1, a non-open-source model, was included. Additionally, comparisons with Gemini and Claude 3 are provided in Table 15 of the Appendix, showing that the model approaches their performance on benchmarks like MathVista and TextVQA, though larger scales are needed for benchmarks like MMMU.", "category": "Method_Disambiguation"}, {"question": "How does the proposed model compare to open-source models like InternLM-XComposer2-VL in terms of performance and design objectives?", "answer": "The proposed model differs in scope and objectives from InternLM-XComposer2-VL. The primary focus of the model is efficiency at inference time, achieved through the use of a perceiver resampler to reduce visual tokens, which improves performance on most tasks except OCR-heavy ones. Additionally, the model does not score lower on the MMMU benchmark, uses only open datasets (as opposed to proprietary data), and emphasizes learning from ablations for the research community rather than solely delivering a model.", "category": "Method_Comparison"}, {"question": "How does the model's performance change with an increased number of in-context demonstrations, and why were further evaluations of in-context learning not pursued?", "answer": "The model showed an improvement of 1-2 points when increasing from 4-shot to 8-shot in-context examples. However, adding more examples primarily boosts scores by helping the model follow the benchmark's template rather than improving its task-solving ability. Further in-context learning evaluations were not pursued because they would involve complex methods like using a LLM as a judge, which is out of the current study's scope and not comparable with other works.", "category": "Experimental_Analysis"}, {"question": "What effect does unfreezing all parameters during pretraining have on the performance of cross-attention architectures?", "answer": "Unfreezing all parameters during pretraining for cross-attention architectures led to unstable training and eventual loss divergence, as stated in Finding 3 of the paper.", "category": "Method_Mechanics"}, {"question": "Does the final model incorporate both LoRA and DoRA during training, and if so, for what specific purposes are each used?", "answer": "The final model primarily uses DoRA for fine-tuning. Initially, pretraining and ablations were conducted with LoRA, but after DoRA's publication, it was tested as a replacement for LoRA and showed similar or slightly better performance. DoRA was chosen for fine-tuning due to its efficiency and lack of additional parameters, without compromising performance.", "category": "Experimental_Setup"}, {"question": "What is the maximum number of images the model can support during inference, and how can this limit be extended?", "answer": "The model can support an arbitrary number of images, limited by the sequence length it was trained with. For a sequence length of 2048, it can handle just over 30 images. Fine-tuning with a longer sequence length can extend this limit to support more images, such as for video processing.", "category": "Experimental_Setup"}, {"question": "Why is the evaluation of the model limited to a small number of benchmarks, and how does the selection of benchmarks like MMMU, MathVista, TextVQA, DocVQA, and MMBench ensure comprehensive assessment of the model's capabilities?", "answer": "The evaluation is limited to a small number of benchmarks due to the challenge of finding benchmarks that effectively distinguish VLM performance. Some open-ended benchmarks are unreliable as they often reward adherence to expected answer templates rather than true capability. For example, the paper obtained 81.2 on VQAv2 while Gemini 1.5 Pro scored 73.2. It doesn’t mean that the model is better than Gemini 1.5 Pro. Benchmarks like MMMU were selected for their comprehensive categories, MathVista for reasoning and math abilities, TextVQA and DocVQA for OCR and document understanding, and MMBench for general VQA evaluation to ensure a thorough assessment of the model's capabilities.", "category": "Experimental_Analysis"}, {"question": "Why did the authors use cross-attention training for the ablation study in Section 3.1 instead of the fully autoregressive training approach, and do the findings from the cross-attention setup hold under the autoregressive approach?", "answer": "The authors used cross-attention training for the ablation study in Section 3.1 to initially show results with the reference approach at that time. They verified that the two most significant findings (replacing Llama1 with Mistral and CLIP with SigLIP) also hold under the fully autoregressive approach. Due to consistency in early training results and to save computational resources, they did not extend these additional training to full completion. These ablations were not included in the final paper, but the authors are willing to add these results for clarity.", "category": "Motivation_Analysis"}, {"question": "Did prior works primarily use LLMs to create synthetic Q/A pairs, which limited diversity and encouraged hallucinations?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are all the findings in the study validated for the model XXXXXX, including those related to the cross-attention architecture?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "tz83Nyb71l", "venue": "NEURIPS 2024", "paper_decision": "Accept (poster)", "title": "YOLOv10: Real-Time End-to-End Object Detection", "authors": ["Ao Wang", "Hui Chen", "Lihao Liu", "Kai CHEN", "Zijia Lin", "Jungong Han", "Guiguang Ding"], "abstract": "Over the past years, YOLOs have emerged as the predominant paradigm in the field of real-time object detection owing to their effective balance between computational cost and detection performance. Researchers have explored the architectural designs, optimization objectives, data augmentation strategies, and others for YOLOs, achieving notable progress. However, the reliance on the non-maximum suppression (NMS) for post-processing hampers the end-to-end deployment of YOLOs and adversely impacts the inference latency. Besides, the design of various components in YOLOs lacks the comprehensive and thorough inspection, resulting in noticeable computational redundancy and limiting the model's capability. It renders the suboptimal efficiency, along with considerable potential for performance improvements. In this work, we aim to further advance the performance-efficiency boundary of YOLOs from both the post-processing and the model architecture. To this end, we first present the consistent dual assignments for NMS-free training of YOLOs, which brings the competitive performance and low inference latency simultaneously. Moreover, we introduce the holistic efficiency-accuracy driven model design strategy for YOLOs. We comprehensively optimize various components of YOLOs from both the efficiency and accuracy perspectives, which greatly reduces the computational overhead and enhances the capability. The outcome of our effort is a new generation of YOLO series for real-time end-to-end object detection, dubbed YOLOv10. Extensive experiments show that YOLOv10 achieves the state-of-the-art performance and efficiency across various model scales. For example, our YOLOv10-S is 1.8$\\times$ faster than RT-DETR-R18 under the similar AP on COCO, meanwhile enjoying 2.8$\\times$ smaller number of parameters and FLOPs. Compared with YOLOv9-C, YOLOv10-B has 46\\% less latency and 25\\% fewer parameters for the same performance. Code and models are available at https://github.com/THU-MIG/yolov10.", "keywords": "YOLO;object detection;computer vision", "qa_pairs": [{"question": "Do the intrinsic ranks in high redundancy stages decrease after replacing certain blocks with CIB, and how does this affect the model's performance and parameters?", "answer": "A lower intrinsic rank indicates higher redundancy, and a higher rank implies more condensed parameters. For example, given a parameter matrix $W$, a low intrinsic rank suggests that it can be replaced through low-rank approximation with fewer parameters. Such an approximation is not to improve the rank of the original $W$, but to reduce its redundancy of parameters. Approach uses this to replace blocks with CIB in high redundancy stages to reduce computational costs while maintaining model capacity. As shown in Tab. 5 and Fig. 2, the resulting models with fewer parameters exhibit similar intrinsic ranks to before, maintaining competitive performance and faster inference speed.", "category": "Experimental_Exposition"}, {"question": "What is the motivation and academic contribution of this paper in improving YOLOs, particularly in post-processing and architectural design?", "answer": "The motivation of this paper is to improve YOLOs by addressing both post-processing and architectural design. (1) For post-processing, the authors identify that one-to-one matching, while eliminating the need for NMS, suffers from limited supervision and leads to inferior performance for YOLOs, as shown in Table 3. To address this, they introduce dual label assignments combining one-to-one and one-to-many matching to enrich supervision during training and simplify post-processing during inference. Additionally, they propose a consistent matching metric to minimize the theoretical supervision gap, improving performance without hyper-parameter tuning. (2) For architectural design, the authors conduct a systematic and comprehensive inspection of various components, similar to previous works like ConvNeXt and MobileLLM, to improve YOLO efficiency and accuracy. They also introduce new designs, such as the rank-guided block design and partial self-attention module, leading to favorable performance and efficiency improvements. These contributions aim to inspire further research in the field. For example, authors think that by simplifying the post-processing, the NMS-free training strategy leaves more room for the model itself, which may facilitate more advanced architectural optimizations and designs.", "category": "Motivation_Analysis"}, {"question": "What is the rationale behind the architectural choices made in the model, and how do these choices contribute to the model's efficiency and accuracy?", "answer": "Like ConvNeXt [1] and MobileLLM [2], the architectural choices are motivated by a comprehensive inspection of various components, focusing on both efficiency and accuracy. The lightweight classification head design aims to reduce computational redundancy in the classification task compared to the regression task. The rank-guided block design adopts a compact block structure based on redundancy indicated by intrinsic ranks. These designs, while less pronounced than the NMS-free training strategy, provide flexibility and inspire further research on the YOLO architecture, serving as an initial exploration for advanced designs.\n[1] A ConvNet for the 2020s, CVPR 2022.\n\n[2] MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases, ICML 2024.", "category": "Motivation_Analysis"}, {"question": "How does the performance of NMS-free training compare to the original one-to-many training as the model size increases, and what factors contribute to the observed performance gap?", "answer": "The performance gap between NMS-free training and the original one-to-many training diminishes as the model size increases. This gap is attributed to the limitations in model capability. NMS-free training requires more discriminative features for one-to-one matching. The YOLOv10-N model, with limited capacity, exhibits a notable performance gap due to insufficiently discriminative features. In Fig.1 in the global response pdf, authors visualize the average cosine similarity of each anchor point's extracted features with those of all other anchor points on the COCO validation set. Conversely, the YOLOv10-X model, with stronger capability, shows no performance gap. Increasing model size reduces feature similarity between anchor points, benefiting one-to-one matching.", "category": "Experimental_Analysis"}, {"question": "What is the latency of YOLOv10 with the original one-to-many training using NMS?", "answer": "The latency of YOLOv10 with the original one-to-many training using NMS is as follows: YOLOv10-N: 6.19ms, YOLOv10-S: 7.15ms, YOLOv10-M: 9.03ms, YOLOv10-B: 10.04ms, YOLOv10-L: 11.52ms, YOLOv10-X: 14.67ms.\n|Model|Latency|\n|---------|-------|\n|YOLOv10-N|6.19ms|\n|YOLOv10-S|7.15ms|\n|YOLOv10-M|9.03ms|\n|YOLOv10-B|10.04ms|\n|YOLOv10-L|11.52ms|\n|YOLOv10-X|14.67ms|", "category": "Experimental_Exposition"}, {"question": "What are the detailed training costs, including total training epochs and throughput, for YOLOv10 compared to other YOLO models when trained with dual label assignments?", "answer": "Authors measured the training throughput on 8 NVIDIA 3090 GPUs using official codebases. YOLOv10, despite having 500 training epochs, achieves a high training throughput of 17.2 epochs/hour, resulting in a total training time of 29.1 hours, making its training cost affordable. Here are the comparison results for medium variants: YOLOv6-3.0-M (300 epochs, 7.2 epochs/hour, 41.7 hours), YOLOv8-M (500 epochs, 18.3 epochs/hour, 27.3 hours), YOLOv9-M (500 epochs, 12.3 epochs/hour, 40.7 hours), Gold-YOLO-M (300 epochs, 4.7 epochs/hour, 63.8 hours), YOLO-MS (300 epochs, 7.1 epochs/hour, 42.3 hours).\n|Model|Epochs|Throughput (epochs/hour)|Total Time (hour)|\n|-------------|------|------------------------|-----------------|\n|YOLOv6-3.0-M|300|7.2|41.7|\n|YOLOv8-M|500|18.3|27.3|\n|YOLOv9-M|500|12.3|40.7|\n|Gold-YOLO-M|300|4.7|63.8|\n|YOLO-MS|300|7.1|42.3|\n|**YOLOv10-M**|500|17.2|29.1|", "category": "Experimental_Exposition"}, {"question": "What are the training outcomes for YOLOv10 when trained for 300 epochs, and how do these compare to other models trained for the same epoch count?", "answer": "In experiments, authors follow previous works [3,4].When trained for 300 epochs, YOLOv10 achieves superior performance and efficiency compared to YOLOv6[5], Gold-YOLO[6], and YOLO-MS[7], as evidenced by higher AP$^{val}$ percentages and lower latency values. Despite being trained for 500 epochs, YOLOv10 also has a lower training cost compared to these models.\n|Model|AP$^{val}$ (%)|Latency (ms)|\n|-------------|--------------|------------|\n|YOLOv6-3.0-N|37.0|2.69|\n|Gold-YOLO-N|39.6|2.92|\n|**YOLOv10-N**|37.7|1.84|\n|YOLOv6-3.0-S|44.3|3.42|\n|Gold-YOLO-S|45.4|3.82|\n|YOLO-MS-XS|43.4|8.23|\n|**YOLOv10-S**|45.6|2.49|\n|YOLOv6-3.0-M|49.1|5.63|\n|Gold-YOLO-M|49.8|6.38|\n|**YOLOv10-M**|50.3|4.74|\n|YOLOv6-3.0-L|51.8|9.02|\n|Gold-YOLO-L|51.8|10.65|\n|YOLO-MS|51.0|12.41|\n|**YOLOv10-B**|51.6|5.74|\n|**YOLOv10-L**|52.4|7.28|\n|**YOLOv10-X**|53.6|10.70|\n[3] YOLOv8.\n\n[4] YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information, arXiv preprint 2024.\n\n[5] YOLOv6 v3.0: A Full-Scale Reloading, arXiv preprint 2023.\n\n[6] Gold-YOLO: Efficient object detector via gather-and-distribute mechanism, NeurIPS 2023.\n\n[7] YOLO-MS: YOLO-MS: rethinking multi-scale representation learning for real-time object detection, arXiv preprint 2023.", "category": "Experimental_Exposition"}, {"question": "How do the specific label assignment mechanisms in your work differ from those in H-DETR and LRANet?", "answer": "While there are structural similarities in the dual assignments among H-DETR, LRANet, and the paper' work, the specific label assignment mechanisms differ. LRANet employs the dense assignment and sparse assignment branches, both of which belong to the one-to-many assignment. In contrast, authors adopt the one-to-many and one-to-one branches. H-DETR introduces extra queries and repeats the ground truth several times to perform one-to-many matching using bipartite matching, while authors leverage prediction-aware one-to-many matching with the spatial prior of anchor points. Additionally, authors introduce the consistent matching metric for YOLOs to narrow the theoretical supervision gap between the one-to-many and one-to-one branches, improving supervision alignment and performance without hyper-parameter tuning.", "category": "Method_Comparison"}, {"question": "What are the key differences between the lightweight classification head and spatial-channel decoupled downsampling proposed in the work and those used in DAMO-YOLO?", "answer": "While lightweight classification head and spatial-channel decoupled downsampling have structural similarities to the ZeroHead and MobBlock used in DAMO-YOLO, they are motivated by different considerations and have distinct architectural designs. Specifically, the ZeroHead in DAMO-YOLO aims to balance computational cost between the neck and head using a single linear layer for both classification and regression tasks, whereas approach reduces redundancy in the classification head relative to the regression head with a lightweight design. Additionally, DAMO-YOLO applies MobBlock only to lightweight models through NAS, resulting in an identical structure for the basic block and downsampling module. In contrast, authors simplify the downsampling process for all model scales to reduce computational redundancy, decoupled from the basic block design. The work also introduces other new architectural designs, such as rank-guided block design and partial self-attention module, to improve performance and efficiency.", "category": "Method_Comparison"}, {"question": "How does the effectiveness of large-kernel convolution vary across different model scales of YOLOs, and how does this influence authors' methodology?", "answer": "Authors' findings indicate that large-kernel convolution provides performance gains for small-scale YOLO models but does not improve larger models. Therefore, authors selectively apply large-kernel convolutions based on model size to achieve a better efficient-accuracy trade-off. Additionally, authors introduce new architectural designs, such as the rank-guided block design and partial self-attention module, which further enhance performance.", "category": "Experimental_Exposition"}, {"question": "Which additional lightweight detectors have been included in the comparison in Table 1, and what are their performance metrics?", "answer": "The additional lightweight detectors included in the comparison are DAMO-YOLO (variants T, S, M, L) and YOLOv7 (variants standard and X). Their performance metrics are as follows: DAMO-YOLO-T (8.5M params, 18.1G FLOPs, 42.0% AP$^{val}$, 2.21ms latency), DAMO-YOLO-S (16.3M params, 37.8G FLOPs, 46.0% AP$^{val}$, 3.18ms latency), DAMO-YOLO-M (28.2M params, 61.8G FLOPs, 49.2% AP$^{val}$, 4.57ms latency), DAMO-YOLO-L (42.1M params, 97.3G FLOPs, 50.8% AP$^{val}$, 6.48ms latency), YOLOv7 (36.9M params, 104.7G FLOPs, 51.2% AP$^{val}$, 17.03ms latency), and YOLOv7-X (71.3M params, 189.9G FLOPs, 52.9% AP$^{val}$, 21.45ms latency).\n|Model|Param. (M)|FLOPs (G)|AP$^{val}$ (%)|Latency (ms)|\n|-------------|----------|---------|--------------|------------|\n|DAMO-YOLO-T|8.5|18.1|42.0|2.21|\n|DAMO-YOLO-S|16.3|37.8|46.0|3.18|\n|**YOLOv10-S**|7.2|21.6|46.3|2.49|\n|DAMO-YOLO-M|28.2|61.8|49.2|4.57|\n|DAMO-YOLO-L|42.1|97.3|50.8|6.48|\n|**YOLOv10-M**|15.4|59.1|51.1|4.74|\n|YOLOv7|36.9|104.7|51.2|17.03|\n|**YOLOv10-B**|19.1|92.0|52.5|5.74|", "category": "Experimental_Exposition"}, {"question": "How does the proposed dual label assignment strategy for YOLOs compare to the hybrid or multiple group label assignment strategies used in recent DETR-based detectors such as H-DETR, Group-DETR, and MS-DETR?", "answer": "The proposed dual label assignment strategy for YOLOs includes one-to-one matching and original one-to-many matching, providing rich supervision during training and high efficiency during inference. Unlike DETR-based detectors such as H-DETR, Group-DETR, and MS-DETR, which introduce extra queries, repeat ground truths, or select top queries for bipartite matching to achieve one-to-many matching, the proposed method uses prediction-aware assignment that incorporates spatial prior. Additionally, the method analyzes the supervision gap between the two heads and introduces a consistent matching metric for YOLOs, which improves performance through better supervision alignment and eliminates the need for hyper-parameter tuning.", "category": "Method_Comparison"}, {"question": "What are the comparative data on training time and costs for YOLOv10-M with and without the auxiliary head, and how does it compare to other detectors?", "answer": "The training cost for YOLOv10-M was measured on 8 NVIDIA 3090 GPUs. With the auxiliary head, there is an increase of about 18 seconds per epoch, resulting in a total training time of 29.1 hours for 500 epochs. Without the auxiliary head, the training time is 26.7 hours for 500 epochs. When extending training to 550 epochs without the auxiliary head, the total time increases to 29.3 hours, but performance is inferior (48.9% AP vs. 51.1% AP with the auxiliary head). YOLOv10 shows relatively low training costs compared to other detectors such as YOLOv6-3.0-M, YOLOv8-M, YOLOv9-M, Gold-YOLO-M, and YOLO-MS.\n|Model|Training Time (min/epoch)|Epochs|Total Time (hour)|\n|--------------------------------|-------------------------|------|-----------------|\n|YOLOv6-3.0-M|8.3|300|41.7|\n|YOLOv8-M|3.3|500|27.3|\n|YOLOv9-M|4.9|500|40.7|\n|Gold-YOLO-M|12.8|300|63.8|\n|YOLO-MS|8.5|300|42.3|\n|**YOLOv10-M** w/o auxiliary head|3.2|500|26.7|\n|**YOLOv10-M** w/ auxiliary head|3.5|500|29.1|", "category": "Method_Comparison"}, {"question": "What evidence supports the claim that the inherent complexity in deploying DETRs impedes its ability to attain the optimal balance between accuracy and speed?", "answer": "The evidence shows that when considering only the forward process of the model, RT-DETR's DETR-based architecture lags behind YOLO series models in inference efficiency for certain scales. For instance, RT-DETR-R34 and RT-DETR-R18 have higher latencies of 1.12ms and 2.16ms compared to YOLOv8-M and YOLOv8-S, respectively. Efficiency of DETRs still needs improvement compared to YOLOs during deployment.", "category": "Experimental_Analysis"}, {"question": "What are the speed comparison results of YOLOv10 and its competitors on CPU hardware?", "answer": "The speed comparison results of YOLOv10 and its competitors on CPU (Intel Xeon Skylake, IBRS) using OpenVINO show that YOLOv10 achieves superior performance and efficiency trade-offs. Specific results are provided in a table in the rebuttal.\n| Model| AP$^{val}$ (%)|Latency (ms)|\n|-------------|--------------|------------|\n|YOLOv6-3.0-N|37.0|32.22|\n|Gold-YOLO-N|39.6|36.35|\n|YOLOv8-N|37.3|32.91|\n|**YOLOv10-N**|38.5|27.84|\n|YOLOv6-3.0-S|44.3|93.11|\n|Gold-YOLO-S|45.4|98.08|\n|YOLO-MS-S|46.2|113.54|\n|YOLOv8-S|44.9|79.56|\n|RT-DETR-R18|46.5|157.40|\n|**YOLOv10-S**|46.3|67.70|\n|YOLOv6-3.0-M|49.1|174.82|\n|Gold-YOLO-M|49.8|188.33|\n|YOLO-MS|51.0|214.28|\n|YOLOv8-M|50.6|172.76|\n|RT-DETR-R34|48.9|222.82|\n|**YOLOv10-M**|51.1|147.99|\n|YOLOv6-3.0-L|51.8|325.85|\n|Gold-YOLO-L|51.8|344.99|\n|YOLOv9-C|52.5|244.14|\n|**YOLOv10-B**|52.5|210.98|\n|YOLOv8-L|52.9|330.02|\n|RT-DETR-R50|53.1|332.76|\n|**YOLOv10-L**|53.2|279.59|\n|YOLOv8-X|53.9|535.23|\n|RT-DETR-R101|54.3|570.25|\n|**YOLOv10-X**|54.4|399.87|", "category": "Experimental_Exposition"}, {"question": "Are DAMO-YOLO and YOLOv7 included in the comparisons presented in Table 1 ?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "SqW42eR2wC", "venue": "NEURIPS 2024", "paper_decision": "Reject", "title": "Offline Inverse Constrained Reinforcement Learning for Safe-Critical Decision Making in Healthcare", "authors": ["Nan Fang", "Guiliang Liu", "Wei Gong"], "abstract": "Reinforcement Learning (RL) applied in healthcare can lead to unsafe medical decisions and treatment, such as excessive dosages or abrupt changes, often due to agents overlooking common-sense constraints. Consequently, Constrained Reinforcement Learning (CRL) is a natural choice for safe decisions. However, specifying the exact cost function is inherently difficult in healthcare. Recent Inverse Constrained Reinforcement Learning (ICRL) is a promising approach that infers constraints from expert demonstrations. ICRL algorithms model Markovian decisions in an interactive environment. These settings do not align with the practical requirement of a decision-making system in healthcare, where decisions rely on historical treatment recorded in an offline dataset. To tackle these issues, we propose the Constraint Transformer (CT). Specifically, 1) utilize causal attention mechanism to incorporate historical decisions and observations into the constraint modeling and employ a non-Markovian layer for weighted constraints to capture critical states, 2) generative world model to perform exploratory data augmentation, thereby enabling offline RL methods to generate unsafe decision sequences. In multiple medical scenarios, empirical results demonstrate that CT can capture unsafe states and achieve strategies that approximate lower mortality rates, reducing the occurrence probability of unsafe behaviors.", "keywords": "Reinforcement Learning;Inverse Constrained Reinforcement Learning;Healthcare", "qa_pairs": [{"question": "How was the patient system modeled to address the limitations of the Markov process and incorporate high-dimensional latent states?", "answer": "The patient system was modeled as a CPOMDP to address the limitations of the Markov process. A non-Markovian model was constructed by incorporating historical state sequences into the RL decision-making process to capture potential states from history. The CPOMDP framework is defined in Section 3 Problem Formulation as $\\left( \\mathcal{S},\\mathcal{A},\\mathcal{O},\\mathcal{P},\\mathcal{Z},\\mathcal{R},\\mathcal{C},\\gamma ,\\kappa,\\rho_0 \\right)$, where $\\mathcal{O}$ is the set of observations $o$, $\\mathcal{P}\\left( s'|s,a \\right) =Pr\\left( s_{t+1}=s'|s_t=s,a_t=a \\right) $ the transition probability, and $\\mathcal{Z}\\left( o|s',a \\right) =Pr\\left( o_{t+1}=o|s_{t+1}=s',a_t=a \\right)$ is the observation probability.", "category": "Method_Mechanics"}, {"question": "What is the definition of $\\hat{\tau}$ in Equation (4), and what is the significance of $x_{t}$ in Equation (8)?", "answer": "In Equation (4), $\\hat{\tau}$ is the trajectory sampled from the executing policy (training policy) $\\mathcal{M}^{\\hat{\\zeta}_{\theta}}$. In an online environment, the execution policy is used to interact with the environment and generate trajectories $\\hat{\tau}$. In Equation (8), $x_t = \\{ h_t \\cup a_t \\} = \\left( R_{-K:t},\\; s_{-K:t},\\; a_{-K:t} \\right)$ is the input which includes the reward $R$, states $s$, and action $a$ from the preceding $K$ timesteps, where $K$ is the context length. The input is encoded by linear layers and passes through the casual transformer to predict hidden tokens $g_t$. These hidden tokens are then used to predict the current reward $r_{t-1}$ and the next state $s_t$ by minimizing the mean squared error for each of the linear layers $\\ell_{\\mu}$ and $\\ell_{\\varphi}$.", "category": "Concept_Understanding"}, {"question": "How are the transformer representations trained in relation to the policy model and the world model within the model-based offline RL framework, and what is the interaction mechanism?", "answer": "In the model-based offline RL framework, the transformer is trained together with the policy model and the world model. The policy model generates actions, the world model generates rewards and the next state, and the transformer extracts historical information for both models. The training process involves simultaneously minimizing Equations (7) and (8), during which the transformer is also trained alongside these objectives until convergence.", "category": "Experimental_Setup"}, {"question": "How does the reward function in authors' study account for the limitations of using mortality rate as a metric, and what measures have been taken to ensure it reflects treatment quality more accurately?", "answer": "Our reward function does not directly use mortality rates but includes intermediate rewards such as the SOFA score for sepsis. Authors designed different reward functions for different diseases based on previous literature. To address hidden variables, authors use a simple reward function with a cost function that accounts for changes in indicators like MAP and NEWS. Supplementary experiments show that the penalty function compensates for the reward function's shortcomings by increasing penalty values when NEWS scores are high or MAP is outside normal ranges.", "category": "Method_Mechanics"}, {"question": "Why is the evaluation metric \\omega not considered ideal, and what additional criteria are used to evaluate the safety of different policies?", "answer": "The evaluation metric \\omega is not ideal because it does not fully capture the complexity of treatment quality. To address this, authors also consider the relationship between the penalty value and medical metrics, as well as the probability of dangerous actions in the policy, to evaluate the safety of different policies from multiple perspectives. All methods in the experiments are based on the same reward function, ensuring a consistent comparison.", "category": "Experimental_Analysis"}, {"question": "Was the clinician's policy $\\hat{\\pi}$ approximated by a neural network in this study, and how was the potential issue of critical flaws in a small number of states addressed to ensure the reliability of off-policy evaluation results?", "answer": "In this study, the clinician's policy $\\hat{\\pi}$ was not initially approximated by a neural network. However, supplementary experiments were conducted where neural networks were used to fit the behavioral policy, the model was corrected accordingly, and offline policy evaluation (OPE) was performed. Additionally, the results of the policy under different evaluation metrics were compared, demonstrating that the CDT+CT method performed better on RMSE_IV, WIS, WIS_b, and WIS_bt metrics.", "category": "Method_Mechanics"}, {"question": "What is the sensitivity of the estimated policy to the generative world model, and how does the target reward influence this sensitivity?", "answer": "The sensitivity of the estimated policy to the generative world model was explored by varying the target reward (1, 5, 10, 20, and 30). As the target reward increases, the world model generates more aggressive data, leading to improved policy performance. However, this improvement has an upper limit and does not increase indefinitely. The results are detailed in Table 2.", "category": "Method_Mechanics"}, {"question": "Does the distribution difference between the expert dataset and the violating dataset significantly impact the learning of the constraint?", "answer": "The impact exists but is not significant. The generative model used for the non-compliant dataset is a reinforcement learning (RL) model that generates data within legal boundaries, reducing the likelihood of extreme values. However, it still presents a potential risk as previous work confirms its ability to generate non-compliant data.", "category": "Experimental_Analysis"}, {"question": "How is the DIFF between the estimated policy and the physicians' policy calculated, and how is this represented graphically in Figure 7?", "answer": "The DIFF is calculated by subtracting the drug dosage (a) under the doctor's policy from the drug dosage (b) under the estimated policy for the same patient state ($b−a$). In Figure 7, the x-axis shows the DIFF values(the difference in drug dosage between the two policies), and the y-axis shows the mortality rate under the doctor's policy. The point where the x-axis is 0, indicating no difference between policies, shows the lowest mortality rate, suggesting the estimated policy is closer to an ideal policy with lower mortality.", "category": "Method_Mechanics"}, {"question": "Why is the term 'unsafe behavior' used in Table 1 to describe a drug dosage that is 'too high', and how has the issue of conditioning on the patient state been addressed?", "answer": "The term 'unsafe behavior' refers to a drug dosage that is dangerous for any patient state. To address the issue of conditioning on the patient state, the authors have developed a personalized cost function $C(\tau)$, which assigns a cost to the current drug dosage based on the patient's historical treatment trajectory.", "category": "Method_Mechanics"}, {"question": "What are the potential issues with incorporating history into the state or redefining the state to address the limitations of the Markov assumption?", "answer": "Incorporating history into the state or redefining the state can lead to redundant calculations and increased complexity. For example, for a patient's trajectory, the state at timestamp t would include the previous t-1 timestamps, increasing the number of timestamps stored from O(n) to O(n^2). Additionally, latent states might be unobservable, making it impossible to model the patient system based on a Markov process due to the high-dimensional latent states that cannot be represented.", "category": "Method_Disambiguation"}, {"question": "What is the definition and selection process of 'top N' patients in the context of Eqn (2) and L144?", "answer": "The 'top N' patients are selected from a dataset of 2N patients, where N is a constant. Among these 2N patients, N patients died under the doctor's treatment, and N patients survived. The DIFF for these 2N patients is calculated and ranked in ascending order, and the top N patients are those with the highest DIFF values.", "category": "Method_Mechanics"}, {"question": "What is the meaning of $R(\\tau)$, $Z_{\\mathcal{M}^{\\mathcal{C}}}$, and $\\beta$ in Eqn (3) and L157?", "answer": "$R(\\tau)$ represents the reward of the trajectory $\\tau$. $Z_{\\mathcal{M}^{\\mathcal{C}}}$ is the correct notation at L157. $\\beta$ is a parameter that describes how close the agent is to the optimal distribution.", "category": "Concept_Understanding"}, {"question": "How is the cost function $C_{\\theta}$ formulated and how does it relate to the MDP $\\mathcal{M}^{\\zeta_{\\theta}}$ and the policy $\\pi_{\\mathcal{M}^{\\hat{\\zeta}_{\\theta}}}$?", "answer": "The cost function $C_{\\theta}$ is formulated as $C_{\\theta}=1–\\zeta_{\\theta}$. The MDP $\\mathcal{M}^{\\zeta_{\\theta}}$ is obtained by adding this cost function to the original MDP $\\mathcal{M}$. The executing policy for this augmented MDP is denoted as $\\pi_{\\mathcal{M}^{\\hat{\\zeta}_{\\theta}}}$.", "category": "Concept_Understanding"}, {"question": "Does the current reward function account for hidden variables such as historical medical conditions and phenotypes?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did supplementary experiments show that the penalty value increases when the NEWS score is too high or when MAP is outside the normal range?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the reward function in the study directly utilize mortality rates?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the evaluation metric $\\omega$ considered the sole ideal measure for evaluating the safety of different policies?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "XII0Wp1XA9", "venue": "COLM 2024", "paper_decision": "Accept", "title": "A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration", "authors": ["Zijun Liu", "Yanzhe Zhang", "Peng Li", "Yang Liu", "Diyi Yang"], "abstract": "Recent studies show that collaborating multiple large language model (LLM) powered agents is a promising way for task solving. However, current approaches are constrained by using a fixed number of agents and static communication structures. In this work, we propose automatically selecting a team of agents from candidates to collaborate in a dynamic communication structure toward different tasks and domains. Specifically, we build a framework named Dynamic LLM-Powered Agent Network ($\\textbf{DyLAN}$) for LLM-powered agent collaboration, operating a two-stage paradigm: (1) Team Optimization and (2) Task Solving. During the first stage, we utilize an agent selection algorithm, based on an unsupervised metric called Agent Importance Score, enabling the selection of best agents according to their contributions in a preliminary trial, oriented to the given task. Then, in the second stage, the selected agents collaborate dynamically according to the query. Empirically, we demonstrate that DyLAN outperforms strong baselines in code generation, decision-making, general reasoning, and arithmetic reasoning tasks with moderate computational cost. On specific subjects in MMLU, selecting a team of agents in the team optimization stage improves accuracy by up to 25.0% in DyLAN.", "keywords": "l;a;r;g;e; ;l;a;n;g;u;a;g;e; ;m;o;d;e;l;,; ;L;L;M;-;p;o;w;e;r;e;d; ;a;g;e;n;t;,; ;m;u;l;t;i;-;a;g;e;n;t; ;c;o;l;l;a;b;o;r;a;t;i;o;n;,; ;t;e;a;m; ;o;p;t;i;m;i;z;a;t;i;o;n", "qa_pairs": [{"question": "What are the reasons for the varying impact of early stopping on API calls across arithmetic reasoning (AR), general reasoning (GR), code generation (CG), and decision-making (DM) tasks?", "answer": "The reasons for the varying impact of early stopping are as follows:\n\n- **GR and AR Tasks:** Early stopping is effective when answers from agents are exactly the same. For GR tasks, problems are simpler, allowing agents to reach consensus earlier. For AR tasks, exact matching is used for mathematical problems, but the complexity makes early stopping less effective than in GR.\n- **CG Tasks:** Early stopping is less effective because answer comparison is challenging for open-ended tasks. A BLEU score with a 0.9 threshold is used for consistency checks, but agents may generate codes in different formats, reducing opportunities to stop early.\n- **DM Tasks:** The large action space and multi-step nature of DM tasks make early stopping impactful. Incorrect actions can lead to agents continuing until the maximum step is reached, significantly affecting the results.", "category": "Experimental_Analysis"}, {"question": "Why are error bars absent in the experimental results, and how was variance accounted for in the absence of error bars?", "answer": "Error bars are absent in the experimental results for two main reasons: 1. The variance is highly influenced by the hyper-parameter 'temperature,' which was varied across different methods(e.g., {0, 0.2, 0.8, 1.0} for GR&AR), and this variance primarily reflects the impact of temperature rather than additional insights about the method. 2. Due to budget constraints, only three runs were conducted for methods with non-zero temperatures, which is insufficient to calculate meaningful error bars(e.g., 1-sigma or 2-sigma bars). As a result, the median result was reported to provide a reliable representation within the resource limitations.", "category": "Experimental_Analysis"}, {"question": "How does MetaGPT compare to DyLAN and bare GPT-4-0613 in terms of performance, methodology, and adaptability to non-programming tasks?", "answer": "In terms of performance, MetaGPT (85.9) performs worse than bare GPT-4-0613 (88.4) with task instructions and requires significantly more API calls for each sample in code generation tasks. DyLAN (92.1) achieves a 7.2% improvement against MetaGPT. In terms of methodology, MetaGPT does not incorporate agent filtering methods, whereas DyLAN selects a team of agents from an initial pool, unlike MetaGPT's fixed framework with five predefined roles. Additionally, MetaGPT's framework is not easily adaptable to non-programming tasks.", "category": "Method_Comparison"}, {"question": "What are the reasons for the limited improvement of DyLAN on the AR task compared to other tasks?", "answer": "The limited improvement of DyLAN on the AR task is due to two main reasons: (1) The AR task uses the MATH dataset, which is highly knowledge-dependent, making it more challenging. Despite this, DyLAN still outperforms baselines with relative improvements of 10.2% and 3.0% compared to the best baseline, as shown in Table 2. (2) Role-playing introduces relatively lower improvements in AR tasks compared to other tasks, as the high knowledge dependency of MATH limits the impact of role assignment. In contrast, selecting appropriate agents for less knowledge-dependent tasks, such as math subjects in MMLU, results in greater performance improvement, as shown in Table 13.", "category": "Experimental_Analysis"}, {"question": "What is the rationale behind the selection of varied agent roles (e.g., 'programmer', 'algorithm developer', 'mathematician', 'doctor', 'economist') in the DyLAN framework, and how does their inclusion impact the performance and generalizability of the system?", "answer": "The selection of varied agent roles is based on the challenge of identifying the most contributory agents in a principled way. While some agents might appear redundant for specific tasks, empirically, fewer properly selected agents can outperform larger teams with redundancy. The Agent Importance Score is used as an unsupervised metric to identify contributory agents, but it is not an oracle. Seemingly irrelevant agents from a human perspective may still make significant contributions, as shown in Appendix C.5. Including various roles enhances generalization, as different tasks benefit from distinct selected teams (Table 10). However, if the majority of agents are irrelevant to the task by design (e.g., coding agents for DM tasks), the results may be negatively impacted. The chosen team size in the settings strikes a balance between effectiveness and efficiency (Fig.3 and Table 8).", "category": "Motivation_Analysis"}, {"question": "How does the DyLAN framework ensure scalability and flexibility across tasks with diverse roles, and what is the necessity for prompt curation in this context?", "answer": "The DyLAN framework is designed to be general and treats each agent at each time step equally. The pattern observed in Figures 5 and 6, where even-numbered layers evaluate responses from odd-numbered layers, is consistent across tasks like AR, DM, and GR. In GR tasks with 57 different subjects, DyLAN improves performance by selecting teams from varied agents. The system prompt for each agent is universal, and task-specific instructions are minimal (usually one sentence), ensuring adaptability without the need for extensive prompt curation.", "category": "Motivation_Analysis"}, {"question": "Are error bars included in the experimental results to indicate variance?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the variance in the experimental results primarily due to the hyper-parameter 'temperature' as detailed in Appendix B.1?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were only three runs conducted for methods with non-zero temperatures due to budget constraints?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the median result reported instead of error bars due to insufficient samples for meaningful error bar calculation?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "zl16jLb91v", "venue": "COLM 2024", "paper_decision": "Accept", "title": "Towards Measuring the Representation of Subjective Global Opinions in Language Models", "authors": ["Esin DURMUS", "Karina Nguyen", "Thomas Liao", "Nicholas Schiefer", "Amanda Askell", "Anton Bakhtin", "Carol Chen", "Zac Hatfield-Dodds", "Danny Hernandez", "Nicholas Joseph", "Liane Lovitt", "Sam McCandlish", "Orowa Sikder", "Alex Tamkin", "Janel Thamkul", "Jared Kaplan", "Jack Clark", "Deep Ganguli"], "abstract": "Large language models (LLMs) may not equitably represent diverse global perspectives on societal issues. In this paper, we develop a quantitative framework to evaluate whose opinions model-generated responses are more similar to. We first build a dataset, GlobalOpinionQA, comprised of questions and answers from cross-national surveys designed to capture diverse opinions on global issues across different countries. Next, we define a metric that quantifies the similarity between LLM-generated survey responses and human responses, conditioned on country. With our framework, we run three experiments on an LLM trained to be helpful, honest, and harmless with Constitutional AI. By default, LLM responses tend to be more similar to the opinions of certain populations, such as those from the USA, and some European and South American countries, highlighting the potential for biases. When we prompt the model to consider a particular country's perspective, responses shift to be more similar to the opinions of the prompted populations, but can reflect harmful cultural stereotypes. When we translate GlobalOpinionQA questions to a target language, the model's responses do not necessarily become the most similar to the opinions of speakers of those languages. We will release our dataset for others to use and build on upon acceptance.", "keywords": "o;p;i;n;i;o;n;s; ;o;n; ;g;l;o;b;a;l; ;i;s;s;u;e;s;,; ;c;u;l;t;u;r;a;l; ;r;e;p;r;e;s;e;n;t;a;t;i;o;n;,; ;s;u;b;j;e;c;t;i;v;i;t;y; ;i;n; ;L;L;M;s;,; ;e;v;a;l;u;a;t;i;o;n;,; ;s;o;c;i;e;t;a;l; ;i;m;p;a;c;t", "qa_pairs": [{"question": "Why was the approach tested on only a single model, and how does this affect the generalizability of the results to other large language model (LLM) families and models of different sizes?", "answer": "Testing on only a single model was due to the scope of the work, which focused on providing an initial benchmark dataset and starting points for experiments on cultural representation in language models (e.g., cross-lingual prompting, linguistic prompting). While this limits the generalizability of the results to other LLM families and model sizes, the authors acknowledge this as an important direction for follow-up work. They aimed to maximize coverage by including experiments on four languages and several countries.", "category": "Method_Mechanics"}, {"question": "What is the novel contribution of GlobalOpinionQA, given that the questions are sourced from the Global Attitudes Survey and World Values Survey?", "answer": "GlobalOpinionQA introduces novel methods that use these questions to measure cultural values of language models, focusing on studying cultural representations in models. Additionally, the data release will provide questions from both sources separately for independent analyses.", "category": "Method_Disambiguation"}, {"question": "What are the implications of the findings in Section 3 regarding the language independence or consistency of model predictions, and how does this relate to the desirability of such properties?", "answer": "The findings in Section 3 suggest that purely linguistic cues do not significantly alter model predictions, which could indicate language independence or consistency. The authors acknowledge that while language consistency might be desirable in contexts like maintaining factual accuracy, it may be less so in areas such as cultural sensitivity. The paper does not take a definitive stance but presents an empirical analysis highlighting that linguistic cues alone may not adequately represent different cultural identities.", "category": "Experimental_Analysis"}, {"question": "Why did the authors choose a multiple-choice setup for evaluating LLMs’ values, given that it has been criticized in previous works, and how did they address its limitations?", "answer": "The authors chose a multiple-choice setup because it provides valuable insights, particularly for a large-scale study, despite its limitations. To address these limitations, they complemented their quantitative analysis with qualitative analysis of open-ended model generations, examples of which are provided in the appendix. Additionally, they will clarify the limitations of the multiple-choice format by referencing relevant sources in the final version.", "category": "Motivation_Analysis"}, {"question": "What are the main takeaways and practical implications of the results presented in the paper?", "answer": "The main takeaways of the work are: 1) Understanding cultural representation in language models is important due to the societal implications of their biases. 2) Methods to measure cultural representations are necessary to mitigate these issues, and the paper proposes a benchmark and methods in this direction. 3) The findings reveal that models often align more with Western perspectives and can propagate harmful stereotypes, emphasizing the need for effective mitigation strategies.", "category": "Experimental_Analysis"}, {"question": "Is individual demographic information available for the human results beyond aggregated selections by country?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "c8McWs4Av0", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Solving Challenging Math Word Problems Using GPT-4 Code Interpreter with Code-based Self-Verification", "authors": ["Aojun Zhou", "Ke Wang", "Zimu Lu", "Weikang Shi", "Sichun Luo", "Zipeng Qin", "Shaoqing Lu", "Anya Jia", "Linqi Song", "Mingjie Zhan", "Hongsheng Li"], "abstract": "Recent progress in large language models (LLMs) like GPT-4 and PaLM-2 has brought significant advancements in addressing math reasoning problems. In particular, OpenAI's latest version of GPT-4, known as GPT-4 Code Interpreter, shows remarkable performance on challenging math datasets. In this paper, we explore the effect of code on enhancing LLMs' reasoning capability by introducing different constraints on the Code Usage Frequency of GPT-4 Code Interpreter. We found that its success can be largely attributed to its powerful skills in generating and executing code, evaluating the output of code execution, and rectifying its solution when receiving unreasonable outputs. Based on this insight, we propose a novel and effective prompting method, explicit $\\underline{\\text{c}}$ode-based $\\underline{\\text{s}}$elf-$\\underline{\\text{v}}$erification (CSV), to further boost the mathematical reasoning potential of GPT-4 Code Interpreter. This method employs a zero-shot prompt on GPT-4 Code Interpreter to encourage it to use code to self-verify its answers. In instances where the verification state registers as \"False\", the model shall automatically amend its solution, analogous to our approach of rectifying errors during a mathematics examination. Furthermore, we recognize that the states of the verification result indicate the confidence of a solution, which can improve the effectiveness of majority voting. With GPT-4 Code Interpreter and CSV, we achieve an impressive zero-shot accuracy on MATH dataset.", "keywords": "mathematical reasoning;large language models;zero-shot learning;code generation;prompting", "qa_pairs": [{"question": "Does the proposed prompt overfit to simple math problems due to the use of only one specific kind of benchmark (i.e., grade school math problems)?", "answer": "The MATH dataset consists of challenging competition-level problems across multiple math fields, including algebra, number theory, geometry, and precalculus. Many of these problems are challenging even for college students. The proposed prompt significantly improves performance on the MATH dataset (69.69% -> 84.32%), demonstrating its potential for solving difficult math problems. For relatively easier datasets like GSM8K and MMLU-Math, the proposed CSV + VW-Voting method shows smaller yet noticeable improvements, which is expected given their already high baseline performance. To fully appreciate the improvement induced by the method, authors would like to calculate Relative error reduction Ratio, which is defined as $Ratio = \\frac{Err_{baseline}-Err_{our}}{Err_{baseline}}$, where $Err = 1 - Acc$. The Relative Error Reduction Ratio further quantifies these improvements, with GSM8K, MMLU-Math, and MATH showing Ratios of 57.7%, 56.0%, and 48.3%, respectively.\n| GSM8K                       | Err (%)| Ratio (%) | \n|:---------------------------|:------:|:---------:| \n| GPT4-Code (baseline)        | 7.1    |  --       |\n | GPT4-Code + CSV + VW-Voting | 3.0    | 57.7      |\n | MMLU-Math                   | Err (%)| Ratio (%) |\n |:---------------------------|:------:|:---------:| \n| GPT4-Code (baseline)        | 12.5   |  --       |\n | GPT4-Code + CSV + VW-Voting | 5.5    | 56.0      |\n | MATH                        | Err (%)| Ratio (%) | \n|:---------------------------|:------:|:---------:| \n| GPT4-Code (baseline)        | 30.31  |  --       |\n | GPT4-Code + CSV + VW-Voting | 15.68  | 48.3      |", "category": "Experimental_Analysis"}, {"question": "What changes were made to Table 1 to address the comparison between GPT4-Code with 1 sample and 16 samples, and the inclusion of Naive Majority Voting results?", "answer": "In Table 1, the authors added a line for **GPT4-Code + CSV + Voting** with an accuracy of 83.54% and a line for **GPT4-Code + Voting** with an accuracy of 79.88%, where Voting represents Naive Majority Voting. GPT4-Code + Voting improved the baseline by 10.19%, demonstrating the diversity and accuracy of GPT4-Code solutions. Additionally, GPT4-Code + CSV + Voting showed a 3.66% improvement over GPT4-Code + Voting, consistent with the 3.85% improvement from GPT4-Code to GPT4-Code + CSV. ", "category": "Method_Mechanics"}, {"question": "Why are there missing rows for GPT4-Code + CSV in Table 2 and GPT4-Code + CSV + Voting in Table 3, and how have these omissions been addressed?", "answer": "The missing rows were due to GPT-4 quota limits at the time of submission, which prevented multiple runs on MMLU-Math. The oversight for GPT4-Code + CSV on GSM8K has been corrected with a result of 94.5, and the result for GPT4-Code + CSV + Voting on MMLU-Math, also 94.5, has been included in Table 3.", "category": "Experimental_Analysis"}, {"question": "What dataset is used in Figure 6a, and why are there five separate paths graphed in the figure?", "answer": "The dataset used in Figure 6a is MATH. The five separate paths represent the results of five instances of the same experiment, intended to demonstrate the robustness of the performance. ", "category": "Experimental_Analysis"}, {"question": "Why are some of the red dots in Figure 2 different sizes, despite the prompt stating that no code is allowed? Does this indicate that the model uses code in some cases?", "answer": "The different sizes of the red dots in Figure 2 indicate that, in some cases, the model used code despite the prompt stating no code is allowed. The Code Usage Frequency values for these dots are 0.025, 0.040, 0.046, 0.038, and 0.106, which are very small and do not pose a serious problem to the analysis. As stated in Section 3.1, 'Prompt 1 results in almost negligible code usage.' However, the prompt cannot completely prevent the model from using code, especially when the problem is very difficult. To clarify, the caption now includes the sentence: 'The red points denoting Prompt 1 show that the model still occasionally uses code, especially when the problem is very difficult. However, even then, the Code Usage Frequency is negligible.'", "category": "Experimental_Analysis"}, {"question": "What are the novel aspects of this work in comparison to existing literature on code writing, self-validating, and majority voting for improving LLMs?", "answer": "The novelty of this work lies in four key aspects: (1) It is the first to quantitatively analyze the role of code in GPT-4 Code Interpreter’s high performance in solving math problems via prompt engineering, including its ability to analyze execution results and self-repair. (2) It employs zero-shot prompting for automatic code-based self-verification, unlike existing methods that use additional validator mechanisms. (3) It introduces verification-guided weighted majority voting, improving upon traditional majority voting, as demonstrated in Figure 6. (4) It explores the potential of using code multiple times, interleaved with natural language reasoning, whereas existing methods like PAL and PoT generally use code only once. These aspects distinguish the work from existing literature and provide novel insights into enhancing LLMs.", "category": "Method_Comparison"}, {"question": "Do the results and findings reported in this work generalize to other publicly available LLMs, such as CodeLlama 7B and 34B, and how do these models perform compared to GPT-4?", "answer": "The method was tested on CodeLlama 7B and 34B using the same prompting of CSV in a zero-shot manner. The results show noticeable improvements on GSM8K and MATH tasks for both models, though their accuracy is lower compared to GPT4-Code. Specifically, CodeLlama-7B achieved 20.85% on GSM8K and 10.18% on MATH, while CodeLlama-34B achieved 37.6% on GSM8K and 13.36% on MATH.\n\n| Model               | GSM8K Acc (%) | MATH Acc (%) |\n|---------------------|----------------|---------------|\n| CodeLlama-7B        | 17.44          | 6.56          |\n| CodeLlama-7B + CSV  | 20.85          | 10.18         |\n| CodeLlama-34B       | 28.96          | 9.12          |\n| CodeLlama-34B + CSV | 37.60          | 13.36         |\n", "category": "Experimental_Exposition"}, {"question": "Why is the voting method not mentioned in the last row of Table 3, and was this an error?", "answer": "The absence of voting results in the last row of Table 3 was not a typo. It was due to GPT-4 quota limits, which prevented running MMLU-Math multiple times to obtain voting results. The table has been updated to include the outcome for GPT4-Code + CSV + VW-Voting on MMLU-Math, achieving 94.5% accuracy.", "category": "Experimental_Analysis"}, {"question": "What factors might prevent the same idea from improving other large language models like Llama 2, and how does the performance of the proposed method on CodeLlama compare to Llama 2 under similar conditions?", "answer": "The factors preventing the same idea from improving Llama 2 include the significant size difference between Llama2-70B and CodeLlama-34B, Llama2's weaker code generation capabilities, and the lack of access to a same-scale CodeLlama model. Performance metrics for the method on CodeLlama-34B exceed those of Llama2-34B on GSM8K and MATH benchmarks by 4.85 and 9.16 points, respectively. Additionally, metrics are based on zero-shot prompting, unlike Llama2's settings, making direct comparisons inequivalent. Significant performance gains observed with the method on CodeLlama-34B (8.64% on MATH and 4.24% on GSM8K) suggest substantial improvements not attributable to variance.", "category": "Method_Comparison"}, {"question": "What is the distribution of consistency between the reasoning process and the verification process, and how does it relate to the correctness of the model's output?", "answer": "The authors analyzed 100 samples and categorized them into four cases based on the correctness of the reasoning and verification processes. The distribution is as follows: 76% of samples have both correct reasoning and verification, 5% have correct reasoning but wrong verification, 7% have wrong reasoning but correct verification, and 12% have both wrong reasoning and verification. This indicates that in most cases, the reasoning and verification processes are consistent. A pie chart visualizing this distribution and additional analysis have been added to Appendix I (page 25) of the paper.\nTo better visualize the results, authors present a pie chart showing the distribution in an anonymous link [1].\n|Case| Reasoning Process | Verification Process| Percentage (%) |\n|:-------------------:|:-------------------:|:-----------:|:---------------|\n| 1 | Correct | Correct     |  76 |\n| 2 |  Correct   | Wrong      | 5 |\n| 3 |  Wrong   | Correct      | 7 |\n| 4 |  Wrong   | Wrong      | 12 |\n[1] https://anonymous.4open.science/r/verify_consistency-F602/README.md", "category": "Experimental_Exposition"}, {"question": "What types of verification code are generated by the CSV method, and what is their distribution among sampled problems?", "answer": "The types of verification code generated by the CSV method are Substitution, Alternative Method, Double Checking, and Approximation. Their distribution among 50 sampled problems is 50% for Substitution, 22% for Alternative Method, 18% for Double Checking, and 10% for Approximation. This analysis and a corresponding pie chart have been included in Appendix J of the updated paper. Authors also present a pie chart showing this distribution through an anonymous link [1]\n|Case| Verification Type | Percentage (%) |\n|:-------------------|:-------------------|:---------------|\n| 1 | Substitution |  50 |\n| 2 | Alternative Method  | 22 |\n| 3 | Double Checking  | 18 |\n| 4 | Approximation  | 10 |\n[1] https://anonymous.4open.science/r/CSV_rebuttal_examples-FE7D/verify_type_code.md", "category": "Method_Disambiguation"}, {"question": "What is the tradeoff between the depth of self-repair (continuing self-repair until verification succeeds) and the breadth of self-verification (repairing once but generating multiple samples for weighted sampling)?", "answer": "Due to the limited window size of GPT4-Code, unlimited self-repair until verification succeeds is not feasible. Instead, authors compared unlimited self-repair with majority voting of limited self-repair. Results show that limiting self-repair to one attempt with voting (k=3) achieves a higher accuracy (77.02%) compared to unlimited self-repair (73.54%). However, the voting approach requires generating three solutions per problem, which is more costly than the average 1.22 attempts in unlimited self-repair.\n|           | MATH Acc (%) |\n|:----------:|:----------:|\n| Unlimited (average depth 1.22) | 73.54 |\n| self-repair once + voting (k=3) | 77.02 |", "category": "Method_Comparison"}, {"question": "What are the novel contributions of this paper, particularly in terms of the role of code in GPT-4 Code Interpreter’s performance, zero-shot prompting for self-verification, weighted voting, and the use of code multiple times interleaved with natural language reasoning?", "answer": "The novel contributions of this paper include: (1) being the first to quantitatively analyze the role of code in GPT-4 Code Interpreter’s high performance in solving math problems via prompt engineering, including its ability to analyze execution results and self-repair; (2) employing zero-shot prompting for automatic code-based self-verification, which differs from existing works that use additional validator mechanisms; (3) introducing verification-guided weighted majority voting, which improves upon existing majority voting methods and has not been studied before; and (4) exploring the potential of using code multiple times interleaved with natural language reasoning, unlike current methods such as PAL and PoT, which generally use code only once. These contributions provide novel insights into the capabilities of GPT-4 Code Interpreter and enhance its performance through self-verification and weighted voting.", "category": "Method_Disambiguation"}, {"question": "What are the specific contributions of the authors to the improvements on the MATH dataset, and how do they acknowledge the role of GPT4-Code in achieving these results?", "answer": "The authors' contributions include analyzing the underlying code mechanism, developing a code-based self-verification (CSV) prompt, and introducing a weighted majority voting method. They also conducted quantitative experiments on various datasets. The authors acknowledge the high performance of GPT4-Code, which achieves 69.69% accuracy on the MATH dataset, largely surpassing the previous SOTA result (53.90%)and use this result as their baseline. They have modified tables and figures and adjusted expressions in the paper to more clearly demonstrate their contributions and consistently mention that their improvements are based on GPT4-Code.", "category": "Experimental_Exposition"}, {"question": "What is the distinction between GPT4-Code and prior frameworks (Lightman et al., 2023; Cobbe et al., 2021) in terms of verification methodologies, and why is zero-shot capability applicable in GPT4-Code?", "answer": "The distinction lies in the fact that prior frameworks relied on an external Large Language Model (LLM) and well-constructed few-shot prompts for natural language verification, whereas GPT4-Code's robust capabilities enable the paper' approach to depend solely on a straightforward prompt, thereby operating in a zero-shot manner. The zero-shot capability is applicable because GPT4-Code leverages OpenAI's instruction-tuned model, which allows for effective zero-shot prompting without the need for extensive few-shot examples.", "category": "Method_Comparison"}, {"question": "Has OpenAI's GPT-4 Code incorporated the PRM methodology (Lightman et al., 2023) for RL tuning, and how does its performance compare to PRM on the 'minival' subset of the MATH dataset?", "answer": "Authors cannot confirm if GPT-4 Code incorporates the PRM methodology. PRM is tailored for math problem-solving, while GPT-4 Code is a general-purpose code interpreter. PRM achieved 78.2% accuracy on the 'minival' subset, whereas GPT-4 Code achieved 70.8% accuracy. The method, GPT-4 Code + CSV, achieved 75.2% accuracy. The approaches differ fundamentally, with PRM using a reward model for step verification and method using CSV prompting for solution validation.\n\n | minival       | Accuracy  |\n |:--------------|:---------:| \n| PRM(Lightman et al., 2023) | 78.2|\n |GPT-4-Code      | 70.8            | \n| GPT4-Code + CSV (Authors)    | 75.2 | ", "category": "Method_Comparison"}, {"question": "Was the absence of voting results in the last row of Table 3 a typo?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the authors modify Tab. 1, Fig. 2, and Fig. 6 to clarify their contributions and the role of GPT4-Code?", "answer": "True", "category": "Claim_Verification"}, {"question": "Has Table 3 been updated to include the outcome for GPT4-Code + CSV + VW-Voting on MMLU-Math?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the method used in the paper to enhance GPT-4 Code's performance involve a reward model similar to PRM (Lightman et al., 2023)?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the initial experiment setup include running MMLU-Math multiple times to obtain voting results due to GPT-4 quota limits?", "answer": "False", "category": "Claim_Verification"}, {"question": "Have the missing setups for GPT4-Code + CSV and GPT4-Code + CSV + Voting on the MMLU-Math dataset been added to Tables 2 and 3, respectively, after the initial submission?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the GPT4-Code + CSV setup initially included in Table 2 for the MMLU-Math dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is it confirmed that GPT-4 Code incorporates the PRM methodology from Lightman et al., 2023?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does GPT-4 Code achieve higher accuracy than PRM (Lightman et al., 2023) on the minival subset of the MATH dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the GPT4-Code + CSV + Voting setup initially included in Table 3 for the MMLU-Math dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Table 1 in the paper now include a line for GPT4-Code + CSV + Majority?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "oKn9c6ytLx", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "WebArena: A Realistic Web Environment for Building Autonomous Agents", "authors": ["Shuyan Zhou", "Frank F. Xu", "Hao Zhu", "Xuhui Zhou", "Robert Lo", "Abishek Sridhar", "Xianyi Cheng", "Tianyue Ou", "Yonatan Bisk", "Daniel Fried", "Uri Alon", "Graham Neubig"], "abstract": "With advances in generative AI, there is now potential for autonomous agents to manage daily tasks via natural language commands. However, current agents are primarily created and tested in simplified synthetic environments, leading to a disconnect with real-world scenarios. In this paper, we build an environment for language-guided agents that is highly realistic and reproducible. Specifically, we focus on agents that perform tasks on the web, and create an environment with fully functional websites from four common domains: e-commerce, social forum discussions, collaborative software development, and content management. Our environment is enriched with tools (e.g., a map) and external knowledge bases (e.g., user manuals) to encourage human-like task-solving. Building upon our environment, we release a set of benchmark tasks focusing on evaluating the functional correctness of task completions. The tasks in our benchmark are diverse, long-horizon, and designed to emulate tasks that humans routinely perform on the internet. We experiment with several baseline agents, integrating recent techniques such as reasoning before acting.  The results demonstrate that solving complex tasks is challenging: our best GPT-4-based agent only achieves an end-to-end task success rate of 14.41%, significantly lower than the human performance of 78.24%. These results highlight the need for further development of robust agents, that current state-of-the-art large language models are far from perfect performance in these real-life tasks, and that \\ours can be used to measure such progress.\\footnote{Code, data, environment reproduction instructions, video demonstrations are available in the supplementary.}", "keywords": "language-guided agents; web automation; benchmark", "qa_pairs": [{"question": "What are the key differences between AndroidEnv and WebArena in terms of reproducibility, task complexity, and application focus, and how do these differences address the challenges mentioned in the related work?", "answer": "The key differences between AndroidEnv and WebArena are: (1) Reproducibility: AndroidEnv uses live Android apps that change over time, leading to challenges like CAPTCHAs, content modifications, and configuration changes, which make fair comparisons difficult. WebArena resolves these challenges by providing a 'frozen' environment. (2) Task Complexity: AndroidEnv faces limitations in diverse or complex task availability, while WebArena allows for the definition of new tasks within its framework. (3) Application Focus: AndroidEnv focuses on Android applications, whereas WebArena focuses on browser-based applications, with differences in UI layouts and task settings. These differences are clarified in the updated related work section.", "category": "Method_Comparison"}, {"question": "Did the authors evaluate agents using prompting methods other than Chain-of-Thought (CoT) and Direct, and if so, what were the findings?", "answer": "Yes, the authors performed preliminary experiments with agents using other prompting methods, such as RCI [1] and Reflexion [2], which involve planning and critique. However, they did not proceed with the full experiment because the performance on a subset of examples was significantly lower. This was due to inaccuracies in planning and critiquing based on the current observation, as the tasks in WebArena span multiple pages and general-purpose LLMs like GPT-4 lack specific knowledge about the websites in the benchmark. For example, in a task to 'Checkout my pending orders, the agent first plans to perform the action `click [1283] to go to the \"My Account\" page.`, its critique component responds that `It assumes that clicking on the \"My Account\" link ([1283]) will take you to the page where pending orders are displayed. However, there is no guarantee that this is the case. The \"My Account\" page could have a different purpose or may not even exist.`. However, there is indeed a 'My order' link inside 'My Account.' ' the agent's plan was vague, leading to repeated actions and task failure. ", "category": "Experimental_Exposition"}, {"question": "What are the success rates of using HTML observation versus accessibility tree observation with `gpt-3.5-turbo-16k-0613`, and what are the qualitative differences between the two approaches?", "answer": "The success rates are 4.43% for HTML observation and 8.75% for accessibility tree observation. Empirically, accessibility trees yield better performance. Qualitatively, HTML is longer and contains more redundancy, such as nested layers of elements, which increases the burden on LLMs to localize and use relevant information.\n|                    | Success Rate (%) |\n|--------------------|------------------|\n| HTML               | 4.43             |\n| Accessibility tree | 8.75             |\n", "category": "Experimental_Exposition"}, {"question": "Is there a difference in the success rate of CS graduate students compared to other evaluators when performing tasks that require domain-specific knowledge?", "answer": "Authors acknowledge that there may be a difference in human performance based on the demographics of the annotators. Many tasks in authors' dataset require domain-specific knowledge, which may not be easily performed by an average user without significant domain expertise. Authors aim to design tasks with easy-to-imagine outcomes rather than those that are easily performed by an average user. Authors will add this discussion to Section 3.2 or the Appendix if space permits. Additionally, authors have implemented a process to resolve potential ambiguities and ensure annotation correctness, as described in Section 3.2. Authors would appreciate concrete examples of tasks that were difficult to perform, if available.", "category": "Experimental_Exposition"}, {"question": "What steps have been taken to address the reproducibility concerns raised by the use of closed-source LLMs in the evaluation, and what are the task success rates of open-source models on WebArena?", "answer": "Follow-up works by other researchers have evaluated open-source models on WebArena after its public release. The task success rates for these open-source models are as follows: CodeLLama-34B-Instruct (4.06%), LLama-2-Chat-7B (1.23%), LLama-2-Chat-34B (1.11%), and LLama-2-Chat-70B (1.72%). These results will be cited in the camera-ready version of the paper.\n| Model                  | Task Success Rate (%) |\n|------------------------|-----------------------|\n| CodeLLama-34B-Instruct | 4.06                  |\n| LLama-2-Chat-7B        | 1.23                  |\n| LLama-2-Chat-34B       | 1.11                  |\n| LLama-2-Chat-70B       | 1.72                  |\n", "category": "Experimental_Exposition"}, {"question": "Why was the POMDP model chosen initially, and why was it later removed from the paper?", "answer": "The POMDP model was initially chosen because it is a standard way of modeling such problems in related works, such as WebShop [2]. However, the authors found the formulation confusing and decided to remove it. Instead, they rephrased the task-solving process in a simpler way, as described in the updated Section 2.1. 'Given a task described as a natural language intent $\\mathbf{i}$, an agent issues an action $a_t \\in \\mathcal{A}$ based on intent $\\mathbf{i}$, the current observation $o_t \\in \\mathcal{O}$, the action history and the observation history $a_{1}^{t-1}, o_{1}^{t-1}$.' \n[2] WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents, Yao el at, 2022", "category": "Motivation_Analysis"}, {"question": "Is it possible to add new tasks to the benchmark without modifying the existing environment, and if so, what is the process?", "answer": "Yes, new tasks can be easily added by following the process described in Sections 3.1 and 3.2. No modifications to the existing environment are required. For example, adding an information-seeking task involves providing the question and the corresponding annotated answer.", "category": "Method_Mechanics"}, {"question": "How does the difficulty of the tasks in WebArena compare to other benchmarks, such as MiniWob++ and Mind2Web?", "answer": "Tasks in WebArena are more challenging than those in simplified synthetic environments like MiniWob++, where GPT-4 achieves a 95% success rate compared to ~15% in WebArena. Additionally, WebArena's task difficulty is on par or more challenging than Mind2Web, as it requires end-to-end task success rather than just verifying correct action strings.", "category": "Method_Comparison"}, {"question": "What are the potential issues with using GPT-4 for evaluation in the proposed framework, and how do the authors address concerns about its commercial nature and accuracy?", "answer": "The authors acknowledge that using GPT-4 for evaluation could raise concerns about its commercial nature and accuracy. They address these concerns by noting that only ~10% (82 out of 812) of the examples require GPT-4 for evaluation, with a cost of less than 1 US dollar. They also manually checked 40 examples and found that 39 matched human judgment. Additionally, 60% of the 82 examples requiring GPT-4 involve evaluating dates or time durations, where GPT-4 is used only to judge answer formats. The authors quantitatively evaluated GPT-4's accuracy in these cases, finding 100% accuracy for both date and time duration formats across 900 examples for GPT-4 and its turbo version. \n|                         | gpt-4-0613 (same as in the paper) | gpt-4-1106-preview (GPT-4 Turbo) |\n|-------------------------|-----------------------------------|----------------------------------|\n| Date (900 examples)     | 100%                              | 100%                             |\n| Time duration (900 examples) | 100%                              | 100%                             |", "category": "Method_Mechanics"}, {"question": "Why is the human success rate only 78% on the designed tasks, and what types of errors contribute to these failures?", "answer": "The 50% of failures are due to errors such as misinterpreting the intent (e.g., providing travel distance instead of time), incomplete answers (e.g., providing only the customer's name instead of name and email), and incomplete executions (e.g., partially filling product information). The other 50% are more severe failures where the human performed the task incorrectly.", "category": "Experimental_Analysis"}, {"question": "What analysis was conducted to evaluate the performance of GPT-4 in the context of the study?", "answer": "To evaluate GPT-4's performance, the authors manually checked 40 examples, finding 39 identical to human judgment. For 82 examples requiring GPT-4, 49 (60%) involved date (e.g., 10/23/2022) or time duration formats (e.g., 15 minutes). GPT-4's accuracy was quantitatively evaluated by generating different date and time formats, achieving 100% accuracy in both date and time duration (900 examples each) for both GPT-4 versions tested. For instance, `Nov 3, 2022`, `November 3, 2022`, `3rd November 2022`, `3 Nov 2022`, `2022-11-03`, and `3rd of November, 2022` are all correct variances of `2022/11/03`. \n\n|                         | gpt-4-0613 (same as in the paper) | gpt-4-1106-preview (GPT-4 Turbo) |\n|-------------------------|-----------------------------------|----------------------------------|\n| Date (900 examples)     | 100%                              | 100%                             |\n| Time duration (900 examples) | 100%                              | 100%                             |\n", "category": "Experimental_Exposition"}, {"question": "What are the reasons for the 78% success rate of human annotators, as described in Section 3.2 of the paper?", "answer": "The 78% success rate is attributed to two main categories of errors: (1) 50% of failures are due to misinterpretation of intent (e.g., providing travel distance instead of travel time), incomplete answers (e.g., providing only the customer's name when both name and email were requested), and incomplete executions (e.g., partially filling product information). (2) The remaining 50% of failures are more severe, where the human annotators performed the task incorrectly.", "category": "Experimental_Analysis"}, {"question": "Is GPT-4's accuracy in evaluating date and time duration formats less than 100% according to the manual checks and quantitative evaluations performed?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the proposed framework's reliance on GPT-4 for evaluation significantly limit its potential use due to commercial constraints?", "answer": "False", "category": "Claim_Verification"}, {"question": "Have open-source models been evaluated on WebArena by follow-up works after the public release?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the evaluation results using GPT-4 considered unreliable due to the tool's potential inaccuracies?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the success rate of CS graduate students compared to other evaluators explicitly measured and reported in the study?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper include direct citations for the follow-up works evaluating open-source models on WebArena?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "RXFVcynVe1", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning", "authors": ["Xiaoxin He", "Xavier Bresson", "Thomas Laurent", "Adam Perold", "Yann LeCun", "Bryan Hooi"], "abstract": "Representation learning on text-attributed graphs (TAGs) has become a critical research problem in recent years. A typical example of a TAG is a paper citation graph, where the text of each paper serves as node attributes. Initial graph neural network (GNN) pipelines handled these text attributes by transforming them into shallow or hand-crafted features, such as skip-gram or bag-of-words features. Recent efforts have focused on enhancing these pipelines with language models (LMs), which typically demand intricate designs and substantial computational resources. With the advent of powerful large language models (LLMs) such as GPT or Llama2, which demonstrate an ability to reason and to utilize general knowledge, there is a growing need for techniques which combine the textual modelling abilities of LLMs with the structural learning capabilities of GNNs. Hence, in this work, we focus on leveraging LLMs to capture textual information as features, which can be used to boost GNN performance on downstream tasks. A key innovation is our use of \\emph{explanations as features}: we prompt an LLM to perform zero-shot classification, request textual explanations for its decision-making process, and design an \\emph{LLM-to-LM interpreter} to translate these explanations into informative features for downstream GNNs. Our experiments demonstrate that our method achieves state-of-the-art results on well-established TAG datasets, including \\texttt{Cora}, \\texttt{PubMed}, \\texttt{ogbn-arxiv}, as well as our newly introduced dataset, \\texttt{tape-arxiv23}. Furthermore, our method significantly speeds up training, achieving a 2.88 times improvement over the closest baseline on \\texttt{ogbn-arxiv}. Lastly, we believe the versatility of the proposed method extends beyond TAGs and holds the potential to enhance other tasks involving graph-text data~\\footnote{Our codes and datasets are available at: \\url{https://github.com/XiaoxinHe/TAPE}}.", "keywords": "large language models (LLM);feature learning;text attributed graphs (TAG);graph neural networks (GNN)", "qa_pairs": [{"question": "How does the memory cost of the proposed method, which involves fully fine-tuning Deberta, compare to other LM-based methods rather than pure GNN methods, and what justifies the increased memory usage?", "answer": "The memory cost of the proposed method, TAPE, should be compared to other LM-based methods like GIANT, GLEM, GraphFormers, and Patton, rather than pure GNN methods, as learning on TAG inherently requires the integration of LMs and GNNs. Despite the increased memory usage, TAPE significantly improves accuracy from 70.83% to 77.50%, justifying the higher resource requirements. Additionally, these demands are manageable with cost-effective academic GPU resources like the A5000 GPU (24GB). A detailed analysis in Table 14 shows that TAPE offers the best performance and comparable or better running time than other LM-based methods.", "category": "Method_Comparison"}, {"question": "What additional evaluation tasks, such as link prediction, have been included to assess the quality of the representations learned by TAPE?", "answer": "The authors have extended their experimental evaluation to include link prediction tasks. They fine-tuned the language model (LM) for link prediction, using the resulting original text encoding and explanation encoding as node features to train a downstream graph neural network (GNN) specifically for link prediction. The results, detailed in Table 16, demonstrate substantial improvements achieved by TAPE compared to GraphFormers, Patton, and GNN with shallow feature methods (without LLM-based features). These results highlight the versatility and applicability of the approach beyond node classification.", "category": "Experimental_Setup"}, {"question": "What is the impact of fine-tuning the language model (LM) in step 2 on performance, and why are different LMs used for original text encoding and explanation encoding?", "answer": "Fine-tuning the LM in step 2 significantly improves performance, as shown in Table 12 where 'without fine-tuning' exhibits inferior results. Regarding the use of different LMs, Table 12 indicates that training the same or distinct LMs for encoding original text and explanations yields similar results, with only a slight advantage for using two LMs, suggesting that one LM can be used to simplify the approach.", "category": "Motivation_Analysis"}, {"question": "What is the detailed time and cost analysis for using the ChatGPT API with the primary dataset ogbn-arxiv, and what cost-effective alternatives have been explored?", "answer": "The detailed time and cost analysis for using the ChatGPT API with the primary dataset ogbn-arxiv is as follows: The dataset has 169,343 nodes and 1,166,243 edges, with average input sequences of 285 tokens and output sequences of 164 tokens. For the ChatGPT-3.5 Turbo API, priced at $0.0015 per 1,000 input tokens and $0.002 per 1,000 output tokens, with a token per minute rate limit of 90,000, the monetary estimation for ogbn-arxiv is as follows: Cost = ((285 * 0.0015) / 1000 + (164 * 0.002) / 1000) * 169,343 ≈ 128 USD. Time = 169,343 / (90,000 / 285) ≈ 536 min ≈ 9h. Additionally, cost-effective alternatives like using open-source LLMs, such as llama2, have been explored, which is free and takes approximately 16 hours with 4 A5000 GPUs. The method also enhances efficiency by requiring only one query to the LLM, with stored predictions and explanations for subsequent use.", "category": "Experimental_Setup"}, {"question": "Does the paper provide evidence that different prompt designs can achieve similar performance in node classification, and if so, what is the range of accuracy observed?", "answer": "The paper provides evidence in Table 8 and Table 9 (in blue) in the appendix, showing that different prompt designs maintain similar accuracy, ranging from 0.7660 to 0.7750, indicating the robustness of the prompt. The higher the zero-shot prediction score, the better the accuracy of TAPE. ", "category": "Experimental_Exposition"}, {"question": "Why were only BERT-based models considered in the ablation study, and how does the performance of SentenceBert Embedding in recent studies compare to au'thors findings?", "answer": "Authors conducted additional experiments with various LMs, including all-roberta-large-v1 and e5-large, to assess their impact on TAPE. authors' findings, presented in Table 13, show that the model is insensitive to the specific LM choice, demonstrating robustness. While recent studies like SimTeG achieve satisfactory results with SentenceBert, their approach also involves fine-tuning, similar to authors, and integrating SimTeG with TAPE establishes new SOTA performance on ogbn-arxiv.", "category": "Method_Comparison"}, {"question": "Why were experimental results for the GIANT model initially missing in Table 1, and how has this been addressed?", "answer": "The initial comparison with GIANT was limited to the ogbn-arxiv dataset because the original GIANT paper did not include experiments on other datasets like Cora/PubMed. In response to the feedback, the authors extended the comparison by applying the GIANT model to additional datasets using its official code. The updated results are now included in Table 1 (highlighted in blue), demonstrating consistent outperformance of the proposed method against GIANT across all datasets.", "category": "Experimental_Setup"}, {"question": "How does the proposed method address the challenges of joint training of GNNs and LMs for TAG, and what is its contribution compared to existing paradigms?", "answer": "The proposed method addresses the challenges of joint training of GNNs and LMs for TAG by employing a two-stage training paradigm, which is more efficient and easier to integrate into existing GNN pipelines compared to end-to-end or joint training. It incorporates an LLM-to-LM interpreter to extract and interpret useful prior knowledge from LLMs, embedding this knowledge into GNN features. This approach achieves equal or better performance than joint training while significantly improving efficiency. The contribution of the method lies in demonstrating that the two-stage training paradigm, enhanced by the LLM-to-LM interpreter, can overcome the training instabilities and resource demands associated with joint training, offering a novel, efficient, and robust solution for tasks involving both text and graphs.", "category": "Method_Mechanics"}, {"question": "Why were GraphFormers [1] and Patton [2] not included as baselines in the experimental section, and how does the proposed method compare to these baselines on relevant datasets?", "answer": "GraphFormers and Patton were not included as baselines because they were either designed for different tasks (e.g., GraphFormers for link prediction) or involved datasets and pre-training methods that do not overlap with the datasets used in this study (e.g., Patton). The proposed method, TAPE, was compared against GLEM and GIANT, which are recognized as the strongest baselines for the datasets ogbn-arxiv and ogbn-products. Specifically addressing GraphFormers, it was primarily designed for link prediction tasks, making a direct comparison less suitable due to differences in task objectives. Similarly, Patton involves pre-training on Microsoft Academic Graph (MAG) or Amazon e-commerce networks, followed by fine-tuning in a few-shot manner, making it also challenging for a direct comparison. Moreover, it is important to note that the datasets used in the GraphFormers and Patton papers do not overlap with the datasets in our study. Consequently, authors are unable to report from their papers the performance on our datasets. additional experiments were conducted on the MAG-Economics dataset, a 40-way node classification task, to facilitate a comparison with Patton, where TAPE outperformed GraphFormers and Patton by at least +13% accuracy. \n[1] J. Yang, et al. GraphFormers: GNN-nested Transformers for Representation Learning on Textual Graph. NeurIPS 2021.\n\n[2] B. Jin, et al. Patton: Language Model Pretraining on Text-Rich Networks. ACL 2023.", "category": "Method_Comparison"}, {"question": "How does the proposed approach using an LLM-to-LM interpreter for two-stage training contribute to the field of integrating LLMs with GNNs for text- and graph-based tasks, and what sets it apart from existing end-to-end or joint training methods?", "answer": "The proposed approach contributes by demonstrating that two-stage training with an LLM-to-LM interpreter can achieve equal or better performance than end-to-end or joint training, with greater efficiency. It allows embedding valuable prior knowledge from LLMs into GNN features without sacrificing performance. This method stabilizes the training process by decomposing the text+graph combination into independently executed parts, similar to the training of LLMs like ChatGPT or Llama2, including self-supervised learning, fine-tuning, and reinforcement learning. This departure from optimizing in a one-shot end-to-end manner, known for its instability, has proven highly successful in NLP. It addresses scalability issues of end-to-end training with LLMs, making it easier to leverage LLMs for graph learning. The approach has been validated by subsequent similar techniques achieving top rankings on the OGB leaderboard, highlighting its pioneering significance.", "category": "Method_Comparison"}, {"question": "Were the experimental results for GIANT initially missing in Table 1 for datasets other than ogbn-arxiv?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the comprehensive results in Table 1 show that the authors' method consistently outperforms GIANT on all tested datasets?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "AOJyfhWYHf", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "OpenChat: Advancing Open-source Language Models with Mixed-Quality Data", "authors": ["Guan Wang", "Sijie Cheng", "Xianyuan Zhan", "Xiangang Li", "Sen Song", "Yang Liu"], "abstract": "Nowadays, open-source large language models like LLaMA have emerged. Recent developments have incorporated supervised fine-tuning (SFT) and reinforcement learning fine-tuning (RLFT) to align these models with human goals. However, SFT methods treat all training data with mixed quality equally, while RLFT methods require high-quality pairwise or ranking-based preference data. In this study, we present a novel framework, named OpenChat, to advance open-source language models with mixed-quality data. Specifically, we consider the general SFT training data, consisting of a small amount of expert data mixed with a large proportion of sub-optimal data, without any preference labels. We propose the C(onditioned)-RLFT, which regards different data sources as coarse-grained reward labels and learns a class-conditioned policy to leverage complementary data quality information. Interestingly, the optimal policy in C-RLFT can be easily solved through single-stage, RL-free supervised learning, which is lightweight and avoids costly human preference labeling.\nThrough extensive experiments on three standard benchmarks, our openchat-13b fine-tuned with C-RLFT achieves the highest average performance among all 13b open-source language models. Moreover, we use AGIEval to validate the model generalization performance, in which only openchat-13b surpasses the base model. Finally, we conduct a series of analyses to shed light on the effectiveness and robustness of OpenChat. Our code, data, and models are publicly available at https://github.com/imoneoi/openchat and https://huggingface.co/openchat.", "keywords": "Open-source Language Models;Fine-tuning;Mixed-quality Data", "qa_pairs": [{"question": "What is the role of 'lower quality data' in the C-RLFT method, and why is it important for the proposed approach?", "answer": "The 'lower quality data' is important in the C-RLFT method as it provides contrast signals to guide policy learning. Unlike supervised fine-tuning (SFT) on GPT-4 data alone, which does not guarantee the best performance, C-RLFT uses the quality difference between datasets as an implicit reward signal. This enables the model to discriminate between good and bad samples during learning, leading to improved performance. Without the lower-quality data, these contrast learning signals would not exist, and the method's effectiveness would be compromised.", "category": "Method_Disambiguation"}, {"question": "How does the sensitivity of the results to the division of 'good' data and 'average' data in C-RLFT change if both sets of data are of the same quality?", "answer": "If the 'good' data and 'average' data in C-RLFT are of the same quality, it reduces to performing SFT on a single source of data. Additional experiments show that SFT tuned with only GPT-4 (~6k) performs better than SFT with both GPT-4 and GPT-3.5 (~70k), confirming that indiscriminately feeding mixed conversations to the base model negatively impacts learning. The performance of C-RLFT remains higher than SFT with GPT-4, demonstrating its effectiveness.\n| Type                     | Average  | AlpacaEval | MT-bench | Vicuna-bench |\n|--------------------------|----------|------------|----------|--------------|\n| C-RLFT (GPT-4 + GPT-3.5) | **77.3** | **89.5**   | **57.5** | **85.0**     |\n| SFT (GPT-4)              | 67.9     | 85.8       | 33.4     | 84.4         |\n| SFT (GPT-4 + GPT-3.5)    | 52.7     | 78.6       | 33.1     | 46.3         |\n| SFT (GPT-3.5)            | 42.8     | 76.5       | 16.9     | 35.0         |", "category": "Experimental_Exposition"}, {"question": "How does the training process of C-RLFT achieve stability compared to traditional Reinforcement Learning methods?", "answer": "C-RLFT achieves stability by transforming the policy optimization into a simple weighted regression objective, which allows solving an RL problem using a single-stage supervised learning scheme. This approach eliminates the need for a reward model, value networks, and the maintenance of a reference policy for policy constraints, making the process extremely lightweight and stable.", "category": "Method_Mechanics"}, {"question": "Does the requirement to divide data into good quality and average quality limit the applications of the proposed method?", "answer": "No, the proposed method has extremely relaxed requirements on datasets and is very friendly for applications. It only requires a small set of expert data and a larger set of medium-quality data, without the need for detailed quality evaluation for each sample nor costly human preference labels. Authors use the widely adopted ShareGPT dataset (~70k conversations), with only 6k expert conversations generated by GPT-4. This finetuning dataset essentially is the same one used to finetune `vicuna-v1.1-13b`, but as shown in paper, the paper' `openchat-13b` achieved significantly better performance. Authors have also summarized the finetuning data sources of OpenChat and other popular open-source language models. It should be noted many existing models have more stringent requirements on the finetuning data, either data quantity or data quality. By contrast OpenChat enables to use the minimal amount of potentially expensive expert data as well as more easily obtainable medium-quality data to achieve improved LLM finetuning performance. The C-RLFT framework provides a flexible and performant approach to address a wide range of applications, including downstream tasks like safe vs not-so-safe data or preferred role vs. less preferred role in role-playing tasks.\n| Model                     | Base Model    | Finetuning | Data                                             |\n| ------------------------- | ------------- | ---------- | ------------------------------------------------ |\n| `vicuna-v1.1-13b`         | `llama-13b`   | SFT        | ~70k ShareGPT (GPT-4 + GPT-3.5)                  |\n| `wizardlm-v1.0-13b`       | `llama-13b`   | SFT        | ~70k GPT-3.5                                     |\n| `ultralm-13b`             | `llama-13b`   | SFT        | ~1.5M UltraChat (GPT-3.5)                        |\n| `llama-2-chat-13b`        | `llama-2-13b` | SFT + RLFT | ~27k high-quality SFT data + ~2.9M preference    |\n| `vicuna-v1.5-13b`         | `llama-2-13b` | SFT        | ~125k ShareGPT (GPT-4 + GPT-3.5)                 |\n| `wizardlm-v1.2-13b`       | `llama-2-13b` | SFT        | ~250k Evol-Instruct + ShareGPT (GPT-4 + GPT-3.5) |\n| **`openchat-13b (ours)`** | `llama-2-13b` | C-RLFT     | ~70k ShareGPT (GPT-4 + GPT-3.5)                  |", "category": "Experimental_Analysis"}, {"question": "How does the proposed C-RLFT method address the challenge of estimating data quality when leveraging mixed-quality data sources, and how does it avoid the need for explicit data quality estimation compared to typical RLFT methods?", "answer": "The proposed C-RLFT method bypasses the explicit estimation of data quality by leveraging the quality difference between an expert dataset ($\\mathcal{D}_{exp}$) and a sub-optimal dataset ($\\mathcal{D}_{sub}$). Unlike typical RLFT methods, which require costly human preference data to learn reward models and explicitly estimate data quality, C-RLFT avoids this two-stage process, reducing instability and cost. The method is not limited to GPT-4 and GPT-3.5 datasets; it can use any expert and sub-optimal datasets. Additional experiments with human-written data (expert) and GPT-4 data (sub-optimal) from the OpenOrca dataset demonstrate that C-RLFT outperforms SFT models trained with human or GPT-4 data, or a combination of both.\n | Type | BBH | \n| ---------------------- | -------- |\n | C-RLFT (Human + GPT-4) | **49.4** |\n | SFT (Human + GPT-4) | 47.9 |\n | SFT (Human) | 47.1 | \n| SFT (GPT-4) | 46.3 |", "category": "Method_Mechanics"}, {"question": "How does the quality of data from different versions of language models, such as GPT-4 and GPT-3.5, impact the performance of fine-tuned models, and what is the optimal quantity ratio of high-quality to lower-quality data required to achieve equivalent performance enhancements?", "answer": "The performance of fine-tuned models is impacted by the quality of data from different versions of language models. C-RLFT (Contrastive Reinforcement Fine-Tuning) achieves better performance compared to SFT (Supervised Fine-Tuning) when trained on datasets with different quality levels, such as GPT-4 and GPT-3.5. The optimal quantity ratio of high-quality to low-quality data is robust to variation, as demonstrated by ablation results in Figure 7. The average performance remains stable when sub-sampling one class (e.g., GPT-3.5 or GPT-4) with ratios varying from 60% to 100%, as long as the expert data size is not overly small. For more details, refer to Section 5.5 in the paper.\n\n | Type | Average | AlpacaEval | MT-bench | Vicuna-bench |\n |--------------------------|----------|------------|----------|--------------| \n| C-RLFT (GPT-4 + GPT-3.5) | **77.3** | **89.5** | **57.5** | **85.0** |\n| SFT (GPT-4) | 67.9 | 85.8 | 33.4 | 84.4 | \n| SFT (GPT-4 + GPT-3.5) | 52.7 | 78.6 | 33.1 | 46.3 |\n | SFT (GPT-3.5) | 42.8 | 76.5 | 16.9 | 35.0 |", "category": "Experimental_Exposition"}, {"question": "What strategies can be implemented within the proposed C-RLFT framework to mitigate the transfer of potentially negative behaviors from GPT-4 examples to the base model, and how can such behaviors be identified and counteracted effectively?", "answer": "The proposed C-RLFT framework is not limited to GPT-4 and GPT-3.5 data. If there are concerns about negative behaviors in GPT-4 examples, the expert dataset $\\mathcal{D}_{exp}$ can be replaced with a small set of carefully curated human expert data, as demonstrated in LIMA [4].\n[4] LIMA: less is more for alignment. arXiv 2023.", "category": "Method_Mechanics"}, {"question": "How does the performance of the C-RLFT method, which is a KL-regularized RL problem rather than a typical supervised learning approach, compare to the SFT+RLFT method, and how does the simplicity of the weighted supervised learning loss reconcile with the advantages of an RL-based approach?", "answer": "C-RLFT is not a typical supervised learning approach but a KL-regularized RL problem: $J_{C-RLFT}(\\theta)=\\mathbb{E}_{y \\sim \\pi_\\theta}[r_c(x,y)]-\\beta D_{\\mathrm{KL}}\\left(\\pi_\\theta, \\pi_c\\right)$ and its optimal policy can be extracted using a weighted supervised learning loss. The `openchat-13b` model, which uses C-RLFT, outperforms the `llama-2-chat-13b` baseline, which uses SFT+RLFT, with an average score of 77.3 compared to 74.4. This demonstrates the strength of the C-RLFT approach despite its simplicity in using a weighted supervised learning loss.\n\n| Models | Base Models | Method | Dataset | AlpacaEval | MT-bench | Vicuna-bench | Average | \n| ------------------ | ------------- | ---------- | --------------------------------------------- | ---------- | -------- | ------------ | -------- | \n| `llama-2-chat-13b` | `llama-2-13b` | SFT + RLFT | ~27k high-quality SFT data + ~2.9M preference | 81.1 | 55.3 | **86.9** | 74.4 |\n | `openchat-13b` | `llama-2-13b` | C-RLFT | ~70k ShareGPT | **89.5** | **57.5** | 85.0 | **77.3** |", "category": "Method_Comparison"}, {"question": "Can the proposed method (C-RLFT) effectively handle mixed-quality data generated by open-sourced models like LLaMA-13b and LLaMA-7b, and how does it perform on datasets other than ShareGPT?", "answer": "Training the base model (`llama-2-13b`) on datasets generated by weaker models like `LLaMA-7b` and `LLaMA-13b` is not appropriate, as these models have inferior capabilities and are unlikely to provide useful information for fine-tuning. However, the performance of C-RLFT on datasets other than ShareGPT has been evaluated using the OpenOrca dataset, which includes human-written and GPT-4-generated data. C-RLFT outperforms SFT on both the human and GPT-4 datasets, as well as their combination, as shown in the provided table.\n| Type                   | BBH      |\n| ---------------------- | -------- |\n| C-RLFT (Human + GPT-4) | **49.4** |\n| SFT (Human + GPT-4)    | 47.9     |\n| SFT (Human)            | 47.1     |\n| SFT (GPT-4)            | 46.3     |", "category": "Experimental_Analysis"}, {"question": "How does the C-RLFT method differ from the methodology used by Xu et al., particularly in terms of training dataset requirements, RL objective formulation, and potential biases?", "answer": "The C-RLFT method differs from Xu et al.'s methodology in several key aspects: 1) C-RLFT uses two different quality datasets (a small set of expert data and a larger set of medium-quality data) without needing preference labels, whereas Xu et al. require pairwise preference labels for every question. 2) C-RLFT reformulates the RL objective to a weighted supervised loss, offering more stability compared to the RL training in existing RLFT methods. Xu et al., like all RLHF methods, require pairwise preference labels for every question $x_i \\mapsto (y_i^+, y_i^-)$, i.e., for every question $x_i$, it needs a preferred answer $y_i^+$ that is better than the worse answer $y_i^-$. 3) Xu et al.'s method involves using two different quality LLMs to generate positive and negative answers, followed by DPO for policy fine-tuning, which is heavier and inherits instability and hyperparameter sensitivity. Additionally, the positive and negative answers generated by LLMs in Xu et al.'s method can introduce model-specific biases to the preference data.", "category": "Method_Comparison"}, {"question": "How does the performance of C-RLFT compare to SFT when distilling from GPT models, and why can't the superior performance of C-RLFT be solely attributed to distillation from better GPT models?", "answer": "C-RLFT outperforms SFT when distilling from GPT models, as shown by experiments comparing C-RLFT and SFT on different GPT data. SFT performances are significantly worse than C-RLFT, even when using both GPT-4 and GPT-3.5 data. This indicates that indiscriminate training with mixed quality data negatively impacts learning, whereas C-RLFT leverages the quality contrast of two datasets for better performance, which cannot be explained by distillation from better GPT models alone.\nTable R1: Comparative performance of C-RLFT and SFT tuned on different data sources.\n\n| Type                     | Average  | AlpacaEval | MT-bench | Vicuna-bench |\n|--------------------------|----------|------------|----------|--------------|\n| C-RLFT (GPT-4 + GPT-3.5) | **77.3** | **89.5**   | **57.5** | **85.0**     |\n| SFT (GPT-4)              | 67.9     | 85.8       | 33.4     | 84.4         |\n| SFT (GPT-4 + GPT-3.5)    | 52.7     | 78.6       | 33.1     | 46.3         |\n| SFT (GPT-3.5)            | 42.8     | 76.5       | 16.9     | 35.0         |\n\nTable R2: Comparing openchat with our C-RLFT to a series of fine-tuned LLMs with SFT + RLFT.\n| Models              | Base Models   | Method     | Dataset                                       | AlpacaEval | MT-bench | Vicuna-bench | Average  |\n| ------------------- | ------------- | ---------- | --------------------------------------------- | ---------- | -------- | ------------ | -------- |\n| **Larger than 13b** |               |            |                                               |            |          |              |          |\n| `gpt-4`             | -             | SFT + RLFT | -                                             | 95.3       | 82.5     | 90.0         | 89.3     |\n| `llama-2-chat-70b`  | `llama-2-70b` | SFT + RLFT | -                                             | 92.7       | 60.0     | 87.5         | 80.1     |\n| `claude`            | -             | SFT + RLFT | -                                             | 88.4       | 65.0     | 76.3         | 76.6     |\n| `gpt-3.5-turbo`     | -             | SFT + RLFT | -                                             | 86.1       | 50.0     | 50.0         | 62.0     |\n| **Equal to 13b**    |               |            |                                               |            |          |              |          |\n| `llama-2-chat-13b`  | `llama-2-13b` | SFT + RLFT | ~27k high-quality SFT data + ~2.9M preference | 81.1       | 55.3     | **86.9**     | 74.4     |\n| `openchat-13b`      | `llama-2-13b` | C-RLFT     | ~70k ShareGPT                                 | **89.5**   | **57.5** | 85.0         | **77.3** |", "category": "Method_Comparison"}, {"question": "How can the proposed C-RLFT framework be adapted to more complex and realistic data environments, such as those involving human-generated data with significant variations in quality and subjective quality assessments?", "answer": "Proposed C-RLFT is not designed to 'estimate data quality' as in typical RLFT methods, but rather provide the cheapest solution to bypass this step in LLM fine-tuning. In typical RLFT methods, authors have to use lots of expensive human preference data to learn reward models, in order to explicitly estimate data quality. This poses major obstacles and instability in LLM fine-tuning, as human preference collection is very costly and the two-stage training process (learn reward first and then run RL algorithm like PPO) can be very unstable. The proposed C-RLFT framework avoids explicit data quality estimation by leveraging the quality difference between an expert dataset ($\\mathcal{D}_{exp}$) and a sub-optimal dataset ($\\mathcal{D}_{sub}$). It is not limited to GPT-4 or GPT-3.5 data and can use any expert and sub-optimal datasets. To demonstrate adaptability, additional experiments were conducted using human-written data (expert) and GPT-4 data (sub-optimal) from the OpenOrca dataset. The results show that C-RLFT outperforms SFT models trained with human or GPT-4 data, or a combination of both, achieving a BBH accuracy of 49.4 compared to 47.9, 47.1, and 46.3 for SFT models.\n| Type | BBH | \n| ---------------------- | -------- | \n| C-RLFT (Human + GPT-4) | **49.4** |\n | SFT (Human + GPT-4) | 47.9 |\n | SFT (Human) | 47.1 | | SFT (GPT-4) | 46.3 |", "category": "Experimental_Analysis"}, {"question": "How does the use of data generated by API models (GPT-4 & GPT-3.5) in the C-RLFT method differ from prior approaches (SFT + RLHF), and how does this impact the fairness of comparisons with baselines that have not used such data?", "answer": "Most open-source SFT models, except for `llama-2-chat-13b`, also leverage training data generated by larger API models. OpenChat uses a minimal amount of expensive expert data and more easily obtainable medium-quality data, which allows for improved LLM finetuning performance despite using similar data sources.\n| Model                     | Base Model    | Finetuning | Data                                             |\n| ------------------------- | ------------- | ---------- | ------------------------------------------------ |\n| **Equal to 13b**          |               |            |                                                  |\n| `vicuna-v1.1-13b`         | `llama-13b`   | SFT        | ~70k ShareGPT (GPT-4 + GPT-3.5)                  |\n| `wizardlm-v1.0-13b`       | `llama-13b`   | SFT        | ~70k GPT-3.5                                     |\n| `ultralm-13b`             | `llama-13b`   | SFT        | ~1.5M UltraChat (GPT-3.5)                        |\n| `llama-2-chat-13b`        | `llama-2-13b` | SFT + RLFT | ~27k high-quality SFT data + ~2.9M preference    |\n| `vicuna-v1.5-13b`         | `llama-2-13b` | SFT        | ~125k ShareGPT (GPT-4 + GPT-3.5)                 |\n| `wizardlm-v1.2-13b`       | `llama-2-13b` | SFT        | ~250k Evol-Instruct + ShareGPT (GPT-4 + GPT-3.5) |\n| **`openchat-13b (ours)`** | `llama-2-13b` | C-RLFT     | ~70k ShareGPT (GPT-4 + GPT-3.5)", "category": "Method_Comparison"}, {"question": "Is the C-RLFT method exclusively dependent on GPT-4 and GPT-3.5 datasets for its functionality?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the method require detailed quality evaluation for each sample in the dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the ShareGPT dataset used for finetuning `vicuna-v1.1-13b` and `openchat-13b` with the same data quantity?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the proposed C-RLFT method explicitly estimate data quality as part of its process?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the authors conduct additional experiments using human-written data as expert data and GPT-4 data as sub-optimal data to validate C-RLFT's performance?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the C-RLFT framework allow for flexible dataset division based on downstream tasks such as safe vs not-so-safe data?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the method limited to datasets that can be strictly divided into good quality and average quality?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "IkmD3fKBPQ", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Large Language Models Cannot Self-Correct Reasoning Yet", "authors": ["Jie Huang", "Xinyun Chen", "Swaroop Mishra", "Steven Zheng", "Adams Wei Yu", "Xinying Song", "Denny Zhou"], "abstract": "Large Language Models (LLMs) have emerged as a groundbreaking technology with their unparalleled text generation capabilities across various applications. Nevertheless, concerns persist regarding the accuracy and appropriateness of their generated content. A contemporary methodology, self-correction, has been proposed as a remedy to these issues. Building upon this premise, this paper critically examines the role and efficacy of self-correction within LLMs, shedding light on its true potential and limitations. Central to our investigation is the notion of intrinsic self-correction, whereby an LLM attempts to correct its initial responses based solely on its inherent capabilities, without the crutch of external feedback. In the context of reasoning, our research indicates that LLMs struggle to self-correct their responses without external feedback, and at times, their performance even degrades after self-correction. Drawing from these insights, we offer suggestions for future research and practical applications in this field.", "keywords": "Large Language Models;Self-Correction;Reasoning", "qa_pairs": [{"question": "Why did the study use a limited set of 200 questions for GPT-4 testing instead of a more extensive dataset, and how does this choice align with previous research and cost considerations?", "answer": "The study used a limited set of 200 questions for GPT-4 testing for two primary reasons: (1) To align with previous studies by Kim et al., 2023; Shinn et al., 2023; and Du et al., 2023, which focused on GPT-3.5, ensuring accurate reproduction of their results on established benchmarks. (2) Due to substantial cost considerations, as a single full test on GSM8K with GPT-4 costs approximately $200, leading to total costs in the thousands. To manage the costs, authors opt to randomly sample 200 examples for authors' tests on GPT-4. The study also considered running tests on the full test set when costs were reasonable. For instance, Du et al. (2023) tested only 100 examples on GSM8K with GPT-3.5, whereas authors chose to run the full test set to make the results more reliable.", "category": "Experimental_Setup"}, {"question": "How does the guessing approach described in section '3.1.3 REFLECTION' work, and can it be applied to non-multiple-choice datasets like GSM8K?", "answer": "Using oracle labels (a setting used in previous work) do not necessarily reflect true self-correction ability. The guessing approach employs oracle labels in multiple-choice settings. For example, if the correct label is (C) and the model initially predicts (A), since (A) ≠ (C), the random baseline selects a new answer from the remaining options (B-E). If the sampled option is (D), and (D) ≠ (C), the random baseline samples again from the remaining options (B, C, E). If the next sampled option is (C), which equals (C), authors then obtain the correct answer and stop. In a 5-choice format like CommonSenseQA, this baseline achieves 100% accuracy within four rounds. For non-multiple-choice datasets like GSM8K, a baseline could generate random numbers, potentially reaching the correct answer through numerous attempts. However, such improvements are not meaningful, and the footnote discussing this has been removed to avoid confusion.", "category": "Method_Mechanics"}, {"question": "Why does section 3.1.3 present findings that differ from the evidence in Table 1, which seems to support self-reasoning?", "answer": "The results in Table 1 use oracle labels to guide self-correction, which is impractical and differs from the intrinsic self-correction setup. Therefore, they cannot support the claim that LLMs can self-correct reasoning intrinsically. The writing has been adjusted to enhance clarity.", "category": "Experimental_Analysis"}, {"question": "What are the specific financial costs associated with conducting a single full test on GSM8K with self-correction using GPT-4, and how does this information enhance the paper's analysis of scalability and resource requirements?", "answer": "Conducting a single full test on GSM8K with self-correction using GPT-4 incurs a cost of approximately $200, with the total cost for all tests amounting to thousands of dollars. This information has been added to the paper to provide concrete financial details and to illustrate the resource requirements for replicating the results, thereby enhancing the analysis of scalability.", "category": "Experimental_Setup"}, {"question": "Did the study rely solely on the prompt 'Review your previous answer and find problems with your answer', and if not, what other prompts were tested?", "answer": "The study did not rely solely on the prompt 'Review your previous answer and find problems with your answer'. Multiple prompts were tested, including three different sets from prior studies (Kim et al., 2023, Shinn et al., 2023, and Du et al., 2023), unbiased self-correction prompts like 'Verify whether your answer is correct, and provide an explanation', and additional tests on GPT-4-Turbo and Llama-2-70b: 'Assume that this answer could be either correct or incorrect. Review the answer carefully and report any serious problems you find.' The results consistently showed decreased model performance after self-correction.", "category": "Method_Mechanics"}, {"question": "Why is it intuitive or expected that a well-aligned model with a thoughtfully designed initial prompt should produce an optimal response, and how does this reasoning differ from the intuition behind the effectiveness of chain-of-thought prompting?", "answer": "The complete sentence is, 'If the model is well-aligned and paired with a thoughtfully designed initial prompt, the initial response should already be optimal relative to the prompt and the specific decoding algorithm.' The critical missing part, 'the response is optimal relative to the given prompt and the specific decoding algorithm', significantly alters the meaning when omitted. The response is optimal relative to the given prompt and the specific decoding algorithm. Chain-of-Thought Prompting provides better guidance for the model to solve the problem, making the model's response optimal to the improved prompt (with CoT), which surpasses the response generated without CoT. This contrasts with intrinsic self-correction, where the feedback prompt (without oracle labels) does not provide helpful benefits for reasoning and may bias the model to generate a worse response.", "category": "Motivation_Analysis"}, {"question": "Why did the authors choose to focus on GPT-3.5 instead of GPT-4, and how did they address the request to include diagrams for GPT-4?", "answer": "The authors chose to focus on GPT-3.5 for two reasons: (1) Alignment with previous studies: The paper critiques studies by Kim et al., 2023; Shinn et al., 2023; and Du et al., 2023, which focus on GPT-3.5, necessitating the use of GPT-3.5 to accurately reproduce their results. (2) Cost considerations: GPT-4 is significantly more expensive, with a single full test on GSM8K costing approximately $200. To manage costs, the authors randomly sampled 200 examples for their tests. Additionally, the authors have included diagrams for GPT-4 in Figure 1 in response to the suggestion.", "category": "Motivation_Analysis"}, {"question": "What comparisons have been made with other baseline methods in the study?", "answer": "The study includes comparisons with multiple prompts and models, such as GPT-4-Turbo and Llama-2-70b, and prompts from prior studies like Kim et al., 2023; Shinn et al., 2023; and Du et al., 2023. Experiments with unbiased prompts were also conducted, showing consistent performance decrease after self-correction. ", "category": "Method_Comparison"}, {"question": "Do the results on the impact of self-correction on model performance generalize to other open-source large language models (LLMs) such as GPT-4-Turbo and Llama-2-70B?", "answer": "Yes, additional experiments with GPT-4-Turbo and Llama-2-70B show that these models' performance also decreases after self-correction. \n|                          |                      | # calls | GSM8K | CommonSenseQA |\n|--------------------------|----------------------|-------|-------|---------------|\n| gpt-4-1106-preview       | Standard Prompting   | 1     | 91.5  | 84.0          |\n|                          | Self-Correct (round 1) | 3   | 88.0  | 81.5          |\n|                          | Self-Correct (round 2) | 5   | 90.0  | 83.0          |\n| Llama-2-70b-chat-hf      | Standard Prompting   | 1     | 62.0  | 64.0          |\n|                          | Self-Correct (round 1) | 3   | 43.5  | 37.5          |\n|                          | Self-Correct (round 2) | 5   | 36.5  | 36.5          |", "category": "Experimental_Exposition"}, {"question": "Did the authors attempt to combine multi-agent debating with self-consistency, and what were the results?", "answer": "Yes, the authors combined multi-agent debating with self-consistency by taking the majority vote of the final responses of all three agents, following the setting of the original paper. They also experimented with taking the majority vote of all responses, including previous responses, but did not observe any performance improvement.", "category": "Experimental_Exposition"}, {"question": "What efforts were made to explore different variations of the self-correct prompt 'Review your previous answer and find problems with your answer', and how do these variations affect the models' performance on reasoning?", "answer": "The prompt 'Review your previous answer and find problems with your answer' was sourced from Kim et al., (2023), and authors adhered to it for experimental replication. Despite different prompts used in previous works, authors' results consistently show that models' reasoning performance does not improve and often decreases after self-correction.", "category": "Experimental_Exposition"}, {"question": "How does the concept of intrinsic self-correction differ from the recall of examples from parameter knowledge as described in the paper 'Large Language Models as Analogical Reasoners'?", "answer": "Analogical prompting, which involves recalling relevant problems and solutions from parameter knowledge, is a form of pre-hoc prompting. In contrast, self-correction is a type of post-hoc prompting. These methods differ fundamentally, so analogical prompting is not directly related to self-correction.", "category": "Method_Disambiguation"}, {"question": "Can authors provide details on where these methods rely on oracle labels, have specific problems with pre-prompts, or present incorrect benchmark results?", "answer": "Multi-Agent Debate (Du et al., 2023) has an unfair comparison to self-consistency, and Self-Refine (Madaan et al., 2023) has suboptimal pre-hoc prompt design.\n\n| Method                                              | Issue                                      |\n|-----------------------------------------------------|--------------------------------------------|\n| RCI (Kim et al., 2023); Reflexion (Shinn et al., 2023) | Use of oracle labels                        |\n| Multi-Agent Debate (Du et al., 2023)                | Unfair comparison to self-consistency       |\n| Self-Refine (Madaan et al., 2023)                   | Suboptimal pre-hoc prompt design            |", "category": "Experimental_Analysis"}, {"question": "Is breaking down a problem into sub-problems and conducting step-wise verification paired with self-consistency considered a form of self-correction, and how does this relate to improving reasoning without external feedback?", "answer": "Yes, breaking down a problem into sub-problems and conducting step-wise verification paired with self-consistency is considered helpful for self-verification, as discussed in Section 5 of the paper. However, enabling LLMs to interact with and learn from external feedback (e.g., interacting with the environment) is a separate direction that falls outside the scope of 'intrinsic self-correction' addressed in the paper.", "category": "Experimental_Analysis"}, {"question": "Did the study rely solely on the prompt 'Review your previous answer and find problems with your answer' without testing other prompts?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper include concrete cost numbers for various runs to illustrate the financial resources required for replication?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the table in Appendix C detail specific problems such as poorly constructed pre-prompts and incorrect benchmark results for prior methods?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "jsWCmrsHHs", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Deep Reinforcement Learning Guided Improvement Heuristic for Job Shop Scheduling", "authors": ["Cong Zhang", "Zhiguang Cao", "Wen Song", "Yaoxin Wu", "Jie Zhang"], "abstract": "Recent studies in using deep reinforcement learning (DRL) to solve Job-shop scheduling problems (JSSP) focus on construction heuristics. However, their performance is still far from optimality, mainly because the underlying graph representation scheme is unsuitable for modelling partial solutions at each construction step. This paper proposes a novel DRL-guided improvement heuristic for solving JSSP, where graph representation is employed to encode complete solutions. We design a Graph-Neural-Network-based representation scheme, consisting of two modules to effectively capture the information of dynamic topology and different types of nodes in graphs encountered during the improvement process. To speed up solution evaluation during improvement, we present a novel message-passing mechanism that can evaluate multiple solutions simultaneously. We prove that the computational complexity of our method scales linearly with problem size. Experiments on classic benchmarks show that the improvement policy learned by our method outperforms state-of-the-art DRL-based methods by a large margin.", "keywords": "Deep Reinforcement Learning;Graph Neural Network;Job Shop Scheduling;Combinatorial Optimization", "qa_pairs": [{"question": "How does the article address the generalization capabilities of the agents, particularly in terms of zero-shot performance and the underlying structural principles?", "answer": "The article addresses generalization capabilities by training the model on synthetic datasets generated using the Talliard method and evaluating it on benchmarking datasets following different distributions. Other benchmarking datasets, i.e., ABZ, FT, LA, SWV, ORB, and YN, are generated by following different distributions (please refer to Taillard, 1993; Adams et al., 1988; Fisher, 1963; Lawrence, 1984; Storer et al., 1992; Applegate & Cook, 1991; Yamada & Nakano, 1992, for details of the generation process of the public benchmarks).Most results in Table 1 report zero-shot generalization performance, except for Taillard benchmark instances of the same size as in training including 15×15 and 20×15. The model's good zero-shot generalization ability is explained by the local isomorphism in graphs, where disjunctive graphs of different sizes share similar local topological structures. Disjunctive graphs of different sizes are structurally similar in that each node has two parent nodes (a precedent constraint parent and a machine clique parent), and each node has corresponding two child nodes (a precedent constraint child and a machine clique child). Such a local topological structure remains constant regardless of the size of the graph. This structural similarity allows the topological features learned on small-scale graphs to generalize to larger-scale graphs, facilitated by the proposed CAM + TPM method, where TPM extracts features of the graph's topological structure.", "category": "Method_Mechanics"}, {"question": "How does the implementation of the proposed MDP in the paper align with the OpenAI Gym API, and what steps are being taken to ensure reproducibility and validation of the results?", "answer": "The implementation of the proposed MDP follows the convention of OpenAI Gym, utilizing interfaces such as `init()`, `step()`, and `reward()`. Additionally, the MDP modeling will be made public and integrated as an independent environment into the OpenAI Gym API, facilitating replication and validation of the results.", "category": "Experimental_Setup"}, {"question": "How is the MDP transition state calculated, and how is the transition space updated with the current status information of all jobs and machines?", "answer": "The MDP transition state is calculated using the $N_5$ neighbourhood structure. When the DRL agent selects an operation pair, such as $(O_{13}, O_{21})$, the $N_5$ operator swaps their processing order. For example, $O_{13}$ and $O_{21}$ becomes $O_{13} \\leftarrow O_{21}$. This may require breaking existing arcs and introducing new disjunctive arcs, such as breaking $O_{21} \\rightarrow O_{32}$ and adding $O_{13} \\rightarrow O_{32}$. In more complex cases, multiple steps are involved, such as breaking and flipping arcs and adding new ones to achieve the new processing order. For example, given a processing order on some machine $O_{1} \\rightarrow O_{2} \\rightarrow O_3 \\rightarrow O_4$, if we swap $O_{2}$ and $O_{3}$ (when $O_{2}$ and $O_{3}$ are the first/last two operations of some critical block), we need to break disjunctive arcs $O_{1} \\rightarrow O_{2}$ and $O_3 \\rightarrow O_4$, flip the orientation of arc $O_{2} \\rightarrow O_{3}$ as $O_{2} \\leftarrow O_{3}$, finally, add new arcs $O_3 \\rightarrow O_1$ and $O_2 \\rightarrow O_4$. So, the path after the transition will be $O_{1} \\rightarrow O_{3} \\rightarrow O_2 \\rightarrow O_4$. After the transition, the new disjunctive graph reflects the updated processing order for the machine assigned to the selected operation pair, while other machines remain unaffected. The starting time of every operation is then re-calculated using the proposed message-passing evaluator (Theorem 4.2).", "category": "Method_Mechanics"}, {"question": "How is the return $R_{t-j}^b$ computed in the n-step REINFORCE algorithm, and how does this approach address the potential issue of ignoring future rewards beyond n steps in the context of a local search method?", "answer": "The return $R_{t-j}^b$ is computed as the sum of rewards from steps $t-j$ to $t$, given by $R_{t-j}^b = \\sum_{t'=t-j}^{t}r(s_{t'}, a_{t'}) = C_{max}(s_{t-j}) - C_{max}(s^*_t)$, where $s_t^*$ is the best solution found till step $t$. The incumbent solution $s^*_t$ remains constant for any $n$ steps. The objective of improving the initial solution over $K$(e.g., $K=5000$) steps is divided into $K//n$ smaller objectives, allowing gradual updates to the DRL agent's policy. The sum of $n$-step returns $R_{t-j}^b$ equals the return $R_{K}^b$ at the end of improvement step $K$, ensuring future rewards are considered.", "category": "Method_Mechanics"}, {"question": "Why was Tabu search chosen as a baseline for comparison, and how does the proposed method compare to state-of-the-art metaheuristic methods for the job shop scheduling problem?", "answer": "Tabu search was chosen as a baseline because it is widely recognized in the manufacturing scheduling research community as one of the most promising metaheuristic methods for solving the job shop scheduling problem (JSSP) with state-of-the-art performance. The proposed method aims to present a framework that learns local operators for JSSP, which are stronger than hand-crafted ones and can potentially be integrated into more complex metaheuristics. While the primary goal is not to compete with state-of-the-art (SOTA) metaheuristics, the method implicitly compares to many SOTA approaches, such as an advanced memetic algorithm. From the results in Table 1, the proposed method achieves comparable performance (with a 5.96% average optimal gap) to SOTA metaheuristics across 26 datasets from various benchmarks. Additionally, the method demonstrates zero-shot generalization capabilities and maintains linear computational complexity with respect to problem size, as established in Theorem 1.", "category": "Method_Comparison"}, {"question": "How does the disjunctive graph model topological relationships among operations, and how does it differ from other visualization methods like Gantt charts?", "answer": "The disjunctive graph models topological relationships among operations using directed arrows that indicate the order of processing, with each arrow pointing from a preceding operation to its subsequent one. The primary purpose of the disjunctive graph is to encapsulate the topological relationships among the nodes, which are defined by these directional connections. This representation excels in illustrating inter-dependencies between operations, unlike Gantt charts, which are better suited for presenting explicit scheduling timelines. In formulating partial solutions within a construction step, the agent must possess accurate work-in-progress information, which encompasses specifics such as the extant load on each machine and the current status of jobs in process. The challenge encountered in this scenario relates to the difficulty in identifying suitable entities within the disjunctive graph that can store and convey such information effectively. While it is conceivable to encode the load of a machine using numerical values, no discrete node within the disjunctive graph explicitly embodies this information. Additionally, the disjunctive graph implicitly represents machine loads through the collective interpretation of processed operations, forming a 'machine clique,' which requires a nuanced approach to provide the agent with necessary information for constructing partial solutions.", "category": "Method_Comparison"}, {"question": "How are the reconfigurations of the disjunctive graph, beyond the two vertices indicated by the action, determined when swapping two operations, as illustrated by the changes to the edges connecting the red machine in Figure 2b?", "answer": "When swapping two operations, the deep reinforcement learning (DRL) agent selects a pair of operations and uses the $N_5$ operator to interchange their processing sequence. For example, swapping $(O_{13}, O_{21})$ changes the sequence from $O_{13}$ and $O_{21}$  to $O_{13} \\leftarrow O_{21}$. This swap may require additional modifications, such as breaking and establishing new disjunctive arcs. In some cases, a multi-step process is needed. For instance, given a sequential order of operations on a particular machine such as $O_{1} \\rightarrow O_{2} \\rightarrow O_3 \\rightarrow O_4$, and a swap between $O_{2}$ and $O_{3}$ within a critical block, multiple arc modifications are required. This involves removing the arcs from $O_{1}$ to $O_{2}$ and from $O_{3}$ to $O_{4}$, inverting the orientation of the arc from $O_{2}$ to $O_{3}$ to render it $O_{2} \\leftarrow O_{3}$, and instituting new arcs from $O_1$ to $O_3$ and from $O_2$ to $O_4$. The transitioned state results in an updated disjunctive graph with an altered processing sequence on the machine associated with the chosen operational pair, while sequences on other machines remain unchanged.", "category": "Method_Mechanics"}, {"question": "How can the proposed method be modified to solve other variants of shop scheduling problems, such as flexible job-shop scheduling problems (FJSSP) and flow shop scheduling problems (FSSP)?", "answer": "The proposed method can be applied to other variants of shop scheduling problems, such as FJSSP and FSSP, with minor modifications. For FJSSP, the method relies on the disjunctive graph-based state representation and critical block-based neighbourhood structure. The action involves reinserting an operation from its current machine into another eligible one or a different position in the same machine, ensuring the resulting disjunctive graph remains a DAG. In other words, the action removes the selected operation (and all its arcs) and then inserts it into another position (connecting all necessary arcs) in the disjunctive graph. The optional operations and their feasible insertion positions are defined by some neighbourhood structure, e.g., $N opt1$ and $N opt2$ proposed in the paper [AR3]. The action set and state transition of the MDP model need to be modified, while other components like the state and reward remain unchanged. The GNN model retains its effectiveness and linear computational complexity  since the state is the disjunctive graph, the same as the one in the manuscript (to see this, compare section 2 of paper [AR3] with the paper's disjunctive graph representation).. For FSSP, the method can be directly applied without changing the neighbourhood structure. Experimental results on FSSP instances from the Taillard dataset demonstrate the method's zero-shot generalization performance and superiority over hand-crafted baselines. From the experiment results, authors can observe that the local operator learned by method preserves the superiority to the hand-crafted ones, and the relative gap of the performance to CP-SAT is small.\n\n|           |  Tai-FSSP 20x5 | Tai-FSSP 20x10 |  Tai-FSSP 20x20 |\n|:---------:|:--------------:|:--------------:|:---------------:|\n|   CP-SAT  |  0.0% (46.3s)  |   0.01% (1h)   |   0.02% (1h)    |\n|  GD-5000  |  3.03% (5.3m)  | 14.32% (8.0m)  | 17.89% (18.3m)  |\n|  FI-5000  |  2.78% (6.1m)  | 12.25% (9.3m)  | 15.31% (22.1m)  |\n|  BI-5000  |  2.69% (5.7m)  | 12.79% (9.9m)  | 16.26% (24.5m)  |\n| Authors-5000 | 1.72% (72.0s)  | 9.24% (83.2s)  |  13.14% (1.8m)  |\n\n[AR3] Mastrolilli, M., \\& Gambardella, L. M. (2000). Effective neighborhood functions for the flexible job shop problem. *Journal of scheduling*, 3(1), 3-20.\n\n[AR4] Eric Taillard. Benchmarks for basic scheduling problems. *European Journal of Operational Research*, 64(2):278–285, 1993.", "category": "Method_Mechanics"}, {"question": "Did authors experiment with different reward calculation methods, and how did these alternatives impact the learning process compared to authors' chosen method?", "answer": "Authors experimented with a reward model based on consecutive state reward differences ($r_{t+1} - r_t$), which resulted in unpredictable learning behaviors due to the lack of a stable comparative reference. The chosen method, which rewards relative improvement over a benchmark solution, promotes continual progress and demonstrates the importance of reward scheme selection in reinforcement learning.", "category": "Experimental_Analysis"}, {"question": "Why was a baseline not used during learning for the REINFORCE algorithm in authors' proposed n-step REINFORCE framework, and what challenges arise in selecting an appropriate baseline for non-episodic Markov Decision Processes (MDPs)?", "answer": "Incorporating a baseline into the REINFORCE algorithm is promising for enhancing training stability, but selecting an appropriate baseline for the n-step REINFORCE framework is challenging due to policy updates occurring after every n steps of Monte Carlo sampling. This challenge is particularly pronounced in non-episodic Markov Decision Processes (MDPs), where conventional methods of computing the baseline at episode termination are not applicable. Currently, the method performs well without a baseline, but the authors will investigate designing a proper baseline to further enhance the method.", "category": "Motivation_Analysis"}]}
{"id": "di52zR8xgf", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis", "authors": ["Dustin Podell", "Zion English", "Kyle Lacey", "Andreas Blattmann", "Tim Dockhorn", "Jonas Müller", "Joe Penna", "Robin Rombach"], "abstract": "We present Stable Diffusion XL (SDXL), a latent diffusion model for text-to-image synthesis. Compared to previous versions of Stable Diffusion, SDXL leverages a three times larger UNet backbone, achieved by significantly increasing the number of attention blocks and including a second text encoder. Further, we design multiple novel conditioning schemes and train SDXL on multiple aspect ratios. To ensure highest quality results, we also introduce a refinement model which is used to improve the visual fidelity of samples generated by SDXL using a post-hoc image-to-image technique. We demonstrate that SDXL improves dramatically over previous versions of Stable Diffusion and achieves results competitive with those of black-box state-of-the-art image generators such as Midjourney.", "keywords": "Image Synthesis;Diffusion;Generative AI", "qa_pairs": [{"question": "What evidence or experiments demonstrate the effectiveness of 'conditioning on cropping' and 'multi-aspect training' in the proposed method?", "answer": "Crop-conditioning is applied as a finetuning technique to address issues in early versions of SD, such as random cropping augmentation leaking into synthesized samples. It is most beneficial when training on a fixed resolution (e.g., 512x512 pixels) but is not required for the final multi-aspect model. The effectiveness of size-conditioning is demonstrated through ablation studies in Table 2.", "category": "Experimental_Exposition"}, {"question": "Which factor contributes more to the better reconstruction performance of the autoencoder: the use of exponential moving average (EMA) or the larger batch size?", "answer": "Autoencoder Ablation: To assess the effect of (1) a larger batch size and (2) using an exponential moving average (EMA) of the weights, the authors conduct two experiments. In these experiments, the autoencoder is trained from scratch with (a) a total batch size of 8 and (b) a total batch size of 256. These models are not trained until convergence. For both experiments, reconstruction performance is evaluated on a fixed 10k subset of the COCO2014 validation set, comparing EMA and non-EMA weights for both models. rFID, PSNR, SSIM, and LPIPS are reported in a table below.\n\n| Model Name  | Batch Size   | EMA   | Global Steps   | rFID [↓]  | PSNR [↑]   | SSIM [↑]  | LPIPS [↓]   |\n| ----------- | ------------ | ----- | -------------- | --------- | ---------- | --------- | ----------- |\n| B           | 256          | ✗     | 500k           | 7.52      | 24.30      | 0.71      | 1.27        |\n| Be          | 256          | ✓     | 500k           | 7.35      | 24.33      | 0.72      | 1.24        |\n| A1          | 8            | ✗     | 500k           | 9.14      | 24.42      | 0.72      | 1.32        |\n| A1e         | 8            | ✓     | 500k           | 8.85      | 24.51      | 0.72      | 1.29        |\n| A2          | 8            | ✗     | 1.9M           | 6.28      | 25.05      | 0.74      | 1.08        |\n| A2e         | 8            | ✓     | 1.9M           | 5.57      | 25.05      | 0.74      | 1.07        |\n\nIt is observed that:\n1.  In all evaluations, the EMA variants demonstrate a consistent improvement compared to the non-EMA variant. This effect is amplified after more update steps (compare A1 and A2).\n2.  To evaluate the effect of the large batch size, the models are first compared when both are trained for the same number of gradient updates (A1(e) and B(e)): In this case, the large batch size provides a clear benefit in terms of rFID (7.34 vs 8.85), while the other metrics are relatively close. However, when the models are evaluated after they have trained for roughly the same wall time (500k (B) and 1.9M (A2) update steps), the larger number of updates is a more important factor, with A2e outperforming Be in all metrics. It is believed that training until convergence might change this outcome, as the B/A1 evaluation indicates that the large batch size provides a better update direction for the weights. This is likely to help in finding a loss-minimum. From this experiment, it is concluded that EMA-tracking has a significant effect on performance.\n", "category": "Experimental_Exposition"}, {"question": "What is the relative impact and priority of larger UNet versus each conditioning mechanism, particularly crop-conditioning, on the model's performance?", "answer": "Crop-conditioning acts as a finetuning technique to fix issues in early SD versions related to random cropping augmentation. It is most beneficial for training on fixed resolutions like 512x512 pixels but is not necessary for the final multi-aspect model. The effectiveness of size-conditioning is detailed in Tab. 2.", "category": "Experimental_Exposition"}, {"question": "How has the autoencoder (AE) been improved or advanced in this study?", "answer": "The autoencoder has been improved by adding EMA-based weight tracking and increasing the batch size used during training. EMA consistently provides better reconstruction metrics, and larger batch sizes enable better gradient updates. Ablation experiments confirmed that EMA variants improve performance across all metrics, and larger batch sizes lead to better weight update directions, enhancing performance in terms of rFID.", "category": "Method_Mechanics"}, {"question": "What are the reasons behind the 'no immediate benefit' observed when using Transformer architectures like DiT[1] and UViT[2] to train diffusion models?\n\n[1]: William Peebles and Saining Xie: Scalable Diffusion Models with Transformers\n\n[2]: Emiel Hoogeboom and Jonathan Heek and Tim Salimans: simple diffusion: End-to-end diffusion for high resolution images", "answer": "The images produced by DiT and UViT architectures were not of higher quality, possibly due to suboptimal hyperparameter choices during a brief experimentation phase. Despite this, the authors remain optimistic about these architectures' potential for scaling diffusion models to larger sizes and simplifying engineering efforts.", "category": "Experimental_Analysis"}, {"question": "What are the results of the ablation experiments on the autoencoder, specifically regarding the effects of larger batch size and the use of exponential moving average (EMA) of the weights?", "answer": "The ablation experiments on the autoencoder evaluated the effects of (1) larger batch size and (2) EMA of the weights. Two experiments were conducted: (a) training with a batch size of 8 and (b) training with a batch size of 256. These models are not trained until convergence because of limited resources and time..The models were evaluated on a fixed 10k subset of the COCO2014 validation set, comparing EMA and non-EMA variants. Key findings include: 1. EMA variants consistently improved performance compared to non-EMA variants, with the effect amplified after more update steps(compare A1 and A2). 2. When comparing models trained for the same number of gradient updates(A1(e) and B(e)), the larger batch size (256) provided a clear benefit in rFID (7.34 vs 8.85), while other metrics were close. However, when evaluated after similar wall time (500k vs 1.9M update steps), the larger number of updates (1.9M (A2)) led to better performance across all metrics. It is possible that training until convergence might change this outcome, as the B/A1 evaluation indicates that the large batch size provides a better update direction for the weights. This is likely to help in finding a loss-minimum. The authors concluded that EMA-tracking significantly improves performance.\n\n| Model Name  | Batch Size   | EMA   | Global Steps   | rFID [↓]  | PSNR [↑]   | SSIM [↑]  | LPIPS [↓]   |\n| ----------- | ------------ | ----- | -------------- | --------- | ---------- | --------- | ----------- |\n| B           | 256          | ✗     | 500k           | 7.52      | 24.30      | 0.71      | 1.27        |\n| Be          | 256          | ✓     | 500k           | 7.35      | 24.33      | 0.72      | 1.24        |\n| A1          | 8            | ✗     | 500k           | 9.14      | 24.42      | 0.72      | 1.32        |\n| A1e         | 8            | ✓     | 500k           | 8.85      | 24.51      | 0.72      | 1.29        |\n| A2          | 8            | ✗     | 1.9M           | 6.28      | 25.05      | 0.74      | 1.08        |\n| A2e         | 8            | ✓     | 1.9M           | 5.57      | 25.05      | 0.74      | 1.07        |", "category": "Experimental_Exposition"}, {"question": "What is the influence of the individual components, specifically the crop-conditioning and size-conditioning, on the overall performance of the proposed techniques?", "answer": "Crop-conditioning acts as a finetuning technique to address issues from previous SD versions related to random cropping expansion. It is especially beneficial for training at fixed resolutions like 512x512 pixels but is not necessary for the final multi-aspect model. The effectiveness of size conditioning is demonstrated in Table 2.", "category": "Experimental_Exposition"}, {"question": "What does the notation (512, 512) in Fig. 5 represent, and why do the rightmost examples show different behaviors (e.g., the top example resembling the leftmost image while the bottom example shows mainly the background)?", "answer": "The notation (512, 512) in Fig. 5 refers to crop conditioning inputs, but no actual cropping is applied. The rightmost examples are out-of-distribution (OOD), leading to less predictable behavior, which explains the differences in the images.", "category": "Concept_Understanding"}, {"question": "What are the effects of using a larger batch size and an exponential moving average (EMA) of the weights on the autoencoder's performance, as assessed through ablation experiments?", "answer": "The ablation experiments on the autoencoder evaluated the effects of (1) larger batch size and (2) EMA of the weights. Two experiments were conducted: (a) training with a batch size of 8 and (b) training with a batch size of 256. These models are not trained until convergence because of limited resources and time..The models were evaluated on a fixed 10k subset of the COCO2014 validation set, comparing EMA and non-EMA variants. Key findings include: 1. EMA variants consistently improved performance compared to non-EMA variants, with the effect amplified after more update steps(compare A1 and A2). 2. When comparing models trained for the same number of gradient updates(A1(e) and B(e)), the larger batch size (256) provided a clear benefit in rFID (7.34 vs 8.85), while other metrics were close. However, when evaluated after similar wall time (500k vs 1.9M update steps), the larger number of updates (1.9M (A2)) led to better performance across all metrics. It is possible that training until convergence might change this outcome, as the B/A1 evaluation indicates that the large batch size provides a better update direction for the weights. This is likely to help in finding a loss-minimum. The authors concluded that EMA-tracking significantly improves performance.\n\n| Model Name  | Batch Size   | EMA   | Global Steps   | rFID [↓]  | PSNR [↑]   | SSIM [↑]  | LPIPS [↓]   |\n| ----------- | ------------ | ----- | -------------- | --------- | ---------- | --------- | ----------- |\n| B           | 256          | ✗     | 500k           | 7.52      | 24.30      | 0.71      | 1.27        |\n| Be          | 256          | ✓     | 500k           | 7.35      | 24.33      | 0.72      | 1.24        |\n| A1          | 8            | ✗     | 500k           | 9.14      | 24.42      | 0.72      | 1.32        |\n| A1e         | 8            | ✓     | 500k           | 8.85      | 24.51      | 0.72      | 1.29        |\n| A2          | 8            | ✗     | 1.9M           | 6.28      | 25.05      | 0.74      | 1.08        |\n| A2e         | 8            | ✓     | 1.9M           | 5.57      | 25.05      | 0.74      | 1.07        |", "category": "Experimental_Exposition"}, {"question": "Is the effectiveness of size-conditioning ablated in Tab. 2 of the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the final multi-aspect model require the application of crop-conditioning?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is crop-conditioning applied as a finetuning technique to address issues with random cropping augmentation in early versions of SD?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "vqIH0ObdqL", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Can Large Language Models Infer Causation from Correlation?", "authors": ["Zhijing Jin", "Jiarui Liu", "Zhiheng Lyu", "Spencer Poff", "Mrinmaya Sachan", "Rada Mihalcea", "Mona T. Diab", "Bernhard Schölkopf"], "abstract": "Causal inference is one of the hallmarks of human intelligence. While the field of CausalNLP has attracted much interest in the recent years, existing causal inference datasets in NLP primarily rely on discovering causality from empirical knowledge (e.g., commonsense knowledge). In this work, we propose the first benchmark dataset to test the pure causal inference skills of large language models (LLMs). Specifically, we formulate a novel task Corr2Cause, which takes a set of correlational statements and determines the causal relationship between the variables. We curate a large-scale dataset of more than 200K samples, on which we evaluate seventeen existing LLMs. Through our experiments, we identify a key shortcoming of LLMs in terms of their causal inference skills, and show that these models achieve almost close to random performance on the task. This shortcoming is somewhat mitigated when we try to re-purpose LLMs for this skill via finetuning, but we find that these models still fail to generalize – they can only perform causal inference in in-distribution settings when variable names and textual expressions used in the queries are similar to those in the training set, but fail in out-of-distribution settings generated by perturbing these queries. Corr2Cause is a challenging task for LLMs, and can be helpful in guiding future research on improving LLMs’ pure reasoning skills and generalizability. Our data is at https://huggingface.co/datasets/causalnlp/corr2cause. Our code is at https://github.com/causalNLP/corr2cause.", "keywords": "Large Language Models;Natural Language Inference;Causal Reasoning;Correlation-to-Causation Inference;Benchmark Dataset;Causal Discovery", "qa_pairs": [{"question": "What are the fine-grained evaluation results for non-fine-tuned models similar to those presented in Table 6?", "answer": "The fine-grained evaluation results for non-fine-tuned models are provided in the table below. These results reveal that GPT-3.5 cannot recognize ancestor relations, GPT-4 fails at all direct causation recognition with parents and children, and RoBERTa MNLI only performs relatively correctly on collider relations. When the F1 score is zero, the positive accuracy results from predicting the negative class of that relation.\n\n| Model            | Relation Type   | F1    | P     | R     | Acc   |\n|------------------|-----------------|-------|-------|-------|-------|\n| GPT-3.5          | All             | 21.69 | 17.79 | 27.78 | 69.46 |\n| GPT-3.5          | Is-Parent       | 8.82  | 100   | 4.62  | 83.47 |\n| GPT-3.5          | Is-Ancestor     | 0     | 0     | 0     | 90.67 |\n| GPT-3.5          | Is-Child        | 9.84  | 100   | 5.17  | 85.33 |\n| GPT-3.5          | Is-Descendant   | 14.29 | 11.9  | 17.86 | 84    |\n| GPT-3.5          | Has-Collider    | 34.24 | 25.51 | 52.07 | 35.12 |\n| GPT-3.5          | Has-Confounder  | 15.33 | 8.86  | 56.76 | 37.8  |\n| GPT-4            | All             | 29.08 | 20.92 | 47.66 | 64.6  |\n| GPT-4            | Is-Parent       | 0     | 0     | 0     | 82.67 |\n| GPT-4            | Is-Ancestor     | 30.77 | 31.25 | 30.3  | 88    |\n| GPT-4            | Is-Child        | 0     | 0     | 0     | 84.53 |\n| GPT-4            | Is-Descendant   | 26.98 | 17.35 | 60.71 | 75.47 |\n| GPT-4            | Has-Collider    | 44.1  | 30.18 | 81.82 | 32.71 |\n| GPT-4            | Has-Confounder  | 20.67 | 11.53 | 100   | 23.86 |\n| RoBERTa MNLI     | All             | 22.79 | 34.73 | 16.96 | 82.5  |\n| RoBERTa MNLI     | Is-Parent       | 0     | 0     | 0     | 82.67 |\n| RoBERTa MNLI     | Is-Ancestor     | 0     | 0     | 0     | 91.2  |\n| RoBERTa MNLI     | Is-Child        | 0     | 0     | 0     | 84.53 |\n| RoBERTa MNLI     | Is-Descendant   | 0     | 0     | 0     | 92.53 |\n| RoBERTa MNLI     | Has-Collider    | 43.45 | 39.73 | 47.93 | 59.52 |\n| RoBERTa MNLI     | Has-Confounder  | 0     | 0     | 0     | 84.45 |\n", "category": "Experimental_Exposition"}, {"question": "Would incorporating instructions such as stating the rules of causal inference or including few-shot examples improve the performance of language models in the Corr2Cause task?", "answer": "There is merit in both the paper’s approach—evaluating LLMs based on simple queries a typical user might pose—and in incorporating instructions, which explores the potential for maximizing performance with tailored prompts. New experiments were conducted incorporating system prompts, few-shot examples, and chain-of-thought reasoning. Combining all three strategies into the enhanced query improved GPT-3.5’s F1 score by 4 percentage points compared to its original performance. However, despite these strategies, GPT-3.5 still performed poorly on the challenging Corr2Cause task.\n\n| Model                           | F1    | P | R | Acc |\n| ------------------------------- | ----- | --------- | ------ | -------- |\n| GPT-3.5 (Plain Query; Original) | 21.69 | 17.79     | 27.78  | 69.46    |\n| GPT-3.5 (Enhanced Query)        | 25.44 | 17.29     | 48.11  | 52.01    |\n\n As for experimental details, for the enhanced query, the authors (1) use a system prompt to indicate the expertise (\"You are a highly intelligent question-answering bot with profound knowledge of causal inference.\"); (2) include two few-shot examples, one positive, one negative; and (3) add the chain-of-thought prompting with \"Let's think step by step.\") to induce language models to produce step-by-step reasoning.", "category": "Experimental_Exposition"}, {"question": "What justifies the use of 'pure causal inference' as an appropriate task for assessing the causal reasoning capabilities of large language models, especially considering the historical development of formal reasoning approaches like those used in Corr2Cause?", "answer": "The paper's proposal for the Corr2Cause task is inspired by interdisciplinary insights, particularly from philosophy, where reasoning is categorized into pure (rationalism) and empirical (empiricism) reasoning. The latter, empirical reasoning, is usually framed as commonsense reasoning in existing NLP literature, which people can use to reason before e.g., DAGs were invented. In contrast, pure reasoning, or a priori reasoning, involves logic and abstract thought, independent of sensory experience. In the causal domain, the authors believe that the necessity of pure causal reasoning is analogous to that of math, which aids human understanding independent of sensory experience. Based on the natural story realizations in Appendix B, the paper's symbolic Corr2Cause reasoning pattern holds regardless of the instantiation of the variables, which makes the paper rely on these inference rules independent of experience, or even overcome cognitive biases that confuse correlation with causation. Upon reviewing existing NLP datasets on causal reasoning, the authors found a predominant focus on empirical commonsense reasoning, with little to no emphasis on pure reasoning in the literature. Recognizing the longstanding philosophical value of pure reasoning, which dates back to early Greek philosophers like Parmenides and Pythagoras and is further developed by figures like Plato, Descartes, and Kant, pure reasoning possesses significant potential to enrich human knowledge beyond empirical experiences. Both rational and empirical forms of knowledge are crucial, complementing each other in expanding the realm of human understanding. This goal is what the paper aims to highlight and explore through the Corr2Cause task.", "category": "Motivation_Analysis"}, {"question": "How do the design choices of Corr2Cause compare with those of other existing causal reasoning benchmarks, and what categories do these benchmarks fall into?", "answer": "Corr2Cause differs from other benchmarks by focusing on pure causal reasoning, whereas most existing datasets fall into two categories: commonsense causality, which relies on empirical experiences, and an information extraction framing to identify cause and effect from text, which is more of a linguistics task. A summary table of the existing body of literature is provided below.\n\n|                                                              | Pure Causal Reasoning | Commonsense Causality/Empirical Causal Reasoning | Causal Relation Extraction (Relying on Linguistics) |\n| ------------------------------------------------------------ | --------------------- | ------------------------------------------- | --------------------------------------------------- |\n| **Type 1**                                                   |                       |                                             |                                                     |\n| COPA [(2012)](https://people.ict.usc.edu/~gordon/copa.html)  | ✗                     | ✓                                           | ✗                                                   |\n| Event2Mind [(2018)](https://uwnlp.github.io/event2mind/)     | ✗                     | ✓                                           | ✗                                                   |\n| ATOMIC [(2019)](https://allenai.org/data/atomic)             | ✗                     | ✓                                           | ✗                                                   |\n| SocialIQA [(2019)](https://allenai.org/data/socialiqa)       | ✗                     | ✓                                           | ✗                                                   |\n| TimeTravel [(2019)](https://paperswithcode.com/dataset/timetravel) | ✗                     | ✓                                           | ✗                                                   |\n| Goal-Step [(2020)](https://aclanthology.org/2020.emnlp-main.374/) | ✗                     | ✓                                           | ✗                                                   |\n| Abductive (ART) [(2020)](https://paperswithcode.com/dataset/art-dataset) | ✗                     | ✓                                           | ✗                                                   |\n| Com2Sense [(2021)](https://aclanthology.org/2021.findings-acl.78/) | ✗                     | ✓                                           | ✗                                                   |\n| CRASS [(2022)](https://aclanthology.org/2022.lrec-1.229/)    | ✗                     | ✓                                           | ✗                                                   |\n| **Type 2**                                                   |                       |                                             |                                                     |\n| SE2021 Task8 [(2010)](https://paperswithcode.com/dataset/semeval-2010-task-8) | ✗                     | ✗                                           | ✓                                                   |\n| EventCausality [(2011)](https://aclanthology.org/D11-1027/)  | ✗                     | ✗                                           | ✓                                                   |\n| Causal-TimeBank [(2014)](https://github.com/paramitamirza/Causal-TimeBank) | ✗                     | ✗                                           | ✓                                                   |\n| CaTeRS ([2016](https://paperswithcode.com/paper/caters-causal-and-temporal-relation-scheme)) | ✗                     | ✗                                           | ✓                                                   |\n| BECauSE ([2017](https://aclanthology.org/W17-0812/))         | ✗                     | ✗                                           | ✓                                                   |\n| TellMeWhy [(2021)](https://aclanthology.org/2021.findings-acl.53/) | ✗                     | ✗                                           | ✓                                                   |\n| **Type 3**                                                   |                       |                                             |                                                     |\n| **Corr2Cause Inference (Ours)**                              | ✓                     | ✗                                           | ✗                                                   |", "category": "Method_Comparison"}, {"question": "Could small-scale GPT-style models be fine-tuned to evaluate their generalization performance on paraphrases and new variables, and what were the results of these experiments?", "answer": "The authors conducted experiments with small-scale GPT-style models, including gpt2 (124M parameters), gpt2-large (774M), gpt2-xl (1.5B), and the LLaMa series (LLaMa-7B and LLaMa2-7B) using LoRA versions. The results show that larger or later models in each series achieve better performance, with GPT2-XL after fine-tuning (gpt2-xl-ft) achieving the best performance, comparable to the fine-tuned Roberta model's 94.74% F1.\nThe table below shows the overall performance results:\n\n| Model          | F1     | Precision | Recall  | Accuracy |\n|----------------|--------|-----------|---------|----------|\n| gpt2-ft        | 89.18  | 88.03     | 90.35   | 96.66    |\n| gpt2-large-ft  | 94.29  | 92.18     | 96.49   | 98.22    |\n| gpt2-xl-ft     | 94.30  | 91.94     | 96.78   | 98.22    |\n| llama-ft       | 91.98  | 88.62     | 95.61   | 97.46    |\n| llama2-ft      | 92.92  | 90.11     | 95.91   | 97.77    |\n\n\nHere are the fine-grained performance data for the models, as shown in the table below. Due to space limitations, only a portion of the GPT-2 model series is displayed.\n\n| Model       | Relation Type    | F1    | P     | R     | Acc   |\n|-------------|------------------|-------|-------|-------|-------|\n| gpt2-ft     | All              | 89.18 | 88.03 | 90.35 | 96.66 |\n| gpt2-ft     | Is-Parent        | 91.04 | 88.41 | 93.85 | 96.8  |\n| gpt2-ft     | Is-Ancestor      | 90.91 | 90.91 | 90.91 | 98.4  |\n| gpt2-ft     | Is-Child         | 95    | 91.94 | 98.28 | 98.4  |\n| gpt2-ft     | Is-Descendant    | 87.72 | 86.21 | 89.29 | 98.13 |\n| gpt2-ft     | Has-Collider     | 82.64 | 82.64 | 82.64 | 88.74 |\n| gpt2-ft     | Has-Confounder   | 97.3  | 97.3  | 97.3  | 99.46 |\n| gpt2-large-ft | All             | 94.29 | 92.18 | 96.49 | 98.22 |\n| gpt2-large-ft | Is-Parent       | 95.52 | 92.75 | 98.46 | 98.4  |\n| gpt2-large-ft | Is-Ancestor     | 95.65 | 91.67 | 100   | 99.2  |\n| gpt2-large-ft | Is-Child        | 94.12 | 91.8  | 96.55 | 98.13 |\n| gpt2-large-ft | Is-Descendant   | 98.25 | 96.55 | 100   | 99.73 |\n| gpt2-large-ft | Has-Collider    | 93.88 | 92.74 | 95.04 | 95.98 |\n| gpt2-large-ft | Has-Confounder  | 89.47 | 87.18 | 91.89 | 97.86 |\n| gpt2-xl-ft  | All              | 94.3  | 91.94 | 96.78 | 98.22 |\n| gpt2-xl-ft  | Is-Parent        | 95.52 | 92.75 | 98.46 | 98.4  |\n| gpt2-xl-ft  | Is-Ancestor      | 92.75 | 88.89 | 96.97 | 98.67 |\n| gpt2-xl-ft  | Is-Child         | 94.92 | 93.33 | 96.55 | 98.4  |\n| gpt2-xl-ft  | Is-Descendant    | 94.92 | 90.32 | 100   | 99.2  |\n| gpt2-xl-ft  | Has-Collider     | 93.98 | 91.41 | 96.69 | 95.98 |\n| gpt2-xl-ft  | Has-Confounder   | 93.15 | 94.44 | 91.89 | 98.66 |\n\n Robustness tests on these models, including Variable Refactoring (VR) and Paraphrasing (P), revealed a significant drop in performance across all fine-tuned models, consistent with the paper's earlier observations. Variable refactorization caused a larger performance drop than paraphrasing.\n\n| Model          | F1     | Precision | Recall | Accuracy |\n|----------------|--------|-----------|--------|----------|\n| gpt2-ft-P      | 56.76  | 57.10     | 56.43  | 86.91    |\n| gpt2-large-ft-P| 55.95  | 56.97     | 54.97  | 86.82    |\n| gpt2-xl-ft-P   | 60.32  | 52.25     | 71.35  | 85.71    |\n| llama-ft-P     | 56.41  | 55.00     | 57.89  | 86.38    |\n| llama2-ft-P    | 52.24  | 67.28     | 42.69  | 88.11    |\n| gpt2-ft-VR     | 31.70  | 25.63     | 41.52  | 72.75    |\n| gpt2-large-ft-VR| 31.99 | 34.11     | 30.12  | 80.50    |\n| gpt2-xl-ft-VR  | 43.95  | 48.25     | 40.35  | 84.33    |\n| llama-ft-VR    | 53.92  | 77.90     | 41.23  | 89.27    |\n| llama2-ft-VR   | 49.47  | 91.34     | 33.92  | 89.45    |\n", "category": "Experimental_Analysis"}, {"question": "What approach was used for dataset construction regarding the inclusion of examples and hypotheses in the Corr2Cause dataset?", "answer": "All possible examples and hypotheses were included in the Corr2Cause dataset to ensure comprehensive coverage of inference tasks involving all combinations of correlations among the given set of variables.", "category": "Experimental_Setup"}, {"question": "What do the variables in the dataset construction process represent, and how are they used to distill reasoning patterns?", "answer": "The variables in the dataset are symbolically named and represent entities, such as 'eating junk food (A), obesity (C), and watching television (B).' The inference problem can be: Premise: Let’s consider three factors: eating junk food (A), obesity (C), and watching television (B). There is a correlation between eating junk food and obesity, and between watching television and obesity. However, eating junk food and watching television are independent from each other. Hypothesis: Eating junk food directly affects obesity.They are used to distill reasoning patterns by setting up inference problems with premises and hypotheses, primarily employing symbolic forms to improve pure causal reasoning. This approach allows for future instantiation of variables using various real-world entities to generate natural stories for practical applications like combating misinformation.", "category": "Concept_Understanding"}, {"question": "How does the paper ensure the reliability and confidence of the results generated by the PC algorithm for causality discovery, given the absence of explicit data samples for statistical testing?", "answer": "The paper selected the PC algorithm for its suitability in processing inputs formatted as correlations. Traditionally, there are two main types of methods for causal discovery: constraint-based methods (based on the dependence and independence relations among variables), and score-based methods (which take in the data distribution and rank the possible causal graphs by their scores given the data, e.g., Bayesian Information Criterion (BIC) score). In the paper’s dataset, the authors made a design choice that natural language texts usually do not explicitly contain data samples to run statistical tests on (which is where the confidence scores come from), but instead they typically contain observations of which variables are related or independent. Therefore, the authors designed the paper’s task as the Corr2Cause inference task, which looked at the core task of inferring causal conclusions based on the presented correlations, and bypasses the step of statistical testing from raw data samples. For this task with correlations as input, constraint-based methods, specifically the most classic PC algorithm, are a great fit. As the dependencies and independences are provided, there will be no statistical tests to produce confidence scores, but only algorithmic reasoning to identify the causal graphs deterministically. The remaining major requirement of the PC algorithm is the faithfulness assumption, which the paper introduces in the preliminaries section. The paper adheres to this faithfulness assumption, which is generally met in real-world scenarios, although exceptions can occur if the causal mechanism is intricately designed to obscure statistical dependencies.", "category": "Method_Mechanics"}, {"question": "Were all possible examples and hypotheses included in the construction of the Corr2Cause dataset?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "CfXh93NDgH", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "WizardLM: Empowering Large Pre-Trained Language Models to Follow Complex Instructions", "authors": ["Can Xu", "Qingfeng Sun", "Kai Zheng", "Xiubo Geng", "Pu Zhao", "Jiazhan Feng", "Chongyang Tao", "Qingwei Lin", "Daxin Jiang"], "abstract": "Training large language models (LLMs) with open-domain instruction following data brings colossal success. However, manually creating such instruction data is very time-consuming and labor-intensive. Moreover, humans may struggle to produce high-complexity instructions. In this paper, we show an avenue for creating large amounts of instruction data with varying levels of complexity using LLM instead of humans. Starting with an initial set of instructions, we use our proposed Evol-Instruct to rewrite them step by step into more complex instructions. Then, we mix all generated instruction data to fine-tune LLaMA. We call the resulting model WizardLM. Both automatic and human evaluations consistently indicate that WizardLM outperforms baselines such as Alpaca (trained from Self-Instruct) and Vicuna (trained from human-created instructions). The experimental results demonstrate that the quality of instruction-following dataset crafted by Evol-Instruct can significantly improve the performance of LLMs.", "keywords": "Large Language Model;Instruction Fine-tuning", "qa_pairs": [{"question": "What is the purpose and methodology of the In-breadth Evolving analysis, and how does it demonstrate the diversity of the dataset compared to ShareGPT and Alpaca (Self-Instruct)?", "answer": "The Analysis of In-breadth Evolving section aims to enhance topic coverage, skill coverage, and overall dataset diversity. The methodology involves using BERT to encode instructions into 768-dimensional embeddings, applying t-SNE to reduce the embeddings to 2 dimensions, and using k-means clustering to partition instructions into 20 clusters for visualization. As shown in Figure 7, the data points of the dataset are more dispersed than those of ShareGPT and Alpaca (Self-Instruct), indicating better topic diversity.", "category": "Method_Mechanics"}, {"question": "What is the novelty and technical contribution of the proposed Evol-Instruct method compared to Self-Instruct[1], and how does it improve the performance of pre-trained language models?\n\n[1] Self-Instruct: Aligning Language Models with Self-Generated Instructions", "answer": "Motivation & Novelty: Self-Instruct [1] aims to address the problem of 'human-written instruction data that is often limited in quantity', aiming to 'provide an almost annotation-free method' based on LLM to automatically generate more instruction data. The paper's Evol-Instruct is not aimed at increasing the quantity of instructions but rather explores whether increasing the complexity of individual instruction contributes to improving the supervised fine-tuning performance of LLM. The paper's novelty lies in: (a) The paper is the first to explicitly propose that increasing the complexity of instructions is beneficial for improving the instruction-following capabilities of pre-trained language models. (b) The paper proposes a brand-new automated method to enhance the complexity and difficulty of existing instruction data without the need for human annotators. Both automatic and human evaluations consistently indicate that the model trained using Evol-Instruct (WizardLM) surpasses the baseline model (Alpaca) trained using Self-Instruct with a large margin in a broad spectrum of benchmarks. In the experiment, the baselines, Alpaca-13b and Vicuna-13b, were trained on data that also came from ChatGPT. The paper's experiment shows that under the condition of using the same powerful LLM, the model trained by Evol-Instruct will perform better on a wide range of benchmarks. Further, when the paper uses a weaker model LlaMA-2-70B-Chat to replace ChatGPT to execute Evol-Instruct, the trained model is still stronger than the models (i.e., Alpaca-13b and Vicuna-13b) trained by ChatGPT outputs.", "category": "Method_Comparison"}, {"question": "What is the token consumption for generating the datasets (ShareGPT, Alpaca, and WizardLM), and how does the performance of WizardLM compare to Alpaca and Vicuna when token consumption is controlled?", "answer": "The paper's motivation is to investigate whether, under the condition of keeping the number of instructions constant, simply increasing the complexity of each instruction can be beneficial to the SFT performance of LLMs. An increase in token consumption is a reasonable expectation for this approach. The token consumption for generating 70k ShareGPT data is 118.3M, for Alpaca data is 3.02M, and for 70k WizardLM data is 18.84M. To verify that the paper’s method can still outperform the self-instruct method for generating Alpaca when the token consumption is exactly the same, the authors randomly selected 11,300 samples from the paper's 70k data, totaling 3.02M tokens, equivalent to Alpaca. The Wizard-13b(3M) model was then trained. The results, as shown in the following table, indicate that Wizard-13b(3M) still significantly surpasses Alpaca-13b in all benchmarks. It's worth noting that the model the authors trained with only 3M Tokens can surprisingly perform comparably to the Vicuna-13b trained with 118M Tokens.\n\n\n| Model             | Avg.  | MMLU   | ARC    | HellaSwag | TruthfulQA | HumanEval | GSM8k | AlpacaEval | MT-Bench | WizardEval |\n|-------------------|-------|--------|--------|-----------|------------|-----------|-------|------------|----------|------------|\n| Alpaca-13b        | 43.44 | 46.63  | 51.20  | 76.31     | 41.62      | 9.2       | 8.35  | 33.25      | 4.78     | 76.6       |\n| Vicuna-13b        | 54.60 | 50.84  | 51.71  | 79.94     | 52.68      | 12.5      | 24.34 | 70.43      | 6.21     | 86.9       |\n| WizardLM-13b (3M) | 54.92 | 50.73  | 53.48  | 78.51     | 49.06      | 19.3      | 29.96 | 70.17      | 5.93     | 83.8       |\n| WizardLM-13b      | 58.96 | 52.92  | 57.25  | 80.88     | 50.55      | 24.0      | 37.15 | 75.31      | 6.35     | 89.1       |\n", "category": "Method_Comparison"}, {"question": "How does the performance of WizardLM-13b, using Evol-Instruct, compare with other instruction data generation methods such as Baize, CAMEL, and Tulu, and what is the role of seed instructions in these comparisons?", "answer": "Seed instructions are very important for the model performance. So the paper chooses the training data of Alpaca-13b as the seed instructions for its execution of Evol-Instruct. The  Table 1 experimental results show that, based on the same seed instructions, WizardLM-13b after executing Evol-Instruct significantly outperforms Alpaca-13b on a series of benchmarks with a large margin. Although ShareGPT has different seed instructions from Alpaca data, its quality is widely recognized as far superior to Alpaca data. The paper chooses to compare with Vicuna-13b, which uses ShareGPT as training data, in order to demonstrate that even a relatively low-quality instruction-following dataset, after applying the paper's proposed Evol-Instruct, can surpass higher-quality data.To further verify the effectiveness of the paper’s method, ShareGPT is used as a new seed, and the paper’s evol-instruct method is applied to it to produce a new 70k dataset, then a model named WizardLM-13b (ShareGPT Seed) is trained. As shown in the following table, the average performance across nine benchmarks of this model is even higher than the previous version of WizardLM (the seed is Alpaca data). The consistent results of Alpaca data and ShareGPT as evolutionary seeds indicate that Evol-Instruct can significantly improve the performance of fine-tuning pre-trained language models using specific instruction data. \n\n\n| Model                      | Avg.  | MMLU  | ARC   | HellaSwag | TruthfulQA | HumanEval | GSM8k | AlpacaEval | MT-Bench | WizardEval |\n|----------------------------|-------|-------|-------|-----------|------------|-----------|-------|------------|----------|------------|\n| WizardLM-13b               | 58.96 | 52.92 | 57.25 | 80.88     | 50.55      | 24.0      | 37.15 | 75.31      | 6.35     | 89.1       |\n| WizardLM-13b (ShareGPT Seed) | 61.87 | 50.92 | 60.24 | 81.39     | 54.56      | 25.0      | 31.46 | 86.32      | 6.76     | 99.3       |\n\nAdditionally, the paper adds more baselines such as Baize, CAMEL, and Tulu for comparisons with other methods for instruction data generation. The extensive results indicate that the paper's WizardLM also significantly surpasses these open source models, which demonstrates the convincing quality of instruction-following datasets crafted by Evol-Instruct.", "category": "Experimental_Analysis"}, {"question": "How was the proposed Evol-Instruct method tested to ensure it is not dependent on ChatGPT?", "answer": "To ensure the Evol-Instruct method is not dependent on ChatGPT, LlaMA-2-70B-Chat was used as a substitute. The entire evolution process was re-executed starting from Alpaca data, and the WizardLM(LlaMA-2-70B-Chat) model was trained with 70k randomly selected data. The results showed that WizardLM(LlaMA-2-70B-Chat) significantly outperformed Alpaca-13b trained solely on Alpaca data and even outperformed Vicuna-13B trained with more ChatGPT tokens.\n\n\n| Model                              | Avg.  | MMLU  | ARC   | HellaSwag | TruthfulQA | HumanEval | GSM8k | AlpacaEval | MT-Bench | WizardEval |\n|------------------------------------|-------|-------|-------|-----------|------------|-----------|-------|------------|----------|------------|\n| Alpaca-13b                         | 43.44 | 46.63 | 51.20 | 76.31     | 41.62      | 9.2       | 8.35  | 33.25      | 4.78     | 76.6       |\n| Vicuna-13b                         | 54.60 | 50.84 | 51.71 | 79.94     | 52.68      | 12.5      | 24.34 | 70.43      | 6.21     | 86.9       |\n| WizardLM-13b                       | 58.96 | 52.92 | 57.25 | 80.88     | 50.55      | 24.0      | 37.15 | 75.31      | 6.35     | 89.1       |\n| WizardLM-13b(LlaMA-2-70B-Chat Evol) | 56.27 | 51.09 | 57.34 | 79.12     | 48.76      | 19.5      | 33.83 | 70.47      | 6.18     | 84.5       |\n\n", "category": "Experimental_Exposition"}, {"question": "Were the created datasets compared with diverse instruction tuning datasets (such as Natural Instructions, Supernatural Instructions, and the training data of FLAN-T5)?", "answer": "Yes, the authors extracted 70k data from the Supernatural Instructions dataset and fine-tuned llama-13b with the same training hyperparameters as theirs, resulting in the model LlaMA(SNI). The experimental results, as shown in the table, indicate that LlaMA(SNI) performs well on classic QA benchmarks, such as MMLU, surpassing Alpaca, Vicuna, and WizardLM. However, its performance is less impressive on math, code, and open-domain test-beds like MT-Bench.\n\n| Model          | Avg.  | MMLU  | ARC   | HellaSwag | TruthfulQA | HumanEval | GSM8k | AlpacaEval | MT-Bench | WizardEval |\n|----------------|-------|-------|-------|-----------|------------|-----------|-------|------------|----------|------------|\n| WizardLM-13b   | 58.96 | 52.92 | 57.25 | 80.88     | 50.55      | 24.0      | 37.15 | 75.31      | 6.35     | 89.1       |\n| LlaMA(SNI)     | 37.73 | 54.9  | 54.95 | 80.4      | 38.69      | 4.2       | 5.79  | 13.67      | 2.86     | 58.4       |", "category": "Experimental_Exposition"}, {"question": "What are the results regarding the applicability of the Evol-Instruct method to large language models other than Llama?", "answer": "The study investigated the applicability of Evol-Instruct to other LLMs by training Mistral-7B on Alpaca data and evolved Alpaca data. The results showed that WizardLM (Mistral) significantly outperformed Alpaca (Mistral) across various benchmarks, demonstrating the method's effectiveness.\n\n| Model             | Avg.  | MMLU   | ARC    | HellaSwag | TruthfulQA | HumanEval | GSM8k | AlpacaEval | MT-Bench | WizardEval |\n|-------------------|-------|--------|--------|-----------|------------|-----------|-------|------------|----------|------------|\n| Alpaca (Mistral)  | 52.87 | 56.34  | 55.38  | 79.49     | 43.92      | 19.2      | 32.05 | 54.26      | 5.47     | 80.5       |\n| WizardLM (Mistral)| 65.81 | 60.7   | 57.47  | 82.08     | 51.79      | 37.8      | 59.49 | 80.7       | 7.1      | 91.3       |\n", "category": "Experimental_Exposition"}, {"question": "How was the evaluation of difficulty and diversity by ChatGPT validated, and what evidence supports the consistency of its results with GPT4 and human annotators?", "answer": "The evaluation of difficulty by ChatGPT was validated by sampling 600 instructions (100 from each dataset) and comparing the results with assessments from GPT4 and 5 human annotators. The results, presented in a table, show a high degree of consistency in the trend of difficulty changes across ChatGPT, GPT4, and manual annotation.\n\n\n|               | ShareGPT | Alpaca | C1  | C2  | C3  | C4  |\n|----------|--------|-----|-----|-----|-----|-----|\n| GPT-3.5  | 4.63   | 3.00| 5.48| 6.35|6.84 |7.08|\n| GPT-4    | 4.31   | 2.69| 4.68| 4.90|5.37| 5.54|\n| Human    | 4.55   | 3.15| 5.51| 5.86|6.49| 6.82|\n\n", "category": "Experimental_Exposition"}, {"question": "Why is only the average score of difficulty and complexity reported in Figure 5, and how is diversity addressed separately in the analysis?", "answer": "In Figure 5, only the average score of 'difficulty and complexity' (which are synonyms) is reported, and this score does not include 'diversity.' For 'diversity,' the analysis is presented in 'Section 4.5 - Analysis of In-breadth Evolving' and Figure 7 in Appendix J, where t-SNE and k-means are applied to BERT embeddings of instructions to indicate topic diversity across self-instruct, ShareGPT, and evol-instruct datasets.", "category": "Experimental_Setup"}, {"question": "How does the size and diversity of the WizardEval dataset compare to other benchmark datasets, and what evidence supports its suitability as a sound and fair evaluation benchmark?", "answer": "(1) Figure 6 in Appendix H illustrates the distribution of the instances and skills in The paper's WizardEval set. It contains 29 distinct skills that represent the main requirements of humanity, such as Coding Generation & Debugging, Math, Reasoning, Complex Formats, Writing, Extensive Disciplines. It consists of 218 instances, each of which is an instruction for a specific skill. (2) Compared the paper's test set with Vicuna’s test set, which is a benchmark dataset for evaluating instruction-following models. It can be found that Vicuna’s test set has only 80 instances and 9 skills and is much smaller and less diverse than the paper's. (3) Figure 4 in the paper shows how the difficulty and complexity of the test data vary across different instances. The paper's test data has a more uniform distribution, meaning that it contains instructions with different levels of difficulty and complexity. On the other hand, Vicuna and Alpaca have a skewed distribution, meaning that Vicuna and Alpaca mostly contain instructions with low difficulty and complexity. This indicates that these two corpora are not able to handle the evaluation on more complex and demanding scenarios.", "category": "Method_Comparison"}, {"question": "How was the correctness of the difficulty and diversity scores measured, and what evidence supports the reliability of ChatGPT's difficulty score?", "answer": "The diversity score was measured using BERT, t-SNE, and k-means. To validate the correctness of ChatGPT's difficulty score, a new experiment was conducted where 300 instruction pairs were randomly selected with each pair formed by independently sampling two instructions from six datasets. ChatGPT and five human annotators judged the difficulty of each pair . The Kappa score between humans was 0.68, and between ChatGPT and human majority voting was 0.66, indicating strong agreement and reliability of ChatGPT's difficulty score.", "category": "Experimental_Exposition"}, {"question": "How does the repeated application of the Evol-Instruct process address the risk of overfitting the model to specific types of instructions?", "answer": "Executing Evol-Instruct multiple times may lead to overfitting to specific topics, which is mitigated by introducing In-Breadth Evolving alongside In-Depth Evolving. Chapter 4.5 and Appendix J show that the evolved instruction data has better topic diversity than the original data, indicating an attempt to address overfitting.", "category": "Method_Mechanics"}, {"question": "What methods are used to ensure that the evolved instructions maintain their intended purpose and do not deviate from the original goal?", "answer": "Two methods are used: (1) Extensive case studies and prompt refinement were conducted, with the authors testing nearly 300 cases to ensure quality. For instance, it is found that the Complicating Input of In-Depth Evolving is difficult for ChatGPT to accurately perform evolution in a zero-shot setting. To address this, a few-shot approach, which involves providing a few appropriate examples, is adopted to ensure that this evolutionary method can work properly with ChatGPT. (2) An 'Elimination Evolving' method was proposed, where ChatGPT evaluates whether evolved instructions retain the same constraints, requirements, depth, and breadth as the original. If they match, the evolved instructions are eliminated.", "category": "Method_Mechanics"}, {"question": "What measures are proposed to prevent potential biases introduced during the instruction evolution process?", "answer": "Two measures are proposed: (1) Using ChatGPT to detect toxic or biased content in new instructions and answers, and (2) sampling evolved instructions and hiring human annotators to identify potential biases. A classifier is then trained using this data to filter out biased instructions during the evolution process.", "category": "Method_Mechanics"}, {"question": "How does WizardLM address instructions that are not included in the evolved set?", "answer": "WizardLM addresses instructions not included in the evolved set through In-Breadth Evolving, which evolves entirely new instructions to enhance topic coverage, skill coverage, and overall diversity. This approach is demonstrated in Section 4.5 and Appendix J, showing improved diversity in the evolved instruction data compared to the original data.", "category": "Method_Mechanics"}, {"question": "How does incorporating more human feedback during the instruction evolution process help refine the quality of instructions?", "answer": "Incorporating more human feedback during the instruction evolution process helps refine the quality of instructions by using a 'human-in-the-loop evolution' approach.Similar to OpenAI's RLHF method. This involves evolving instructions multiple times, hiring people to score and rank the evolved instructions, and training a reward model for instruction quality based on this human-annotated data. During fine-tuning, high-scoring instructions are selected, and the gradient update weight is adjusted according to the instruction score, with high-scoring cases receiving larger gradient updates and low-scoring cases receiving smaller updates.", "category": "Method_Mechanics"}, {"question": "What evidence supports the scalability of the Evol-Instruct process for larger models or datasets?", "answer": "The scalability of the Evol-Instruct process is supported by experiments with three new trained wizard models: WizardLM 65b and WizardLM 70b, trained from Llama-1 65 and Llama-2 70B, respectively, and WizardLM-13b using 70K ShareGPT data as the seed. The results in the provided table demonstrate scalable performance for larger models and datasets.\n\n| Model                    | Avg.  | MMLU   | ARC    | HellaSwag | TruthfulQA | HumanEval | GSM8k | AlpacaEval | MT-Bench | WizardEval |\n|--------------------------|-------|--------|--------|-----------|------------|-----------|-------|------------|----------|------------|\n| WizardLM-13b            | 58.96 | 52.92  | 57.25  | 80.88     | 50.55      | 24.0      | 37.15 | 75.31      | 6.35     | 89.1       |\n| WizardLM-13b (ShareGPT Seed) | 61.87 | 50.92  | 60.24  | 81.39     | 54.56      | 25.0      | 31.46 | 86.32      | 6.76     | 99.3       |\n| WizardLM-65b            | 69.40 | 62.09  | 65.83  | 85.48     | 52.19      | 36.5      | 66.39 | 87.50      | 7.12     | 97.5       |\n| WizardLM-70b            | 71.33 | 63.32  | 64.52  | 83.21     | 54.60      | 42.1      | 70.61 | 89.32      | 7.46     | 99.7       |\n", "category": "Experimental_Exposition"}, {"question": "What are the correlations, differences, pros, and cons between learning with Evol-Instruct and in-context learning with demonstrations?", "answer": "In-context learning can allow Large Language Models (LLMs) to learn a new task with only a few examples given. For tasks like Evol-Instruct that make instructions more complex, which are not common in the training data for LLM, enabling in-context learning is expected to help LLM perform better in evolution. In fact, in the paper's In-Depth Evolving of Complicating Input, the paper uses in-context learning to enhance the evolution effect. In-context learning is a process where the LLM learns a new task based solely on a few given examples, without doing any Back Propagation (BP). Evol-Instruct is aimed at constructing complex instructions to enhance the performance of large pre-trained language model supervised fine-tuning. Therefore, for learning a new task, the advantage of in-context learning is that it requires very few resources, while Evol-Instruct requires more resources but can achieve better results.", "category": "Method_Comparison"}, {"question": "Can the proposed data generation method, originally designed for ChatGPT, be effectively applied to other LLMs using the same experimental design?", "answer": "Yes, the proposed data generation method is effective for other LLMs. The authors tested the method using LlaMA-2-70B-Chat as a substitute for ChatGPT, following the same prompts and design choices.The authors re-executed the entire evolution process starting from Alpaca data and trained the WizardLM-13b(LlaMA-2-70B-Chat) model with randomly selected 70k data. The resulting WizardLM-13b(LlaMA-2-70B-Chat) model significantly outperformed Alpaca-13b trained solely on Alpaca data and even Vicuna-13B trained with much more ChatGPT tokens, despite LlaMA-2-70B-Chat being weaker than ChatGPT. Detailed results are provided in the referenced table.\n\n| Model                                    | Avg.  | MMLU  | ARC    | HellaSwag | TruthfulQA | HumanEval | GSM8k  | AlpacaEval | MT-Bench | WizardEval |\n| :--------------------------------------- | :---- | :---- | :----- | :-------- | :--------- | :-------- | :----- | :--------- | :------- | :--------- |\n| Alpaca-13b                               | 43.44 | 46.63 | 51.20  | 76.31     | 41.62      | 9.2       | 8.35   | 33.25      | 4.78     | 76.6       |\n| Vicuna-13b                               | 54.60 | 50.84 | 51.71  | 79.94     | 52.68      | 12.5      | 24.34  | 70.43      | 6.21     | 86.9       |\n| WizardLM-13b                             | 58.96 | 52.92 | 57.25  | 80.88     | 50.55      | 24.0      | 37.15  | 75.31      | 6.35     | 89.1       |\n| WizardLM-13b(LlaMA-2-70B-Chat Evol) | 56.27 | 51.09 | 57.34  | 79.12     | 48.76      | 19.5      | 33.83  | 70.47      | 6.18     | 84.5       |\n", "category": "Experimental_Analysis"}, {"question": "How was the validity of the difficulty estimation in Figure 5 ensured, given that the instructions were generated by the same LLM used for difficulty assessment?", "answer": "To ensure the validity of the difficulty estimation, the authors sampled 600 instructions (100 per dataset) and used GPT-4 and 5 human annotators for difficulty assessment. The results show a high degree of consistency in difficulty trends across ChatGPT, GPT-4, and manual annotation.\n\n|      | ShareGPT | Alpaca | C1   | C2   | C3   | C4   |\n|-----------|----------|--------|------|------|------|------|\n| GPT-3.5   | 4.63     | 3.0    | 5.48 | 6.35 | 6.84 | 7.08 |\n| GPT-4     | 4.31     | 2.69   | 4.68 | 4.9  | 5.37 | 5.54 |\n| Human     | 4.55     | 3.15   | 5.51 | 5.86 | 6.49 | 6.82 |\n\n", "category": "Experimental_Exposition"}, {"question": "How many annotators were involved in the human evaluation, and what were the exact inter-labeler agreement Kappa scores for the comparisons between WizardLM and other models?", "answer": "Ten well-educated annotators were recruited to judge the whole test set. The Kappa scores for inter-labeler agreement were: WizardLM vs. Alpaca (0.67), WizardLM vs. Vicuna (0.64), and WizardLM vs. ChatGPT (0.63), all indicating good agreement among the annotators.", "category": "Experimental_Setup"}, {"question": "How was the concept of 'difficulty' defined and standardized for the experiments, particularly to address its subjective nature and ensure consistency among labelers?", "answer": "Difficulty was defined by having labelers provide an overall score on a scale of 1 to 10, with higher scores indicating higher difficulty. To standardize understanding, topics were prepared with three related instructions at difficulty levels 2, 5, and 9, along with examples for general, code, and math topics.\n\nExample-1: (General topic)\n| Difficulty | Instruction                                                                                                                               |\n|------------|------------------------------------------------------------------------------------------------------------------------------------------|\n| 2          | What is global warming?                                                                                                                  |\n| 5          | Create a question about the consequences and governance of global warming with English, then answer it with French.                     |\n| 9          | I need detailed data on temperature and carbon dioxide changes in North America over the past decade. Please present it in a markdown format and use Python Matplotlib to plot the trend of both. Finally, based on the phenomena you observe, provide your predictions and suggestions with more than 2000 words. |\n\nExample-2: (Code topic)\n| Difficulty | Instruction                                                                                                                                                                                                                                                               |\n|------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| 2          | Solve with C++: You are given an array prices where prices[i] is the price of a given stock on the i-th day. You want to maximize your profit by choosing a single day to buy one stock and choosing a different day in the future to sell that stock. Return the maximum profit you can achieve from this transaction. If you cannot achieve any profit, return 0. |\n| 5          | Solve with C++: You are given an integer array prices where prices[i] is the price of a given stock on the i-th day. On each day, you may decide to buy and/or sell the stock. You can only hold at most one share of the stock at any time. However, you can buy it then immediately sell it on the same day. Find and return the maximum profit you can achieve. |\n| 9          | Solve with C++: You are given an array prices where prices[i] is the price of a given stock on the i-th day. Find the maximum profit you can achieve. You may complete at most two transactions. Note: You may not engage in multiple transactions simultaneously (i.e., you must sell the stock before you buy again). |\n\nExample-3: (Math topic)\n| Difficulty | Instruction                                                                                                                                                                                             |\n| :--------- | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |\n| 2          | If x - 2 = 0, what is x?                                                                                                                                                                               |\n| 5          | How would you solve for the roots of $x^2 - 4x + 4 = 0$?                                                                                                                                                |\n| 9          |$x = {1+\\frac{\\sqrt{2}}{1+\\frac{\\sqrt{2}}{1+...}}}$ . Find $\\frac{1}{(x+1)(x-2)}$. When your answer is in the form $\\frac{A+\\sqrt{B}}{C}$ , where A, B, and C are integers, and is not divisible by the square of a prime, what is |A| + |B| + |C|?               |\n\n\n\n\n", "category": "Experimental_Setup"}, {"question": "Does the model LlaMA(SNI), fine-tuned with Supernatural Instructions data, surpass Alpaca, Vicuna, and WizardLM on classic QA benchmarks like MMLU?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is Orca included in the performance comparison due to its open-source availability?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the human evaluation involve 10 annotators judging the entire test set?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "v8L0pN6EOi", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Let's Verify Step by Step", "authors": ["Hunter Lightman", "Vineet Kosaraju", "Yuri Burda", "Harrison Edwards", "Bowen Baker", "Teddy Lee", "Jan Leike", "John Schulman", "Ilya Sutskever", "Karl Cobbe"], "abstract": "In recent years, large language models have greatly improved in their ability to perform complex multi-step reasoning. However, even state-of-the-art models still regularly produce logical mistakes. To train more reliable models, we can turn either to outcome supervision, which provides feedback for a final result, or process supervision, which provides feedback for each intermediate reasoning step. Given the importance of training reliable models, and given the high cost of human feedback, it is important to carefully compare the both methods. Recent work has already begun this comparison, but many questions still remain. We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision. To support related research, we also release PRM800K, the complete dataset of 800,000 step-level human feedback labels used to train our best reward model.", "keywords": "LLMs;chain of thought;verifiers;human feedback;reasoning", "qa_pairs": [{"question": "Under the same budget, is it more effective to use step-level rewards on a small set or outcome-level rewards on a large set?", "answer": "Under the same budget, step-level rewards on a small set are more effective as process-level labels are equivalent in cost to outcome-level labels, and Figure 4 illustrates the size of outcome-level labels needed to match the performance of the process-level model.", "category": "Concept_Understanding"}, {"question": "Did the experiments include results on datasets significantly different from examination datasets, such as chemistry and physics?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "d4UiXAHN2W", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention", "authors": ["Renrui Zhang", "Jiaming Han", "Dongyang Liu", "Aojun Zhou", "Pan Lu", "Yu Qiao", "Hongsheng Li", "Peng Gao"], "abstract": "With the rising tide of large language models (LLMs), there has been a growing interest in developing general-purpose instruction-following models, e.g., ChatGPT. To this end, we present LLaMA-Adapter, a lightweight adaption method for efficient instruction tuning of LLaMA. Using 52K self-instruct demonstrations, LLaMA-Adapter only introduces 1.2M learnable parameters upon the frozen LLaMA 7B model, and costs less than one hour for fine-tuning. Specifically, a zero-initialized attention mechanism is proposed. It adopts a learnable zero gating to adaptively inject the instructional cues into LLaMA within self-attention layers, contributing to a stable training process and superior final performance. In this way, LLaMA-Adapter can generate high-quality responses to diverse language instructions, comparable to Alpaca with fully fine-tuned 7B parameters. Besides language commands, by incorporating an image encoder, our approach can be simply extended to a multi-modal LLM for image-conditioned instruction following, which achieves superior multi-modal reasoning capacity on several popular benchmarks (MME, MMBench, LVLM-eHub). Furthermore, we also verify the proposed zero-initialized attention mechanism for fine-tuning other pre-trained models (ViT, RoBERTa, CLIP) on traditional vision and language tasks, demonstrating the effectiveness and generalizability of our approach. Code and models are released at https://github.com/OpenGVLab/LLaMA-Adapter.", "keywords": "Large Language Model (LLM);instruction tuning;multi-modality learning", "qa_pairs": [{"question": "What is the main difference between the LLaMA-Adapter and Alpaca-LoRA in terms of their capabilities and efficiency?", "answer": " It is true that Alpaca-LoRA can also achieve parameter- and time-efficiency (4.2M, 1.5h), even though the paper's LLaMA-Adapter is more efficient (1.2M, 1h). More importantly, the main difference of the paper's approach is the generalization capacity for multi-modal input. LoRA appends trainable rank decomposition matrices into each transformer layer, which entirely inherits the original structures of LLMs. Therefore, LoRA cannot be extended to process new input features. In contrast, the paper's LLaMA-Adapter with adaption prompts and zero-initialized attention is generalizable to accept new input conditions for instruction following, e.g., the plugin with visual reasoning expertise. With the paper's approach, LLaMA can be fine-tuned into a multi-modal LLM with competitive visual understanding capability to existing multi-modal LLMs (shown in Table 3 of the paper), which indicates the paper's promising generalizability for wider application scenarios.", "category": "Method_Comparison"}, {"question": "How does the performance of the LLaMA-Adapter change when scaling up the base model from LLaMA-7B to LLaMA-13B, and does this indicate scalability with base model size?", "answer": "When scaling up the base model from LLaMA-7B to LLaMA-13B, the LLaMA-Adapter shows improved performance on most tasks across three benchmarks. This demonstrates that the approach can be adapted to larger LLMs for improved multi-modal instruction-following capacity.\n\n| LLM Size|MME ||  | MMBench |  |  |  |  |  |  | LVLM-eHub |  |  |  |  |\n|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| LLaMA-Adapter | **All** |P | C | **All** | LR | AR | RR | FP-S | FP-C | CP | **All** | VP | VKA | VR | VC |\n| w LLaMA-7B |1222| 973 | 249 |39.5| 13.1 | 47.4 | 23.0 | 45.0 | 33.2 | 50.6 |0.67| 0.81 | 0.44 | 0.83 | 0.59 |\n| w LLaMA-13B |1299|1021|278|40.1|13.9|47.4|23.7|46.1|34.2|51.2|0.64|0.82|0.39|0.85|0.50|\n", "category": "Experimental_Exposition"}, {"question": "How does the proposed method's performance scale with increasing amounts of instruction data used for tuning?", "answer": "The authors utilize a combination of Alpaca's data (52K) and LLaVA-I (158K) for visual instruction tuning. Progressively adding more question-answering data enlarges the instruction datasets of LLaMA-Adapter: the sampled ***83K VQAv2*** by LLaVA-1.5 [1] and the entire ***204K VQAv2***. As shown in the table below, the increasing instruction tuning data leads to better multi-modal reasoning results on three benchmarks, demonstrating the  paper's method’s favorable scalability to data size.\n\n| Data Size | MME ||  | MMBench |  |  |  |  |  |  | LVLM-eHub |  |  |  |  |\n|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| LLaMA-Adapter | **All** |P | C | **All** | LR | AR | RR | FP-S | FP-C | CP | **All** | VP | VKA | VR | VC |\n| Alpaca+LLaVA-I |1222| 973 | 249 |39.5| 13.1 | 47.4 | 23.0 | 45.0 | 33.2 | 50.6 |0.6675| 0.81 | 0.44 | 0.83 | 0.59 |\n| +VQAv2 (83K) |1256| 1007 | 249 |43.4| 22.9 | 44.7 | 31.3 | 46.7 | 46.9 | 50.3 |0.6925| 0.84 | 0.42 | 0.88 | 0.63 |\n| +VQAv2 (204K) |1618| 1272 | 346 |60.1| 34.7 | 65.3 | 48.7 | 63.1 | 57.3 | 69.3 |0.7175| 0.86 | 0.44 | 0.92 | 0.65 |\n\n[1]Improved Baselines with Visual Instruction Tuning. arXiv 2023.", "category": "Experimental_Exposition"}, {"question": "Was the LLaMA-Adapter method applied to the newer LLM LLaMA-2-7B[1], and what results were achieved on LLaMA-2-7B across various benchmark tests?\n\n[1] Llama 2: Open Foundation and Fine-Tuned Chat Models. arXiv 2023.", "answer": "The authors applied the paper's LLaMA-Adapter, including adaption prompts and zero-initialized attention, to LLaMA-2-7B. LLaMA-2 outperforms LLaMA on various tasks. The authors evaluated the multi-modal variant of the paper's approach on MME, MMBench, and LVLM-eHub benchmarks, achieving higher visual instruction-following results, demonstrating superior generalization capacity for different pre-trained LLMs.\n\n| LLM Model|MME ||  | MMBench |  |  |  |  |  |  | LVLM-eHub |  |  |  |  |\n|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| LLaMA-Adapter | **All** |P | C | **All** | LR | AR | RR | FP-S | FP-C | CP | **All** | VP | VKA | VR | VC |\n| w LLaMA-7B|1222| 973 | 249 | 39.5 | 13.1 | 47.4 | 23.0 | 45.0 | 33.2 | 50.6 | 0.67| 0.81 | 0.44 | 0.83 | 0.59 |\n| w LLaMA-2-7B |1267|968|299|42.1|15.9|48.2|23.6|48.9|35.1|51.6|0.68|0.83|0.43|0.86|0.61|", "category": "Experimental_Exposition"}, {"question": "What additional evaluations were conducted to assess counterfactual reasoning and object hallucination issues in the multi-modal LLaMA-Adapter model?", "answer": "The multi-modal LLaMA-Adapter model was evaluated on two benchmarks: C-VQA[1] for counterfactual reasoning and POPE for object hallucinations. For C-VQA, the model was tested on 2k counterfactual question and answer pairs, achieving the best results in the 'Numerical indirect' category with a score of 34.3 and the least performance loss of 5.6. \n\nMethod|Numerical direct$\\uparrow$ (Loss$\\downarrow$)|Numerical indirect$\\uparrow$ (Loss$\\downarrow$)|Boolean$\\uparrow$ (Loss$\\downarrow$)\n-|:-:|:-:|:-:\nViperGPT|80.6 (19.4$\\downarrow$)|31.6 (68.4$\\downarrow$)|21.6 (72.4$\\downarrow$)\nLLaVA-7B|27.0 (9.9$\\downarrow$)|25.0 (15.2$\\downarrow$)|58.5 (4.8$\\downarrow$)\nLLaVA-13B|24.8 (11.9$\\downarrow$)|20.8 (21.2$\\downarrow$)|56.3 (4.7$\\downarrow$)\nLLaMA-Adapter-7B|30.1 (5.8$\\downarrow$)|34.3 (5.6$\\downarrow$)|45.8 (14.5$\\downarrow$)\n\nFor POPE[2], the model was tested on 500 images from MSCOCO with 6 questions per sample, showing competitive accuracy across different evaluation settings, indicating robustness to object hallucination problems.\n\nMethod|Random|Popular|Adversarial\n-|:-:|:-:|:-:\nmPLUG-Owl-7B|53.30|50.63|50.67\nLLaVA-13B|54.43|52.43|50.77\nMM-GPT-7B|50.03|50.00|50.00\nLLaMA-Adapter-7B|75.47|60.43|60.66\n\n[1]What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models. arXiv 2023.\n[2]Evaluating Object Hallucination in Large Vision-Language Models. arXiv 2023.", "category": "Experimental_Exposition"}, {"question": "Why is it not appropriate to judge the capability of LLaMA-Adapter solely based on BLIP-2's MME[1] performance, and how does LLaMA-Adapter perform on other benchmarks like MMBench[2] and MM-Vet[3]?\n\n[1] Li et al., Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models, ICML 2023\n[2] MMBench: Is Your Multi-modal Model an All-around Player?. arXiv 2023.\n[3] MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities. arXiv 2023.", "answer": "1. **Targeting on Different Capabilities.** The paper's LLaMA-Adapter, along with the concurrent LLaVA and MiniGPT-4, all aim to acquire the ***visual instruction-following capacity***, namely, following human instructions to generate human-like answers conditioned on image input. In contrast, BLIP-2 focuses on the ***fundamental vision-language understanding capability***. Therefore, the former visual instruction models tend to output long and detailed sentences like human conversations of daily life, while the latter BLIP-2 prefers to generate accurate answers with concise language.\n\n2. **Evaluation Bias in MME.** MME mainly evaluates the fundamental perception and cognition abilities of multi-modal LLMs, for which the ground-truth answer contains only one word: \"Yes\" or \"No\". As analyzed above, such short-answer evaluation is more advantageous to BLIP-2 than visual instruction models (LLaMA-Adapter, LLaVA, and MiniGPT-4). For example, MiniGPT-4 is instruction-tuned based on the pre-trained vision encoder and Q-Former of BLIP-2, while performing worse than BLIP-2 in MME, as shown in the table below. In addition, InstructBLIP [4] further conducts instruction tuning upon the pre-trained BLIP-2, and also underperforms BLIP-2 on MME.\n\n| Method | MME |  |  | MMBench |  |  |  |  |  |  | MM-Vet |  |  |  |  |  |  |\n|---|:-:|---|---|:-:|---|---|---|---|---|---|:-:|---|---|---|---|---|---|\n|  | **All** | P | C | **All** | LR | AR | RR | FP-S | FP-C | CP | **All** |Rec | OCR | Know | Gen | Spat | Math |\n| BLIP2 | 1584 | 1294 | 290 | - | - | - | - | - | - | - | 22.4 |27.5 | 11.1 | 11.8 | 7.0 | 16.2 | 5.8 |\n| InstructBLIP | 1505 | 1213 | 292 | 33.9 | 21.6 | 47.4 | 22.5 | 33.0 | 24.4 | 41.1 | 26.2 |32.4 | 14.6 | 16.5 | 18.2 | 18.6 | 7.7 |\n| Mini-GPT4 | 1159 | 867 | 292 | 23.0 | 13.6 | 32.9 | 8.9 | 28.7 | 11.2 | 28.3 | 22.1 |27.5 | 11.1 | 11.8 | 7.0 | 16.2 | 5.8 |\n| LLaVA | 718 | 503 | 215 | 36.2 | 15.9 | 53.6 | 28.6 | 41.8 | 20.0 | 40.4 | 23.8 | 28.0 | 17.1 | 16.3 | 18.9 | 21.2 | 11.5 |\n| LLaMA-Adapter | 1222 | 973 | 249 | 39.5 | 13.1 | 47.4 | 23.0 | 45.0 | 33.2 | 50.6 | 32.8 |38.5 | 20.3 | 31.4 | 33.4 | 22.9 | 3.8 |\n\nTherefore, ***the instruction-following ability of visual instruction models***, e.g., LLaMA-Adapter, ***cannot be entirely reflected by MME scores***. Instead, on other multi-modal benchmarks, e.g., MMBench [2] and MM-Vet [3], the paper's approach exhibits significant advantages over BLIP-2 or BLIP-2-based instruction models (InstructBLIP and MiniGPT-4).\n\n[4] InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning. arXiv 2023.", "category": "Experimental_Analysis"}, {"question": "What quantitative comparisons were conducted between Alpaca, Alpaca-LoRA, and LLaMA-Adapter in terms of instruction-following evaluation, and what were the results?", "answer": "Quantitative comparisons were conducted using the GPT-4 evaluation benchmark[1], which assessed response quality on 80 questions, and the Open LLM benchmark[2], which evaluated generative abilities across four tasks: AI2 Reasoning Challenge[3], HellaSwag[4], MMLU[5], and TruthfulQA[6]. The results, as shown in the provided table, indicate that LLaMA-Adapter achieved the best average performance (52.2) compared to Alpaca (49.2) and Alpaca-LoRA (50.7), demonstrating its strong language instruction-following ability.\n\n | Method | Avg | ARC | HellaSwag | MMLU | TruthfulQA |\n|---|:-:|:-:|:-:|:-:|:-:|\n| Alpaca | 49.2 | 49.1 | 77.7 | 33.8 | 36.3 |\n| Alpaca-LoRA | 50.7 | 53.0 | 77.9 | 37.1 | 34.9 |\n| LLaMA-Adapter | 52.2 | 54.7 | 78.8 | 34.9 | 40.4 |\n\n[1] https://lmsys.org/blog/2023-03-30-vicuna/.\n\n[2] https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/.\n\n[3] Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge. arXiv 2018.\n\n[4] HellaSwag: Can a Machine Really Finish Your Sentence? ACL 2019.\n\n[5] Measuring Massive Multitask Language Understanding. ICLR 2021.\n\n[6] TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL 2022.\n", "category": "Experimental_Exposition"}, {"question": "What is the rationale behind the choice of the number of insertion layers as a hyperparameter, and how does inserting prompts into higher transformer layers compare to earlier layers?", "answer": "Inserting prompts with zero-initialized attention into higher transformer layers is more effective than earlier layers because it allows for better injection of instructional cues into layers that capture higher-level semantics for language generation. There is an optimal number of insertion layers: too few layers may underfit the training data, while too many layers can disrupt the early encoding of input words. An ablation study on ScienceQA shows that inserting prompts into the last 30 layers yields better results (83.85% validation accuracy) compared to the early 30 layers (79.69%) or all 32 layers (81.03%). If limited resources prevent identifying the best insertion layer number, inserting modules into all transformer layers is a generally good solution.\n\n| Layers| Params|Val Acc.|\n|:---|:---:|:---|\n|Last 30|1.79|83.85|\n|Early 30|1.79|79.69|\n|All 32|1.83|81.03|", "category": "Experimental_Analysis"}, {"question": "What happens to the softmax outputs in equation 7 when the gating factor \\(g\\) exceeds 1, and why are additional normalization steps unnecessary?", "answer": "In the actual implementation, a *tanh* activation function is applied to the gating factor \\(g\\), which constrains its scale to the range -1 to 1. This ensures that the impact of adaptation prompts on the pre-trained attention mechanism is well-controlled in both forward and reverse directions, eliminating the need for additional normalization steps.", "category": "Concept_Understanding"}, {"question": "How does the LLaMA-Adapter method, which has few learnable parameters, perform when scaling to larger datasets and a wider range of tasks?", "answer": "The LLaMA-Adapter method demonstrates favorable scalability to larger datasets and a wide range of tasks.\n1. ***More Instruction Data.*** By default, the authors utilize a combination of Alpaca's data (52K) and LLaVA-I (158K) for visual instruction tuning. Progressively adding more question-answering data enlarges the instruction datasets of LLaMA-Adapter: the sampled **83K VQAv2** by LLaVA-1.5 and the entire **204K VQAv2**. As shown in the table below, the increasing instruction tuning data contributes to better multi-modal reasoning results on all benchmarks, demonstrating the paper's method’s favorable scalability to a larger amount of data.\n\n| Data Size | MME ||  | MMBench |  |  |  |  |  |  | LVLM-eHub |  |  |  |  |\n|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| LLaMA-Adapter | **All** |P | C | **All** | LR | AR | RR | FP-S | FP-C | CP | **All** | VP | VKA | VR | VC |\n| Alpaca+LLaVA-I |1222| 973 | 249 |39.5| 13.1 | 47.4 | 23.0 | 45.0 | 33.2 | 50.6 |0.67| 0.81 | 0.44 | 0.83 | 0.59 |\n| +VQAv2 (83K) |1256| 1007 | 249 |43.4| 22.9 | 44.7 | 31.3 | 46.7 | 46.9 | 50.3 |0.69| 0.84 | 0.42 | 0.88 | 0.63 |\n| +VQAv2 (204K) |1618| 1272 | 346 |60.1| 34.7 | 65.3 | 48.7 | 63.1 | 57.3 | 69.3 |0.72| 0.86 | 0.44 | 0.92 | 0.65 |\n\n2. ***More Learning Tasks.*** The paper's LLaMA-Adapter has superior generalization capacity over a wide range of tasks, including ***language instruction following***, e.g., question answering and code generation, ***multi-modal reasoning tasks***, e.g., visual question answering and image captioning, and ***fine-tuning conventional vision or language models***. For example, the specific results on the task of optical character recognition (OCR) on several benchmarks are shown in the table below: DocVQA [1], TextVQA [2], and OCR-VQA [3]. As shown, the paper's LLaMA-Adapter achieves competitive performance compared to other methods. The experiments indicate that, given a small number of learnable parameters, the paper's approach is generalizable to comprehend diverse vision-language tasks.\n\n| Datasets | LLaMA-Adapter | LLaVA | MiniGPT-4 | BLIP2 |\n|---|:---:|:---:|:---:|:---:|\n| DocVQA | 8.13 | 6.26 | 2.65 | 4.75 | \n| TextVQA | 43.76 | 38.92 | 19.40 | 31.98 | \n| OCR-VQA | 38.12 | 23.40 | 16.85 | 38.85 |\n\n[1] DocVQA: A Dataset for VQA on Document Images. WACV 2021.\n[2] Towards VQA Models That Can Read. CVPR 2019.\n[3] OCR-VQA: Visual Question Answering by Reading Text in Images. ICDAR 2019.", "category": "Experimental_Exposition"}, {"question": "Is it possible to use full insertion layers in LLaMA, and what are the advantages and disadvantages of inserting prompts into higher transformer layers versus earlier layers?", "answer": "Yes, it is possible to use full insertion layers in LLaMA. Inserting prompts into higher transformer layers is more effective than earlier layers because higher layers capture higher-level semantics for language generation. There is an optimal insertion layer number; too few layers may underfit the training data, while too many layers may disturb the early encoding of input words. An ablation study on ScienceQA shows that inserting at the last 30 layers yields better results (83.85% validation accuracy) compared to the early 30 layers (79.69%) or all 32 layers (81.03%). If resources are limited, inserting modules into all transformer layers is generally a good solution.\n\n| Layers| Params|Val Acc.|\n|:---|:---:|:---|\n|Last 30|1.79|83.85|\n|Early 30|1.79|79.69|\n|All 32|1.83|81.03|", "category": "Experimental_Exposition"}, {"question": "How does the zero-initialized attention in LLaMA-Adapter differ from the zero-initialized convolution in ControlNet[1] in terms of target task, methodology, and parameter efficiency?\n\n[1]Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023c.", "answer": "The zero-initialized attention in LLaMA-Adapter differs from ControlNet in three key aspects: (1) **Target Task**: ControlNet injects visual conditions into pre-trained diffusion models for text-to-image generation, while LLaMA-Adapter incorporates pre-trained LLMs with new instruction signals to create instruction-following models. (2) **Methodology**: On top of the CNN-based U-Net, ControlNet zero-initializes traditional convolutional layers, and directly adds new conditions with the pre-trained features by residual connections. In contrast, considering the transformer architecture of LLMs, the paper proposes to achieve zero-initialization within the self-attention layer by a learnable gating factor. Then, the paper injects the new instruction cues into LLMs naturally through the attention process, which adaptively controls the interaction between prompts and word tokens(3) **Parameter Efficiency**: ControlNet requires over 300M parameters for training, while LLaMA-Adapter uses only 1.2M parameters for adaption prompts and gating factors (1.8M for the multi-modal variant), making it more parameter-efficient.\n\nTherefore, the zero-initialized attention of LLaMA-Adapter is distinct from ControlNet in both ideas and technical details.", "category": "Method_Comparison"}, {"question": "What are the academic contributions of LLaMA-Adapter that distinguish it from being merely an engineering design?", "answer": "The paper's LLaMA-Adapter is not just an engineering design, but brings new insight to the field of LLM-based instruction tuning. \n\n1. **Progressively injecting instruction cues.** Existing methods do not involve explicit designs to control how much instruction knowledge is injected into LLMs. This might disrupt the pre-trained weights of LLMs, especially at the early training stage. To this end, a learnable gating factor within LLM’s self-attention layers is proposed, which is initialized as zeros to progressively incorporate new instruction cues into LLMs. Such a parameter-efficient design contributes to a more stable training process (***Figure 7 of the paper***), and can be utilized for future instruction-tuning methods.\n\n2. **Generalizability for multi-modal reasoning.** In addition to language instructions, by simply adding multi-scale image features, the paper's LLaMA-Adapter can be extended to a visual instruction model and conduct multi-modal reasoning. In contrast, the previous Alpaca and Alpaca-LoRA only support language instruction-following. Also, the paper reports its downstream performance to fine-tune conventional vision and language models (ViT, RoBERTa, CLIP), demonstrating superior generalization capacity.\n\nTherefore, the paper's LLaMA-Adapter not only verifies the importance of zero-initially injecting instruction cues, but also exhibits good generalizability over a wide range of application scenarios.", "category": "Method_Disambiguation"}, {"question": "What are the results of the ablation study on reducing the rank of LoRA, and how do they compare to LLaMA-Adapter in terms of parameters, training time, and performance on the Open LLM benchmark[1] tasks?\n\n[1] https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/.", "answer": "The ablation study on reducing the rank of LoRA shows results for ranks 2, 4, 8, and 16. Lower ranks reduce learnable parameters from 8.4M to 1.0M and slightly decrease training time from 1.5h to 1.48h. Performance on the Open LLM benchmark[1] tasks (AI2 Reasoning Challenge [2], HellaSwag [3], MMLU [4], and TruthfulQA [5]) is evaluated, with average scores ranging from 50.7 to 50.9 across ranks. LLaMA-Adapter, with 1.2M parameters and 1h training time, achieves the best average result of 52.2, demonstrating a better trade-off between performance and training efficiency compared to Alpaca-LoRA.\n\n| Model | Rank | Param | Time | Avg | ARC | HellaSwag | MMLU | TruthfulQA |\n|---|---|---|---|---|---|---|---|---|\n| Alpaca-LoRA | 2 | 1.0 | 1.48 | 50.9 | 53.6 | 77.9 | 37.9 | 34.0 |\n| Alpaca-LoRA | 4 | 2.1 | 1.49 | 50.8 | 53.5 | 77.8 | 37.5 | 34.4 |\n| Alpaca-LoRA | 8 | 4.2 | 1.49 | 50.7 | 53.2 | 78.1 | 37.1 | 34.5 |\n| Alpaca-LoRA | 16 | 8.4 | 1.5 | 50.8 | 53.0 | 78.0 | 37.1 | 34.9 |\n| LLaMA-Adapter | - | 1.2 | 1 | 52.2 | 54.7 | 78.8 | 34.9 | 40.4 |\n\n[2] Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge. arXiv 2018.\n\n[3] HellaSwag: Can a Machine Really Finish Your Sentence? ACL 2019.\n\n[4] Measuring Massive Multitask Language Understanding. ICLR 2021.\n\n[5] TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL 2022.", "category": "Experimental_Exposition"}, {"question": "Why do the rescaled attention scores in Eq. 7 not represent a probability distribution, and why is the formulation [softmax($S_a)\\cdot g;$ softmax($S_b)\\cdot (1-g)$] not used?", "answer": "The rescaled attention scores in Eq. 7 do not represent a probability distribution because the formulation aims to preserve the pre-trained knowledge of the LLM (softmax($S_b$)) without multiplying any coefficient by it, and adaptively incorporate the new instruction cues, namely, softmax$(S_a)$, ensuring its originally pre-trained probability distribution remains undisturbed. The zero-initialized gating factor $g$ is only multiplied to softmax($S_a$) to adaptively determine how much new information is injected into the LLM, rather than using the proposed formulation [softmax($S_a)\\cdot g;$ softmax($S_b)\\cdot (1-g)$].", "category": "Method_Disambiguation"}, {"question": "What are the effects of inserting adaptation prompts across different transformer layers, and what is the optimal strategy for layer insertion?", "answer": "An ablation study on insertion layer numbers shows that inserting adaptation prompts into the last 30 layers yields better results (83.85% validation accuracy) compared to the early 30 layers (79.69%) or all 32 layers (81.03%). This suggests that injecting instructional cues into layers capturing higher-level semantics is more effective for language generation. There is an optimal insertion number from the higher layers, as too few layers may underfit the training data, while too many layers may disrupt early input encoding. If limited resources prevent identifying the best insertion layer number, inserting modules into all transformer layers is a generally effective solution.\n\n| Layers| Params|Val Acc.|\n|:---|:---:|:---|\n|Last 30|1.79|83.85|\n|Early 30|1.79|79.69|\n|All 32|1.83|81.03|", "category": "Experimental_Exposition"}]}
{"id": "KuPixIqPiq", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Teaching Large Language Models to Self-Debug", "authors": ["Xinyun Chen", "Maxwell Lin", "Nathanael Schärli", "Denny Zhou"], "abstract": "Large language models (LLMs) have achieved impressive performance on code generation. However, for complex programming tasks, generating the correct solution in one go becomes challenging, thus some prior works have designed program repair approaches to improve code generation performance. In this work, we propose self-debugging, which teaches a large language model to debug its predicted program. In particular, we demonstrate that self-debugging can teach the large language model to perform rubber duck debugging; i.e., without any human feedback on the code correctness or error messages, the model is able to identify its mistakes by leveraging code execution and explaining the generated code in natural language. Self-debugging achieves the state-of-the-art performance on several code generation benchmarks, including the Spider dataset for text-to-SQL generation, TransCoder for C++-to-Python translation, and MBPP for text-to-Python generation. On the Spider benchmark where there are no unit tests to verify the correctness of predictions, self-debugging with code explanation consistently improves the baseline by 2-3%, and improves the prediction accuracy on problems of the hardest level by 9%. On TransCoder and MBPP where unit tests are available, self-debugging improves the baseline accuracy by up to 12%. Meanwhile, by leveraging feedback messages and reusing failed predictions, self-debugging notably improves sample efficiency, and can match or outperform baseline models that generate more than 10$\\times$ candidate programs.", "keywords": "large language model;self-debugging", "qa_pairs": [{"question": "What is the significance of the 2-3% improvement on Spider and the improvements on Transcoder and MBPP, as discussed in the paper?", "answer": "The 2-3% improvement on Spider is non-trivial, as achieving a 3% improvement without self-debugging requires 16 samples, as shown in Figure 4 (a). Additionally, self-debugging improves accuracy on the hardest SQL tasks by 9%, as depicted in Figure 4 (b). For Transcoder and MBPP, the improvement with explanation feedback is generally less significant due to the availability of unit tests. However, explanation feedback provides interpretable rationales on how LLMs infer code correctness and conduct fixes, which helps investigate model weaknesses and potentially improves debugging and coding abilities through finetuning.", "category": "Experimental_Analysis"}, {"question": "What is the impact of self-debugging on the computational time and latency in inference, especially for large language models, and how does it compare to baseline methods?", "answer": "Self-debugging requires extra computational time by design compared to the initial code sampling for debugging, as it generates more samples than the greedy decoding baseline (Table 2). This extra computation is also observed in related work like self-refine and Reflexion. However, compared to baselines that generate multiple samples from scratch, self-debugging achieves better performance with lower inference cost. For example, in text-to-SQL generation, the baseline requires 16 samples (16 forward passes) to achieve similar performance as self-debugging (Figure 4a). Similarly, for code translation, self-debugging from greedy decoding achieves comparable performance to the baseline with 10 samples (Appendix E.1). Since self-debugging with one turn brings most of the performance gain (Section 5.2.1), it is more cost-effective than baseline methods without debugging steps.", "category": "Method_Comparison"}, {"question": "Why did the paper not compare Self-Debug with baseline strategies such as Self-Refine in their experiments?", "answer": "The paper's work is concurrent with self-refine. As self-refine did not evaluate on code generation tasks, their prompts are not directly applicable to the paper's benchmarks, and they did not design feedback formats tailored to coding such as code execution and explanations. Specifically, self-refine utilizes GPT-3.5 and GPT-4 to generate feedback to improve its performance on several tasks, where they mainly show performance improvement on text generation tasks that do not require complex reasoning. On the other hand, self-debugging focuses on code generation, and the paper presents a comprehensive study on different feedback formats across different coding benchmarks. In particular, one key difference between self-debugging and other concurrent and followup work on using LLM-generated feedback to improve test-time performance is the paper demonstrates that the self-debugging ability can be taught with few-shot prompting, even if the model is not specially tuned for zero-shot debugging. Specifically, with code-davinci-002 that is not instruction tuned, the paper shows that few-shot prompting achieves similar performance gains as GPT-3.5 and GPT-4, while its self-debugging performance is much better for text-to-SQL generation than GPT-3.5 and GPT-4.", "category": "Experimental_Analysis"}, {"question": "What are the key differences between self-verification and self-debugging in terms of their processes and target tasks?", "answer": "There are two major differences: 1. Self-verification involves a backward verification process to compute verification scores after generating multiple candidate answers, without a subsequent response improvement process. In contrast, self-debugging includes a debugging step after the initial code generation, which can enhance greedy decoding. 2. Self-verification is designed for reasoning tasks such as arithmetic, commonsense, and logical reasoning, with more significant improvements in arithmetic reasoning. Self-debugging, however, focuses on code generation.", "category": "Method_Comparison"}, {"question": "What are the key differences between self-debugging and other methods like Reflexion and Self-Refine, particularly in terms of task focus, access to oracle rewards, and prompt formats?", "answer": "Self-debugging differs from Reflexion and Self-Refine in several ways. Reflexion assumes access to an oracle reward that indicates whether the current state or answer is correct, while Self-Refine uses GPT-3.5 and GPT-4 to generate feedback for improving performance on text generation tasks without complex reasoning. In contrast, self-debugging focuses on code generation tasks where the LLM may not have access to an oracle reward about code correctness, such as on text-to-SQL generation where the unit tests are not available, the LLM needs to determine completely by itself when the debugging process should terminate. Additionally, self-debugging proposes different prompt formats to enable its application across diverse tasks and LLMs, and it demonstrates that the self-debugging ability can be taught with few-shot prompting, even for models not specifically tuned for zero-shot debugging. For example, code-davinci-002, which is not instruction-tuned, shows that self-debugging with few-shot prompting achieves similar or better performance gain than GPT-4.", "category": "Method_Comparison"}, {"question": "Is the 2-3% improvement over the baseline for the Spider benchmark statistically significant, and could it be attributed to accidental changes in the prompt rather than the self-debugging process?", "answer": "The improvement from self-debugging is complementary to the improvement from baseline prompt changes. As supporting evidence, the authors evaluated a variant of full prompt where the authors removed the demonstration of one data row per table (deleting the INSERT INTO statements in Appendix H). With greedy decoding, the accuracy of this baseline prompt is 75.6%, which is 1.9% lower than the paper's full prompt. From this baseline, self-debugging improves the accuracy by 2.1%. Removing INSERT INTO statements weakens both the baseline and the self-debugging prompt, but a consistent improvement is still observed from the self-debugging process. Additionally, the paper has shown results of integrating different feedback information in Section 5, especially Table 2, and conducted extensive evaluation on different wordings for each feedback type without observing significant performance differences that change the relative order when comparing feedback types.", "category": "Experimental_Analysis"}, {"question": "Why do the Codex results for MBPP in Table 2 not correspond to the Codex results for MBPP in Table 1?", "answer": "The Codex results in Table 2 were produced using greedy decoding, which shows that self-debugging improves performance with higher sample efficiency. In contrast, the Codex baseline results in Table 1 were obtained using multiple samples, demonstrating that self-debugging outperforms other baseline methods that leverage multiple samples. ", "category": "Concept_Understanding"}, {"question": "Did the paper experiment with various self-debugging prompts, and what range of performance variability did you observe?", "answer": "Regarding different self-debugging prompts, the paper has shown results of integrating different feedback information in Section 5, especially Table 2. During the development of the paper's approach, the paper also conducted extensive evaluation on different wordings of each feedback type, but no significant performance differences were observed that change the relative order when comparing different feedback types.", "category": "Experimental_Exposition"}, {"question": "What is the role and significance of unit test execution in the self-debugging process, and how does the absence of unit tests impact performance?", "answer": "Unit test execution is important for self-debugging, as its absence results in less significant performance improvements. However, even without unit tests, LLMs can still improve performance through self-generated feedback. For example, code-davinci-002 shows a performance improvement of up to 5%, and GPT-4 improves MBPP accuracy by 3.6% without unit test execution. Additionally, utilizing unit tests whenever applicable is also part of the paper’s self-debugging approach, and the ability to leverage the execution of past generated programs is a key advantage of the paper’s self-debugging approach.", "category": "Method_Mechanics"}, {"question": "How does the self-debugging approach compare to other methods in terms of sample efficiency and the number of tokens generated for solving each problem, particularly in scenarios with and without unit tests?", "answer": "Compared to the sampling cost for obtaining the initial code for debugging, self-debugging requires the LLM to generate more tokens by design. For example, in Table 2, self-debugging generates more samples than the greedy decoding baseline. This extra cost is also needed for other related work, e.g., self-refine and Reflexion. On the other hand, as discussed in Section 5.2.1, the paper demonstrates that to improve the accuracy over greedy decoding, self-debugging achieves better performance with a lower inference cost. To compare the cost in terms of the number of tokens, two settings are considered: (1) when unit tests are not available; e.g., on Spider for text-to-SQL generation; and (2) when unit tests are available; i.e., on Transcoder and MBPP for Python code generation. On the Spider benchmark, with explanation feedback, on average a full debugging turn generates 5.5 times more tokens than simply generating the SQL query. By comparing the performance of self-debugging starting from greedy decoding (80.8% accuracy) to the baseline that samples 16 programs without self-debugging (80.7% accuracy), it can be seen that self-debugging also reduces the number of generated tokens compared to the baseline. On Transcoder and MBPP, due to the availability of unit tests, the improvement with explanation feedback is generally less significant than on Spider. In this case, the unit test feedback is a better option to reduce the sampling budget. On average, a full debugging turn with the unit test feedback generates 2.5 times more tokens than simply generating the initial code, thus self-debugging is again more sample-efficient in terms of the token cost. Furthermore, despite the fact that generating explanation and trace feedback increases the number of generated tokens, one benefit of such feedback is to provide more interpretable rationales on how LLMs infer the code correctness, and how the fixes are conducted. Such natural language feedback helps the paper investigate the models’ weaknesses, and can potentially be used to finetune the model and improve its overall debugging and coding ability.", "category": "Method_Comparison"}, {"question": "What are the key differences between the paper's approach and Reflexion, particularly in terms of task focus, access to oracle rewards, and self-debugging capabilities?", "answer": "Reflexion evaluates its approach on decision-making tasks (AlfWorld) and question-answering tasks (HotPotQA), assuming access to an oracle reward that indicates success or correctness. In contrast, the paper's work focuses on code generation tasks, where the LLM often lacks access to an oracle reward for code correctness, such as on text-to-SQL generation where the unit tests are not available, the LLM needs to determine completely by itself when the debugging process should terminate. Additionally, the paper proposed diverse prompt formats to enable self-debugging across tasks and LLMs, demonstrating that self-debugging can be taught with few-shot prompting, even for models not specifically tuned for zero-shot debugging. For instance, code-davinci-002, which is not instruction-tuned, shows that self-debugging with few-shot prompting achieves similar or better performance gain than GPT-4.", "category": "Method_Comparison"}, {"question": "Can self-debugging effectively identify and fix logic errors in code, such as those where the implementation is correct but does not solve the intended problem?", "answer": "Self-debugging primarily relies on execution accuracy, which considers code correct if it passes unit tests or matches ground truth outputs. In scenarios where only unit tests are used, e.g., for Transcoder experiments where all unit tests are available, programs that pass the unit tests but have other logic errors would not be detected. However, when the LLM is tasked with determining code correctness, e.g., on Spider and MBPP, the LLM can correct such logical errors to some extent. The paper's manual investigation of correct predictions from various benchmarks did not reveal such false positives.", "category": "Experimental_Analysis"}, {"question": "Does the self-debugging framework utilize additional input-output pairs or unit tests during evaluation compared to the baselines, and how does this affect the reported accuracy improvements?", "answer": "The self-debugging framework does not utilize additional input-output pairs or unit tests during evaluation compared to the baselines. For text-to-SQL generation, no unit tests are used for either self-debugging or the baselines, requiring the LLM to infer code correctness independently. Therefore, the LLM also needs to infer the code correctness, since passing the given unit test does not necessarily mean that the predicted code is fully accurate and can pass the hidden unit tests. During the self-debugging process, the hidden unit tests are never used, and the generated code is only executed on the 1 given unit test that already shows up in the problem description for generating the initial code. For code translation, the paper considers the setting where all unit tests are available, since the ground truth execution results can already be obtained by executing the code in the source programming language, which also follows the evaluation setup in prior work.", "category": "Experimental_Setup"}, {"question": "How does the self-debugging approach compare to baselines that generate multiple programs and use code execution to select the final response, particularly in terms of sample efficiency and performance on benchmarks like Spider and Transcoder?", "answer": "The paper compared self-debugging to baselines that generate multiple programs, and uses code execution to select the final response. On the Spider benchmark for text-to-SQL generation, since the unit tests are not available, the baseline takes the majority vote of execution results as the final prediction. As shown in Figure 4 (a), self-debugging from greedy decoding matches the performance of the execution-based baseline that uses 16 samples, and self-debugging consistently improves the sample efficiency. The baseline for Transcoder utilizes all unit tests to determine code correctness, which is equivalent to modifying the baseline to be able to access to the test oracle: when the baseline generates code that fails on test oracle, simply resample a solution. In this case, the baseline would also benefit from oracle, and the resample will be guaranteed to improve the test accuracy (but probably not as much as self-repair approach), as the prediction is considered accurate if any one of the generated samples is correct. Again, the paper shows that self-debugging from greedy decoding matches the baseline that generates 10 samples, and the full analysis is presented in Appendix E.1.", "category": "Method_Comparison"}, {"question": "What percentage of initially correct solutions in text-to-SQL generation are determined as 'incorrect' and proceed to the self-debug phase, and what is the outcome of these solutions after self-debugging?", "answer": "For text-to-SQL generation, Self-debugging only draws 2 samples (1 greedy + 1 self debug output), while outperforming the majority voting based on 16 solutions. Also, LLM indeed might wrongly infer the solution correctness and turn correct solutions into wrong ones.Using Codex, the statistics are as follows: 3.3% of initially correct solutions are wrongly inferred as incorrect and proceed to the self-debug phase. Among these, 1.0% are wrongly changed into incorrect solutions, while 4.3% of initially wrong predictions are fixed into correct ones. The majority of modified solutions remain accurate, indicating that incorrect predictions of code correctness do not significantly degrade performance.", "category": "Experimental_Exposition"}, {"question": "How does the performance of an enhanced baseline, which resamples a second program when the initial greedy decoding produces incorrect code, compare to the existing greedy baseline and self-debugging methods for MBPP and Transcoder tasks?", "answer": "The authors evaluated the enhanced baseline using Codex, and the accuracies are as follows: * Transcoder: greedy decoding = 80.4%, enhanced baseline = 84.5%, self-debugging = 92.5% * MBPP: greedy decoding = 61.4%, enhanced baseline = 65.4%, self-debugging = 70.8%. These results show that self-debugging outperforms the enhanced baseline, which highlights the effect of self-debugging feedback. The design of the baseline with multiple samples in Table 1 in the main paper and Figure 5 (b) in Appendix E.1 is equivalent to the enhanced baseline, except that these results were obtained with more than two trials. For example, Figure 5 (b) shows that when the baseline uses 5 samples (equivalent to the enhanced baseline with 5 trials), the accuracy is 90.9%, which is still worse than self-debugging that reaches 92.5% accuracy.", "category": "Experimental_Exposition"}, {"question": "How does the oracle baseline compare to self-debug feedback in terms of effectiveness, and would it be a better baseline to highlight the effect of self-debug feedback?", "answer": "The oracle baseline results with different LLMs further justify that self-debugging is more effective than completely discarding wrong predictions and sampling another program independently. On MBPP, the oracle baseline has access to all unit tests to determine the code correctness, while self-debugging only has access to 1 unit test. Nevertheless, the oracle baseline still generally performs worse than self-debugging with the simple feedback, let alone other feedback types.", "category": "Method_Comparison"}, {"question": "Was the 2-3% improvement over the baseline for the Spider benchmark shown to be statistically significant?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the paper observe significant performance differences when varying the wordings of feedback types in self-debugging prompts?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Section 4 now include figures (Figure 2 and 3) that make it more self-contained and reduce the need to refer to the appendix?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the authors evaluate the impact of removing demonstration data rows from the prompt on the baseline accuracy for text-to-SQL generation?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "ZG3RaNIsO8", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers", "authors": ["Qingyan Guo", "Rui Wang", "Junliang Guo", "Bei Li", "Kaitao Song", "Xu Tan", "Guoqing Liu", "Jiang Bian", "Yujiu Yang"], "abstract": "Large Language Models (LLMs) excel in various tasks, but they rely on carefully crafted prompts that often demand substantial human effort. To automate this process, in this paper, we propose a novel framework for discrete prompt optimization, called EvoPrompt, which borrows the idea of evolutionary algorithms (EAs) as they exhibit good performance and fast convergence. To enable EAs to work on discrete prompts, which are natural language expressions that need to be coherent and human-readable, we connect LLMs with EAs. This approach allows us to simultaneously leverage the powerful language processing capabilities of LLMs and the efficient optimization performance of EAs. Specifically, abstaining from any gradients or parameters, EvoPrompt starts from a population of prompts and iteratively generates new prompts with LLMs based on the evolutionary operators, improving the population based on the development set. We optimize prompts for both closed- and open-source LLMs including GPT-3.5 and Alpaca, on 31 datasets covering language understanding, generation tasks, as well as BIG-Bench Hard (BBH) tasks. EvoPrompt significantly outperforms human-engineered prompts and existing methods for automatic prompt generation (e.g., up to 25% on BBH). Furthermore, EvoPrompt demonstrates that connecting LLMs with EAs creates synergies, which could inspire further research on the combination of LLMs and conventional algorithms.", "keywords": "large language model;evolutionary algorithm;prompt engineering;natural language processing", "qa_pairs": [{"question": "What is the process for repairing infeasible offspring in GA and DE, and how does this process impact the timing complexity?", "answer": "For both GA and DE, after each iteration including selection, mutation and crossover, updating upon the current population, the infeasible offspring will be eliminated. The elimination is up to their fitness value, i.e., the performance on the given development set in the paper's work. Eliminating infeasible offspring based on certain rules or a small subset of the development set, rather than evaluating them on the entire development set, could potentially reduce costs.", "category": "Method_Mechanics"}, {"question": "What is the ratio of manually-created prompts to LLM-generated prompts in the initial population when Differential Evolution (DE) performs better than Genetic Algorithm (GA), and how does the performance of EvoPrompt change if the initial prompts are entirely generated by LLMs?", "answer": "Before selecting top-k, the ratio of manually written prompts is around 50%. The ratio after top-k and variations are listed in the table below, with no significant correlation observed between the ratio of manual prompts and performance. \n\n|                 | SST-2 | CR    | MR    | SST-5 | AG's News | TREC  | Subj  | ASSET |\n| --------------- | ----- | ----- | ----- | ----- | --------- | ----- | ----- | ----- |\n| Score (GA)      | 95.13 | 91.27 | 90.07 | 49.91 | 72.81     | 64.00 | 70.55 | 46.43 |\n| Score (DE)      | 94.75 | 91.40 | 90.22 | 49.89 | 73.82     | 63.73 | 75.55 | 46.21 |\n| Ratio of manual | 0.37  | 0.50  | 0.20  | 0.23  | 0.47      | 0.00  | 0.30  | 0.33  |\n\nWhen all initial prompts are generated by LLMs, for SST-5, the performance remains similar, indicating little impact of initialization method.\n\n| Method             | SST-5 |\n| ------------------ | ----- |\n| APE                | 46.32 |\n| GA                 | 49.91 |\n| GA (-manual)| 49.98 |\n| DE                 | 49.89 |\n| DE (-manual) | 50.06 |", "category": "Experimental_Exposition"}, {"question": "Why did the paper choose evolutionary algorithms over binary PSO or ant colony optimization for the prompt optimization problem?", "answer": "The authors chose evolutionary algorithms because their mutation and crossover operators, initially designed for gene sequences, are well-suited for natural language prompts. Binary PSO and ant colony optimization were not used because defining direction of motion, velocity, etc., on discrete languages requires further exploration. Additionally, natural language demands high coherence, making alignment with traditional algorithms challenging. It is essential to make appropriate adaptations and adjustments, whether these traditional algorithms are handling continuous or discrete vectors.", "category": "Motivation_Analysis"}, {"question": "What are the key differences in cost analysis between the paper's methods and APE, particularly in terms of overhead and performance?", "answer": "The overhead of the paper's methods mainly comes from prompt evaluation and generation. For evaluation, the paper's overhead is $N * |D| * T$, where $N$ is the size of the population, $|D|$ is the size of the development set, and $T$ is the number of iterations. These parameters differ from the task and can be found in Appendix B.3. For the cost from prompt generation, the cost mainly depends on the number of API results, $T * N$. So the total number of API requests is $N * T * (1 + |D|)$, the same as APE. Moreover, given that the API of LLMs is typically billed based on the number of tokens used, the total number of tokens used in the API requests during the prompt optimization process is also estimated. The overhead is analyzed mainly from two aspects: 1) the performance of the paper's methods compared with APE under the same number of iterations; 2) the performance until convergence measured by the average score on the dev set. The scores on the test set are reported, as shown below. \n\n|   | SST-5 |     |    | Subj |    |    |\n|-------|-------|-------|------------|------|------|------|\n|       | APE   | GA    |  DE       | APE   | GA   | DE  | DE   |\n|-------|-------|-------|------------|------|------|------|\n| Same iteration | | | | | | |\n| # iterations | 9 | 9 | 9 | 15 | 15 | 15 |\n| # tokens | 5.39 M | 5.40 M | 5.52 M | 5.66 M | 5.73 M | 5.93 M |\n| score | 45.79 | 50.23 | 49.23 | 67.20 | 70.10 | 79.35 |\n|-------|-------|-------|------------|------|------|------|\n| Until convergence | | | | | | |\n| # iterations | 9 | 7 | 11 | 15 | 15 | 17 |\n| # tokens | 5.39 M | 4.20 M | 6.75 M | 5.66 M | 5.73 M | 6.72 M |\n| score | 45.79 | 50.23 | 51.13 | 67.20 | 70.10 | 79.35 |\n\nIt can be observed that with the same number of iterations, both GA and DE outperform APE significantly while introducing only a slight overhead in terms of the number of tokens. The convergence rates of APE and GA are similar while DE is slightly slower, but it delivers better performance. This implies a relatively high ceiling for the paper's methods.", "category": "Method_Comparison"}, {"question": "What is the motivation for focusing on prompt engineering in the context of large language models (LLMs), given the concern that prompt engineering may become less relevant as models become more robust to prompts?", "answer": "Prompts for large language models (LLMs) are crucial, as they provide a necessary interface to interact with black-box LLMs. In the era of LLMs, the fine-tuning paradigm is computationally demanding, while the prompt in natural language is effective, interpretable, and human-readable, especially practical for those limited by the computational budget. Previous works find prompt order [1], templates, and examples can highly affect [2] the performance. The field of prompt engineering has recently gained significant popularity, evolving into a highly sought-after career. Automating this process, and achieving superior performance compared to human-designed prompts, signifies a major step forward in the area of language model optimization.Although there is a concern that prompt engineering might not be useful as the language model scales, it has not been proved that the larger the model is, the more robust it is to the given prompt. In the paper's experiments, even GPT3.5 is sensitive to prompts. Lu et al. [1] demonstrate that models of different scales are sensitive to the prompt, with the largest LM being 175 billion. Meanwhile, the paper's proposed EvoPrompt also achieves significant improvement on multiple tasks (e.g., up to 25% improvement on BBH compared with the famous Chain-of-Thought, which shows the influence of different prompts). Moreover, as the model's scale increases, the corresponding rise in computational cost renders fine-tuning methodologies less feasible. This shift amplifies the importance and relevance of prompts,which are known for their effectiveness and better interpretability. One major benefit of employing LLMs as evolutionary operators is their ability to account for the coherence within a sequence. Consequently, the prompt in natural language is a particularly appropriate application. The paper begins with NLP tasks because they come with standard objective evaluation criteria, which are more effective in showcasing the performance of EvoPrompt. It would indeed be intriguing to apply EvoPrompt to multi-modality applications (e.g., game level and image generation) by adjusting the evaluation and selection strategies.\n\n[1] Lu, Yao, et al. \"Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivity.\" Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.\n\n[2] Zhao, Zihao, et al. \"Calibrate before use: Improving few-shot performance of language models.\" International Conference on Machine Learning. PMLR, 2021.", "category": "Motivation_Analysis"}, {"question": "What is the relevance of the LLM-based mutation/crossover operator and the prompt optimization framework to ICLR, given that the main contribution seems more aligned with evolutionary algorithms conferences like GECCO?", "answer": "ICLR considers a broad range of subject areas. One contribution of the paper's work is to connect LLMs with evolutionary algorithms, where the most innovative design is the LLM-based mutation/crossover operator. This connection does not contribute to improving the evolutionary algorithms, instead, it sparks new applications for both of these fields. Accordingly, besides evolutionary algorithms, it is also highly related to LLMs. For example, the works considering evolutionary algorithms with reinforcement learning [2, 3] were also accepted by ICLR’22 and ICLR’23. Another contribution is the prompt optimization framework. While the paper's main focus is on prompts for NLP tasks, potential applications could also include prompts for vision tasks. This is in line with the 'applications in language and vision' in the call for papers, as well as other previously accepted papers at ICLR'23 like APE [1].\n\n[1] Zhou, Yongchao, et al. \"Large Language Models are Human-Level Prompt Engineers.\" The Eleventh International Conference on Learning Representations. 2023.\n\n[2] Wang, Yutong, Ke Xue, and Chao Qian. \"Evolutionary diversity optimization with clustering-based selection for reinforcement learning.\" International Conference on Learning Representations. 2022.\n\n[3] Hao, Jianye, et al. \"ERL-Re \n: Efficient Evolutionary Reinforcement Learning with Shared State Representation and Individual Policy Representation.\" The Eleventh International Conference on Learning Representations. 2023.", "category": "Method_Disambiguation"}, {"question": "What is the computational and financial cost of applying EvoPrompt, and how does it compare to the baseline method APE in terms of cost and performance?", "answer": "The computational and financial cost of applying EvoPrompt is analyzed in terms of prompt evaluation and generation. The evaluation overhead is $N * |D| * T$, where $N$ is the population size, $|D|$ is the development set size, and $T$ is the number of iterations. The cost from prompt generation depends on the number of API results, $T * N$, resulting in a total of $N * T * (1 + |D|)$ API requests, which is the same as APE. Additionally, the total number of tokens used in API requests is estimated. EvoPrompt outperforms APE in performance under the same number of iterations, with only a slight overhead in token usage. GA and DE outperform APE significantly, with similar convergence rates for APE and GA, while DE is slightly slower but delivers better performance. \n\n|   | SST-5 |     |    | Subj |    |    |\n|-------|-------|-------|------------|------|------|------|\n|       | APE   | GA    |  DE       | APE   | GA   | DE  | DE   |\n|-------|-------|-------|------------|------|------|------|\n| Same iteration | | | | | | |\n| # iterations | 9 | 9 | 9 | 15 | 15 | 15 |\n| # tokens | 5.39 M | 5.40 M | 5.52 M | 5.66 M | 5.73 M | 5.93 M |\n| score | 45.79 | 50.23 | 49.23 | 67.20 | 70.10 | 79.35 |\n|-------|-------|-------|------------|------|------|------|\n| Until convergence | | | | | | |\n| # iterations | 9 | 7 | 11 | 15 | 15 | 17 |\n| # tokens | 5.39 M | 4.20 M | 6.75 M | 5.66 M | 5.73 M | 6.72 M |\n| score | 45.79 | 50.23 | 51.13 | 67.20 | 70.10 | 79.35 |", "category": "Method_Comparison"}, {"question": "How does the quality of the initial prompts affect EvoPrompt's performance, and what evidence supports the claim that EvoPrompt is robust even with randomly selected prompts?", "answer": "In Section 5.3, the paper investigates the effect of initial population quality. The investigation shows that both random-k and top-k initialization results are similar and the random setting is more in line with the real situation, indicating EvoPrompt's performance is not heavily reliant on carefully crafted manual prompts. Even randomly selected prompts achieve good performance. Additionally, DE (48.64) and GA (47.80) starting from the bottom population still outperform APE (46.32), demonstrating EvoPrompt's effectiveness and robustness.", "category": "Experimental_Analysis"}, {"question": "How is the diversity of prompts produced by EvoPrompt evaluated, and what are the key findings regarding the diversity of prompts generated by DE compared to GA?", "answer": "The diversity of prompts is evaluated through three aspects: 1) the average length of the prompt and its variance in the population, 2) the number of new words after each iterative step, and 3) a case study. From Figure 9 (a, b) in the Appendix, DE shows higher length and variance than GA, indicating that DE generates more diverse prompts in terms of length. From Figure 9 (c), DE mutates more new words than GA in later iterations, suggesting better potential for exploration and escaping local optima. Overall, DE tends to generate more diverse prompts than GA, indicating that DE is potentially better at exploration.", "category": "Experimental_Exposition"}, {"question": "Does increasing the number of iterations in EA optimization for a single prompt lead to convergence, and what are the effects of more iterations on performance?", "answer": "Increasing the iteration steps results in a slight performance gain on the Subj dataset, but EvoPrompt (both GA and DE versions) converges around 10 steps for most datasets. A consistent value of 10 was chosen across all tasks to maintain uniformity and reduce hyperparameter elaboration. While the average score continues to increase at 10 iterations, there is a trade-off between the number of steps and performance. Further iterations until convergence could improve the best score of DE. Additional details are provided in Appendix C.1.", "category": "Experimental_Exposition"}, {"question": "What are the runtime details for optimizing Alpaca on a single V100 GPU across different datasets?", "answer": "The runtime varies by dataset, with specific runtimes listed in the table for SST-2, CR, MR, SST-5, AG's News, TREC, Subj, ASSET, and BBH. The evaluation runtime depends on iterations, population size, and dev set size. The following table shows the time under default settings. Evaluation time is similar for APE and the paper's methods, and each BBH task takes about the same amount of time.\n\n|                | SST-2 | CR   | MR   | SST-5 | AG's News | TREC | Subj | ASSET | BBH  |\n| -------------- | ----- | ---- | ---- | ----- | --------- | ---- | ---- | ----- | ---- |\n| Runtime (mins) | 63    | 63   | 65   | 90    | 162       | 99   | 194  | 163   | 86   |", "category": "Experimental_Exposition"}, {"question": "Are there any negative cases where EvoPrompt has failed? If so, what are the reasons for the failure, and what are the limitations of the EvoPrompt approach?", "answer": "In the Big-Bench Hard results, there are tasks where EvoPrompt (GA) or EvoPrompt (DE) fails. This occurs because the distribution of the randomly sampled dev set does not align with the test set, leading to a performance gap between evaluations based on them. Although EvoPrompt generates a better prompt based on the dev set, it may not perform as well on the test set. A limitation of this approach is the need for a more effective dev set to enhance performance.", "category": "Experimental_Analysis"}, {"question": "How are EA adapted to work with discrete prompts? Additionally, how are Large Language Models (LLMs) utilized to generate new candidate prompts?", "answer": "LLMs play a role in generating new prompts based on evolutionary operators including mutation and crossover according to the given candidates. Figure 1 and Figure 2 in the paper provide detailed instructions on guiding LLMs to generate a new discrete prompt. Meanwhile, the paper utilizes the selection and updating strategy in evolutionary algorithms to control convergence and effectiveness.", "category": "Method_Mechanics"}, {"question": "Is the stability of the generated prompts by EvoPrompt consistent across different target models, particularly when the target model is a large language model?", "answer": "EvoPrompt demonstrates stability, as evidenced by average performance and standard deviation over three runs on Alpaca. For GPT-3.5, only a single run was conducted due to budget constraints. The instability issue is expected to be mitigated when the target model is also a large language model, as smaller models like RoBERTa may not effectively follow instructions from LLMs.", "category": "Experimental_Analysis"}, {"question": "What is the cost analysis of the paper's methods compared with APE, including the overhead from prompt evaluation and generation, the number of API requests, and the number of tokens used?", "answer": "The cost analysis of the paepr's methods compared with APE is detailed in Appendix C.3. The paper's overhead primarily comes from prompt evaluation and generation. For evaluation, the overhead is calculated as $N * |D| * T$, where $N$ is the population size, $|D|$ is the development set size, and $T$ is the number of iterations. For prompt generation, the cost depends on the number of API results, $T * N$. The total number of API requests is $N * T * (1 + |D|)$, which is the same as APE. Additionally, the authors estimate the total number of tokens used in API requests during prompt optimization. Their analysis compares performance under the same number of iterations and until convergence, showing that GA and DE outperform APE with only a slight overhead in tokens. The convergence rates of APE and GA are similar, while DE is slightly slower but delivers better performance.This implies a relatively high ceiling for the paper's methods.\n\n|                     | SST-5       |         |         | Subj       |         |         |  \n|---------------------|-------------|---------|---------|------------|---------|---------|  \n|                     | APE         | GA      | DE      | APE        | GA      | DE      |  \n|---------------------|-------------|---------|---------|------------|---------|---------|  \n| **Same iteration**  |             |         |         |            |         |         |  \n| # iterations        | 9           | 9       | 9       | 15         | 15      | 15      |  \n| # tokens            | 5.39 M      | 5.40 M  | 5.52 M  | 5.66 M     | 5.73 M  | 5.93 M  |  \n| score               | 45.79       | 50.23   | 49.23   | 67.20      | 70.10   | 79.35   |  \n|---------------------|-------------|---------|---------|------------|---------|---------|  \n| **Until convergence** |           |         |         |            |         |         |  \n| # iterations        | 9           | 7       | 11      | 15         | 15      | 17      |  \n| # tokens            | 5.39 M      | 4.20 M  | 6.75 M  | 5.66 M     | 5.73 M  | 6.72 M  |  \n| score               | 45.79       | 50.23   | 51.13   | 67.20      | 70.10   | 79.35   |", "category": "Method_Comparison"}]}
{"id": "EnXJfQqy0K", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Building Cooperative Embodied Agents Modularly with Large Language Models", "authors": ["Hongxin Zhang", "Weihua Du", "Jiaming Shan", "Qinhong Zhou", "Yilun Du", "Joshua B. Tenenbaum", "Tianmin Shu", "Chuang Gan"], "abstract": "In this work, we address challenging multi-agent cooperation problems with decentralized control, raw sensory observations, costly communication, and multi-objective tasks instantiated in various embodied environments. While previous research either presupposes a cost-free communication channel or relies on a centralized controller with shared observations, we harness the commonsense knowledge, reasoning ability, language comprehension, and text generation prowess of LLMs and seamlessly incorporate them into a cognitive-inspired modular framework that integrates with perception, memory, and execution. Thus building a Cooperative Embodied Language Agent CoELA, who can plan, communicate, and cooperate with others to accomplish long-horizon tasks efficiently. Our experiments on C-WAH and TDW-MAT demonstrate that CoELA driven by GPT-4 can surpass strong planning-based methods and exhibit emergent effective communication. Though current Open LMs like LLAMA-2 still underperform, we fine-tune a CoELA with data collected with our agents and show how they can achieve promising performance. We also conducted a user study for human-agent interaction and discovered that CoELA communicating in natural language can earn more trust and cooperate more effectively with humans. Our research underscores the potential of LLMs for future research in multi-agent cooperation. Videos can be found on the project website https://vis-www.cs.umass.edu/Co-LLM-Agents/.", "keywords": "Large Language Models;Embodied Intelligence;Multi-Agent Cooperation;Human-AI Interaction;Communication", "qa_pairs": [{"question": "How is the communication cost evaluated, and is there a tradeoff study between communication cost and effectiveness?", "answer": "The communication cost is evaluated using the metrics *Average Steps* and *Transport Rate*. Communication in this setting takes time, impacting the timesteps needed to complete a task, as reflected in these metrics. As shown in B.1 and B.2 Action Space, the agent can send at most 500 characters per frame, meaning a message of 600 characters costs 2 frames. The agents must reason over the pros and cons of each communication step. An extreme tradeoff case is when communication has no cost, allowing agents to share all observations and discuss plans thoroughly before taking the next action, leading to less decentralized collaboration and better efficiency in actions.", "category": "Method_Mechanics"}, {"question": "What is the system's performance in scenarios with an increased number of objects and complex interactions, and what measures ensure its scalability?", "answer": "The system performs well in complex scenarios, as demonstrated in multi-room scenes with 18 types of objects and containers of interest and 6-7 actions to interact with. The prompt uses an average of 453 tokens and a maximum of 591 tokens, which is economical due to the paper's deliberately designed memory module to store and organize the objects of interest effectively. For extreme scenarios, the system can integrate additional filter modules, such as those described in [1], to enhance scalability, leveraging its modular design.\n\n[1] Plan, Eliminate, and Track — Language Models are Good Teachers for Embodied Agents", "category": "Experimental_Exposition"}, {"question": "What happens if the object of interest is on the back of the agent when turning left/right, and how is the visual-related textual input organized? Did the work use oracle information of the whole environment, including objects and relations?", "answer": "The paper presents results from two scenarios: Oracle Perception, where instance segmentation is provided by the environment, and Learned Perception, where only an ego-centric RGBD image from the agent’s camera is given, requiring a neural perception model to segment the objects. Under both scenarios, there exists the partial observation challenge, if the object of interest is out of the agent's view (e.g., on the back of the agent), the agent must actively explore the environment to find it, leading to a navigation and exploration challenge. Partial observation is a key feature in the paper's work, making effective cooperation more challenging and leading to the paper's key insights. The work does not assume the oracle information of the whole environment.", "category": "Method_Disambiguation"}, {"question": "Can the framework of the paper exhibit effective communication behavior in challenging environments?", "answer": "The paper's designed framework can show effective communication behaviors under the challenging setting, especially when cooperating with humans. This setting also leads to many interesting scenarios, such as where agents fail to reach a consensus. The authors have analyzed agent behaviors and found that CoELA tends to cooperate and coordinate plans efficiently without arguing back and forth which can be credited to LLMs trained to follow instructions and trust their cooperators. Additional details on cooperation mechanisms and a discussion on CoELA's communication behavior are provided in Appendices C.4 and D.", "category": "Method_Disambiguation"}, {"question": "How does the method address the limitation of requiring a manually defined skill library for specific domains, and how does it ensure applicability in domains without predefined skill libraries?", "answer": "The low-level skills in the method are derived from motion planning, which is widely used in embodied AI benchmarks[1][2][3][4]. The framework's modular design allows flexibility and adaptability to various low-level control policies, such as reinforcement learning and trajectory optimization, making it applicable across domains without predefined skill libraries and offering an advantage over end-to-end methodologies.\n\n[1] Retrospectives on the embodied ai workshop.\n\n[2] Do As I Can, Not As I Say: Grounding Language in Robotic Affordances. CoRL22\n\n[3] Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR21\n\n[4] The threedworld transport challenge: A visually guided task-and-motion planning benchmark towards physically realistic embodied ai. ICRA22", "category": "Method_Disambiguation"}, {"question": "What evidence is provided to demonstrate the necessity and effectiveness of each module in the CoELA framework?", "answer": "An ablation study on the memory, communication, planning, and execution modules shows that all modules contribute to CoELA's best performance. Without the Memory Module, CoELA requires nearly double the steps to complete tasks. A strong LLM is crucial for the Planning and Communication modules' efficiency. Without the Execution Module, the agent struggles to make low-level control directly and fails most of the task. It's not trivial to ablate on the perception module under the visual image observation setting, so the paper also conducts experiments with the oracle perception given, and the result is shown in Table 1. The performance gap between TDW-MAT and w/ Oracle Perception is rather close, showing the Perception Model is effective as well. On the other hand, the paper's modular framework is not restricted to the number of modules and can be easily extended to deal with all kinds of challenging tasks.", "category": "Experimental_Exposition"}, {"question": "How does the experimental design address concerns about its breadth, particularly in terms of the scenarios and tasks considered?", "answer": "The experimental design includes comprehensive experiments on two challenging embodied multi-agent benchmarks, conducted across two simulation platforms with different tasks and varied observation and action spaces. C-WAH involves 5 activities, and TDW_MAT features 12 scenes. Experiments cover scenarios of collaborating with both AI agents and humans. The rearrangement task, a canonical task for evaluating embodied agents, includes challenges like perception, navigation, and interaction[1]. \n\n[1]Rearrangement: A Challenge for Embodied AI", "category": "Experimental_Setup"}, {"question": "What are the design principles and methodologies used for the execution module, memory module, and perception module, and are they based on language-based approaches?", "answer": "The execution module, memory module, and perception module are designed to mimic human modular cognitive architectures and are not based on language-based approaches.  The Perception Module utilizes a Mask-RCNN to obtain the instance segmentation mask from an ego-centric RGB image observation, and combines it with the depth image and the agent’s position to project each pixel into the 3D world coordinate to obtain a 3D voxel semantic map. The Execution Module is based on an A-star-based planner to find the best path for navigation and robustly interact with the objects through primitive actions according to environmental rules, which is a common design for hierarchical planning in the field[1, 2]. \n\n[1] Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR21\n\n[2] The threedworld transport challenge: A visually guided task-and-motion planning benchmark towards physically realistic embodied ai. ICRA22", "category": "Method_Mechanics"}, {"question": "What are the core contributions and innovations of CoELA compared to previous works using LLMs for implementing embodied agents, and how does CoELA address the challenges of multi-agent cooperation?", "answer": "CoELA addresses the challenges of multi-agent cooperation, such as decentralized control, complex observations, costly communication, and long-horizon multi-objective tasks, by incorporating LLMs into a modular framework that enables direct communication between agents. This approach is both effective and interpretable, distinguishing it from previous works that focused on single-agent and simple tasks. ", "category": "Method_Comparison"}, {"question": "What form do the procedures stored in CoELA's Memory Module take, and how are they obtained for a specific environment?", "answer": "The framework is deliberately designed to incorporate LLMs into modular frameworks mimicking human cognitive architecture to deal with challenging decentralized multi-agent cooperation scenarios. Experiments on two challenging embodied AI benchmarks show the paper's modular framework is effective and can be adapted to different observation and action spaces. The ablation study and in-depth analysis in the paper show all modules contribute to the best performance of CoELA when cooperating with other agents and humans. Detailed information about the execution module and memory module is provided in Appendix A.2 and Appendix A.5.", "category": "Method_Disambiguation"}, {"question": "How does CoELA address the potential challenges of manually designing high-level plans and the associated increase in prompt length, particularly in complex environments?", "answer": "The paper's evaluated scenarios are complex as there are 18 types of objects and containers of interest in the multi-room scenes, and 6-7 actions to interact with. Across the experiments presented in the paper, the prompt utilizes an average of 453 tokens and a maximum of 591 tokens, which is quite economical. This efficiency is attributed to the deliberately designed memory module that effectively stores and organizes the objects of interest. Considering even more complex scenarios with too many objects of interest, the paper's framework can easily integrate some more deliberately designed filter modules for better scalability, thanks to the modular design of the paper's system.", "category": "Method_Mechanics"}, {"question": "Why were the baselines in the experiments relatively simple and heuristic, and how does the framework's modular design allow for adaptation to more complex baselines, such as those designed for embodied agents?", "answer": "The baselines were relatively simple and heuristic because the focus of the study was on high-level reasoning and planning, with low-level skills obtained from motion planning, which is common in embodied AI benchmarks. The framework's modular design allows for flexibility and adaptation to various low-level control policies, including reinforcement learning and trajectory optimization, offering a distinct advantage over end-to-end methodologies. ", "category": "Motivation_Analysis"}, {"question": "What are the mechanisms for cooperation between CoELA and other agents, specifically how CoELA cooperates with MHP agents?", "answer": "Both CoELA and MHP agents take the same environment observation as input and output environment-acceptable primitive actions. CoELA can model the MHP agent's state and plan cooperatively despite different inner working mechanisms. The cooperation follows the MHP design in [1], where MHP infers the other's intention and adapts its subgoal based on the observation of the other agent.\n\n[1] Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR21", "category": "Method_Disambiguation"}, {"question": "How does the paper address the potential challenges of agents failing to reach a consensus through communication in the Discrete Execution setting using Large Language Models (LLMs)?", "answer": "Communication does not guarantee consensus and can lead to inefficiencies due to prolonged arguments. However, the experiments with CoELA in the paper show that agents tend to cooperate and coordinate plans without arguing, likely because LLMs are trained to follow instructions and trust their cooperators. An example in Figure 3a and Section C.4 shows Bob initially suggesting a division of labor, Alice disagreeing and proposing an alternative plan, and Bob accepting the new plan after consideration. While this behavior enhances cooperation, it may reduce efficiency if the cooperator is malicious. ", "category": "Method_Mechanics"}, {"question": "How does the traditional MHP cooperate with MHP or CoELA?", "answer": "Both agents receive the same environment observations and produce environment-acceptable actions. Although their internal mechanisms differ, CoELA can model the MHP agent's state and plan collaboratively. The traditional MHP, as per [1], infers the other agent's intention and adjusts its subgoal based on the observed actions of the other agent.\n\n[1] Watch-and-help: A challenge for social perception and human-ai collaboration. ICLR21", "category": "Method_Mechanics"}, {"question": "How does CoELA manage situations where no consensus is reached between agents, such as when Alice wants Bob to go to location A and Bob wants Alice to go to location B?", "answer": "CoELA tends to avoid prolonged arguments by coordinating plans efficiently, attributed to LLMs trained to follow instructions and trust cooperators. In cases like the one described, CoELA agents adapt their plans considering others' suggestions, as illustrated in Figure 3a and Section C.4. For instance, Bob initially suggests a division of labor, but upon Alice's disagreement and proposal of a new plan, Bob accepts it. While this fosters cooperation, it may reduce efficiency if the cooperator is malicious. ", "category": "Method_Mechanics"}, {"question": "Did the experimental design include only one scenario?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "MLBdiWu4Fw", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "InternVid: A Large-scale Video-Text Dataset for Multimodal Understanding and Generation", "authors": ["Yi Wang", "Yinan He", "Yizhuo Li", "Kunchang Li", "Jiashuo Yu", "Xin Ma", "Xinhao Li", "Guo Chen", "Xinyuan Chen", "Yaohui Wang", "Ping Luo", "Ziwei Liu", "Yali Wang", "Limin Wang", "Yu Qiao"], "abstract": "This paper introduces InternVid, a large-scale video-centric multimodal dataset that enables learning powerful and transferable video-text representations for multimodal understanding and generation. InternVid contains over 7 million videos lasting nearly 760K hours, yielding 234M video clips accompanied by detailed descriptions of total 4.1B words. Our core contribution is to develop a scalable approach to autonomously build a high-quality video-text dataset with large language models (LLM), thereby showcasing its efficacy in learning video-language representation at scale. Specifically, we utilize a multi-scale approach to generate video-related descriptions. Furthermore, we introduce ViCLIP, a video-text representation learning model based on ViT-L. Learned on InternVid via contrastive learning, this model demonstrates leading zero-shot action recognition and competitive video retrieval performance. Beyond basic video understanding tasks like recognition and retrieval, our dataset and model have broad applications. They are particularly beneficial for generating interleaved video-text data for learning a video-centric dialogue system, advancing video-to-text and text-to-video generation research. These proposed resources provide a tool for researchers and practitioners interested in multimodal video understanding and generation.", "keywords": "video-language dataset;video understanding;video generation;multimodal understanding;action recognition;video retrieval", "qa_pairs": [{"question": "What is the potential impact of the language diversity in the InternVID dataset on the generated captions and video distributions, and how was this hypothesis tested?", "answer": "Unlike previous models that are benchmarked mostly on English-based datasets, InternVID encompasses clips from a variety of languages. This necessitates a further analysis to determine the potential impact this diversity might have. Currently, the authors hypothesize that the language of the video may not significantly impact the generated captions as the paper's deployed image caption models generate English descriptions based purely on input frames. However, in terms of video distributions, there may exist differences (such as in behaviors, activities, and events) between videos stemming from different countries due to varied cultural backgrounds. To examine this hypothesis, the authors selected two 2 million subsets of InternVid: one consisting of only English videos (InternVid-2M-EN) and another with only Chinese videos (InternVid-2M-CN). Note that captions are generated in English regardless of whether the source videos are in English or Chinese. The paper's ViCLIP-B model was pretrained on these subsets, and the authors conducted zero-shot experiments as described below. Due to resource constraints, ViCLIP-B was trained with a batch size of 4096 using 8 A100 GPUs with a mask ratio set to 0.9. The remaining training settings were kept consistent with those outlined in the paper. It was found that the model pretrained with InternVid-2M-EN notably outperformed that with InternVid-2M-CN in both zero-shot action recognition on K400/600/700 and video retrieval. This result can be attributed to the fact that InternVid-2M-EN has a data distribution much closer to downstream task data than InternVid-2M-CN, as all used task videos are sourced from English sources.\n\n\n### Table B. Zero-shot action recognition results of ViCLIP using different pretraining sources on Kinetics 400/600/700.\n| Method       | Training Data     | K400 top-1 | K400 AVG | K600 top-1 | K600 AVG | K700 top-1 | K700 AVG |\n|--------------|-------------------|------------|----------|------------|----------|------------|----------|\n| ViCLIP       | +InternVid-2M-EN  | 40.20      | 53.68    | 37.40      | 51.00    | 29.60      | 42.06    |\n| ViCLIP       | +InternVid-2M-CN  | 35.9       | 49.73    | 33.70      | 47.05    | 26.90      | 39.02    |\n### Table C. Results of zero-shot video retrieval of ViCLIP using different pretraining sources on MSR-VTT, LSMDC, DiDeMo, MSVD, and ActivityNet.\n| Method       | Data              | MSR-VTT T2V | MSR-VTT V2T | LSMDC T2V | LSMDC V2T | DiDeMo T2V | DiDeMo V2T | MSVD T2V | MSVD V2T | ANet T2V | ANet V2T |\n|--------------|-------------------|-------------|-------------|-----------|-----------|------------|------------|----------|----------|----------|----------|\n| ViCLIP       | +InternVid-2M-EN  | 24.1        | 24.1        | 9.8       | 9.1       | 10.3       | 15.6       | 31.4     | 50.5     | 7.1      | 12.0     |\n| ViCLIP       | +InternVid-2M-CN  | 22.2        | 22.2        | 8.4       | 8.7       | 9.9        | 15.1       | 29.6     | 48.5     | 5.7      | 9.7      |\n\n", "category": "Experimental_Exposition"}, {"question": "How does the performance of the paper's model (ViCLIP-B) compare to CLIP-ViP-B[1] on the MSR-VTT dataset?\n\n[1]CLIP-ViP: Adapting Pre-trained Image-Text Model to Video-Language Representation Alignment, ICLR 2023.", "answer": "As seen in Table A, when evaluated on MSR-VTT, the paper's model (ViCLIP-B used with InternVid-200M/-10M-FLT) outperforms CLIP-ViP-B used with HD-VILA-100M, giving a better R@1 score (50.7%/49.0% vs. 47.7%). However, when captions and subtitles from HD-VILA-100M are combined, CLIP-ViP-B achieves comparable result of 49.6%. An important note is that CLIP-ViP-B uses more frames for pretraining than the paper (12 vs. 8). \n\nThis suggests that synthetic video captions from the paper's InternVid dataset can compete effectively against HD-VILA-100M's combined subtitles and captions.\n\n- Table A. Fine-tuned video retrieval on MSR-VTT.\n\n|Method | Data | #Frames in Train | T2V | V2T |\n|:---:|:---:|:---:|:---:|:---:|\n|CLIP-ViP-B | +HD-VILA-100M | 12 | 47.7 |  |\n|CLIP-ViP-B | +HD-VILA-100M(sub+cap) | 12 | 49.6 |  |\n|CLIP-ViP-B | +HD-VILA-100M(sub)+Im-text data | 12 | 49.1 |  |\n|ViCLIP-B | +InternVid-200M | 8 | 50.7 | 49.4 |\n|ViCLIP-B | +InternVid-10M-FLT | 8 | 49.0 | 49.2 |", "category": "Method_Comparison"}, {"question": "Why does ViCLIP trained on InternVid-10M-FLT/DIV show better performance in zero-shot action recognition (Table 2) but poorer performance in the fine-tuned setting (Table 3)? And how does the large training data size affect the likelihood of the same video clip appearing in the same batch?", "answer": "The contrasting performance of ViCLIP in zero-shot vs. fine-tuned action recognition is due to differing task demands. Zero-shot performance relies on aligning video and text representations, while fine-tuned action recognition focuses on the discriminative abilities of the video representation. As such, there is no guaranteed correlation between a model's capabilities in these two tasks.\n\nViCLIP trained on InternVid-200M has a more robust video representation due to a higher quantity and diversity of videos. Conversely, the version trained on InternVid-10M-FLT/DIV likely excels in video-text alignment: a result of reducing the number of clips sourced from identical videos. This approach makes their training data more distinct than InternVid-200M’s, which promotes contrastive learning to enhance video-text representation—yielding superior outcomes for zero-shot action recognition and video-text tasks such as retrieval. The same video clip appearing in the same batch is relatively rare given the extensive size of the training data. However, the higher video diversity in InternVid-10M-FLT/DIV and its improved video-text correlation (sorted by CLIP) result in a more effective video-text representation compared to InternVid-10M, which explains the false negatives.", "category": "Experimental_Analysis"}, {"question": "What is the strategy for mixing coarse and fine-level captions in InternVid-200M, and what motivated the inclusion of coarse-level captions? Additionally, why was an ablation study not conducted at a larger scale (e.g., 10M)?", "answer": "The mixing strategy for coarse and fine-level captions involves summarizing framewise captions at 1 fps into a single video caption, where the central frame is processed by BLIP2 and the remaining frames by tag2text. The inclusion of coarse-level captions was motivated by empirical evidence from the BLIP2 paper, which showed superior performance in image captioning benchmarks, suggesting potential benefits for video-text contrastive learning. An ablation study was conducted on smaller subsets (InternVid-2M and InternVid-2M-BLIP) due to computational constraints. Fine-level captions were not included in experiments because they are long and repetitive, making them unsuitable for contrastive learning. The results from Tables B and C demonstrate that combining coarse and fine-level captions yields better zero-shot performance than using coarse-level captions alone.", "category": "Motivation_Analysis"}, {"question": "What evidence supports the effectiveness of combining coarse and fine-level captions in video-text contrastive learning?", "answer": "An ablation study was conducted on smaller subsets (InternVid-2M and InternVid-2M-BLIP) due to computational constraints, each having 2 million video-text pairs. InternVid-2M utilized fused captions, combining both coarse- and fine-level ones. In contrast, InternVid-2M-BLIP only used the coarse-level captions produced by BLIP2 on the central frames.\n\nZero-shot experiments were conducted on these models. Due to computational constraints, the ViCLIP-B was trained with a batch size of 4096 using 8 A100 GPUs, with a mask ratio set to 0.9. All remaining training parameters were consistent with those in the main paper.\n\nContrasting the results from Table B and C, it's evident that the use of combined coarse and fine-level captions in video-text contrastive learning rendered superior zero-shot performance than utilizing the coarse level ones alone. It shows the effectiveness of the paper's given video captioning pipeline.\n\n### Table B. Zero-shot action recognition results of ViCLIP using different captions on Kinetics 400/600/700.\n|Method | Training Data | K400 |  | K600 |  | K700 |  |\n|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| | |top-1 | AVG | top-1 | AVG | top-1 | AVG |\n|ViCLIP | +InternVid-2M | 51.70 | 64.69 | 49.20 | 62.34 | 40.90 | 53.70 |\n|ViCLIP | +InternVid-2M-BLIP2 | 38.40 | 51.58 | 36.40 | 49.19 | 29.10 | 40.68 |\n\n### Table C. Results of zero-shot video retrieval from ViCLIP using different captions on MSR-VTT, LSMDC, DiDeMo, MSVD, and ActivityNet.\n|Method | Data | MSR-VTT | | LSMDC | | DiDeMo | | MSVD | | ANet | |\n|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|\n| | |T2V | V2T | T2V | V2T | T2V | V2T | T2V | V2T | T2V | V2T |\n|ViCLIP | +InternVid-2M | 31.8 | 33.7 | 14.3 | 12.7 | 13.6 | 21.5 | 39.6 | 62.5 | 9.9 | 16.8 |\n|ViCLIP | +InternVid-2M-BLIP2 | 21.7 | 21.9 | 5.2 | 5.1 | 7.1 | 12.7 | 24.6 | 42.1 | 6.4 | 10.2 |", "category": "Experimental_Exposition"}, {"question": "What are the examples from the dataset that show the contrast between coarse-level and fine-level captions?", "answer": "The authors give two randomly picked examples to illustrate the difference between coarse and fine-level captions below. These examples highlight how coarse-level captions from BLIP2 provide a more detailed and comprehensive description of the scene or activity, as compared to fine-level framewise captions from tag2text. For instance, the BLIP2 caption in the second example identifies gender ('woman') and background context ('on a wooden floor'), which are details not included in the tag2text captions.\n\n- Case 1\n  - Fine-level framewise captions from tag2text: 'a screenshot from a game of a pink van on the road at night', 'a screenshot from a game of a pink van on the road at night', 'a screenshot from a game of a pink van on the road at night', 'a pink van on a small street at night', 'a small truck on the road at night in the game', 'a small van on the street at night', 'a small truck on the road at night in the game', 'a screenshot from a game of a pink van on the road at night'\n  - Coarse-level caption from blip2: 'a small truck driving down the street in an online game'\n- Case 2:\n  - Fine-level framewise captions from tag2text: 'a bowl of ice and a bowl of water', 'a pair of hands mixing food in a bowl', 'a pair of hands mixing food in a bowl', 'a pair of hands mixing food in a bowl', 'a person mixing a bowl of food with ice and water', 'a person mixing food in a bowl of water', 'a person mixing food in a bowl of water', 'a person mixing food in a bowl of ice and water'\n  - Coarse-level caption from blip2: 'woman making food on a wooden floor with bowls'", "category": "Method_Disambiguation"}, {"question": "Were the T5-summary and Vicuna models fine-tuned specifically for processing captions, and why was the T5-summary model chosen over models like GPT-3.5-Turbo or GPT-4, which can perform zero-shot caption processing?", "answer": "The [T5-summary]and [Vicuna] models the paper used have been fine-tuned on public NLP datasets for summary and dialogue performance respectively, but not specifically on the paper's dataset for processing captions.\n\nIt's true that larger language models like GPT-3.5-Turbo or GPT-4 could potentially improve the performance of caption handling due to their advanced capacity in capturing semantic meaning from text. However, these models also come with higher computational cost. Considering this trade-off between performance and efficiency, the paper elected to use a T5 summary model as it offers acceptable performance at a reasonable cost.\n\nBelow are some qualitative comparisons between GPT-3.5-Turbo, GPT-4, and T5-summary illustrating their summary capabilities. These examples demonstrate that while the three models offer similar qualitative outcomes, choosing the best one would depend on specific requirements and constraints.\n\n- Case1:\n  - Framewise Captions: 'a woman with glasses talking to a man on the news', 'a woman with glasses talking to a man on the news', 'a woman with glasses talking to a man in an interview', 'a woman with glasses talking to a man in a video', 'a woman with glasses is talking to people in an interview', 'a woman with glasses is talking to a man', 'a woman with glasses talking to a man on the news', 'a woman with glasses talking to a man on the news'\n  - T5-Summary: 'Woman with glasses talking to a man on the news.'\n  - GPT-3.5-Turbo: 'A woman with glasses is speaking to a man in an interview or on the news.'\n  - GPT-4: 'A woman with glasses is frequently seen talking to a man in interviews, videos, and on the news.'\n\n- Case2:\n  - Framewise Captions: 'a screenshot of a ship in a large body of water', 'a large ship in a large body of water', 'a screenshot of a ship in a large body of water', 'a large ship in a large body of water', 'a screen shot of a boat in a large body of water', 'a screen shot of a boat in a large body of water', 'a screen shot of a boat in a large body of water', 'a screen shot of a boat in a large body of water'\n  - T5-Summary: 'a screenshot of a boat in a large body of water'\n  - GPT-3.5-Turbo: 'a screenshot of a large ship or boat in a vast body of water'\n  - GPT-4: 'a large ship or boat in  a large body of water'", "category": "Motivation_Analysis"}, {"question": "What is the justification for the choice of a 0.9 masking ratio in ViCLIP, and how does it balance efficiency gains and performance drops?", "answer": "The primary aim of applying ViCLIP on InternVid is to validate the efficacy of the paper's proposed dataset in video understanding tasks. The architecture and training methodology for ViCLIP are not novel; it follows the design principles of CLIP, with the addition of mask modeling for training efficiency. The masking scheme, specifically the choice of a 0.9 masking ratio from VideoMAE, was adopted to balance computational efficiency and performance. The Table A, which lists the trade-off between efficiency gains and performance drops at different masking ratios for ViCLIP-B. As can be seen, decreasing the masking ratio improves performance but increases GPU memory usage significantly. The masking ratio of 0.9 was chosen because it provides the highest efficiency with a tolerable performance drop on downstream tasks.\n\n- Table A. Fine-tuned video retrieval on MSR-VTT from ViCLIP with different masking ratios. ``Mem\" denotes single GPU memory usage with per GPU batch size 128.\n\n|Method | Data | Masking ratio | T2V | Mem/G|\n|:---:|:---:|:---:|:---:|:---:|\n|ViCLIP-B | +InternVid-10M | 0.7 | 48.0 | 27.9 |\n|ViCLIP-B | +InternVid-10M | 0.8 | 47.7 | 19.2 |\n|ViCLIP-B | +InternVid-10M | 0.9 | 47.4 | 12.1 |", "category": "Experimental_Analysis"}, {"question": "Does generating video captions from frame-level captions using a language model significantly reduce the number of motion-related words captured for video-based understanding?", "answer": "Generating video captions from frame-level captions using a language model has a negligible effect on the number of motion-related words. A comparison of unique verbs (using nltk package) in captions from a 10m subset of InternVid under two settings shows only a small discrepancy (109,859 in video captions vs. 109,895 in frame-wise captions), indicating that almost no motion-related words are lost. Two settings :\n\n1. In the first setting, the captions are video captions generated by the language model.\n\n2. In the second setting, the captions are frame-wise ones from BLIP2 and tag2text.", "category": "Experimental_Exposition"}, {"question": "Does the minor gain from 50M and 200M INTERNVID pretraining data compared to 10M data suggest that more pretraining data is not always better? Will the performance of vision and language models saturate when pretraining data reaches a certain scale?", "answer": "For the subset of 10M pretraining data, ViCLIP performs better in both zero-shot and fine-tuned action recognition when utilized with InternVid-10M-FLT/DIV compared to WebVid. This performance differences between models using InternVid-10M-FLT/DIV and InternVid-10M underscores the importance of sampling unique clips from videos - a hypothesis supported by the paper's initial explorations.\n\nThe paper's findings from Tables 2 and 5 along with Figures 7 and 8 demonstrate a consistent performance improvement in zero-shot action recognition and fine-tuned video retrieval as the pretraining data size escalates. However, the enhancement in zero-shot video retrieval is only marginal, which can be believed to stem from the limited increase in video-text diversity during pretraining. In particular, the more clips sampled from the same video, the higher tendency they have to share similar semantics. Furthermore, while a generative captioning method aids in video descriptions, the limited capacity of the captioning model restricts the diversity of produced captions, making them somewhat inferior to human annotations. These findings motivate the paper to explore scaling InternVid-10M-FLT/DIV and improving video captions through advanced captioning methods, language modeling, and human feedback techniques.\n\nFrom this standpoint, the paper's current experimental data does not conclusively suggest a saturation point for vision and language model performance as pretraining data scales up. The findings can only affirm that zero-shot video retrieval performance will reach a plateau at a certain scale of pretraining data when further diversification of the data used becomes unavailable.", "category": "Experimental_Analysis"}, {"question": "How does ViCLIP's performance compare to other video or video-language models pretrained on datasets such as HD-VILA, HT100M, or YT100M, rather than image-text models like CLIP?", "answer": "To address this concern, the authors compared ViCLIP with CLIP-ViP, a version of CLIP pretrained on HD-VILA with coarse synthetic captions(one frame caption used as the video caption) and ASR text.However, this comparison still has limitations as CLIP-ViP only offers a ViT-base version and some of its variants are trained with different settings. On MSR-VTT, ViCLIP-B, trained on InternVid-200M/-10M-FLT, outperforms CLIP-ViP-B, trained on HD-VILA-100M with subtitles, achieving a higher R@1 score (50.7%/49.0% vs. 47.7%). CLIP-ViP-B reaches similar performance (49.6%) only with a combination of captions and subtitles from HD-VILA-100M. However, this comparison has limitations due to different training settings and frame usage (CLIP-ViP-B uses 12 frames vs. ViCLIP's 8 frames).\n\nTable A. Fine-tuned video retrieval on MSR-VTT.\n| Method             | Data                           | #Frames in Train | T2V  | V2T  |\n|--------------------|--------------------------------|------------------|------|------|\n| CLIP-ViP-B         | +HD-VILA-100M                  | 12               | 47.7 |      |\n| CLIP-ViP-B         | +HD-VILA-100M(sub+cap)         | 12               | 49.6 |      |\n| CLIP-ViP-B         | +HD-VILA-100M(sub)+Im-text data | 12               | 49.1 |      |\n| ViCLIP-B           | +InternVid-200M                | 8                | 50.7 | 49.4 |\n| ViCLIP-B           | +InternVid-10M-FLT             | 8                | 49.0 | 49.2 |\n```\n\n\n", "category": "Method_Comparison"}, {"question": "Why does pretraining with InternVid-10M-FLT or InternVid-10M-DIV result in better performance in zero-shot action recognition compared to InternVid-200M, but not in the fine-tuned setting?", "answer": "The varying performance of ViCLIP (trained on InternVid-10M-FLT/DIV and InternVid-200M) in zero-shot vs. fine-tuned action recognition can be attributed to the differing demands of these tasks. Zero-shot action recognition primarily depends on the alignment between video and text representations, while fine-tuned action recognition focuses on discriminating video representation as we fine-tune the video encoder using supervised video-label data and discard the text encoder. So there's no theoretical guarantee for correlation between a model's zero-shot and fine-tuned capabilities. The interpretation for why InternVid-10M-DIV/FLT performs better in zero-shot action recognition (Table 2), but not in the fine-tuned setting (Table 3), when compared to InternVid-200M is as follows: 1. ViCLIP trained on InternVid-200M likely has a more robust video representation due to a larger quantity and diversity of videos. 2. On the other hand, ViCLIP trained on InternVid-10M-FLT/DIV might excel in video-text alignment—made possible by limiting the number of clips drawn from identical videos. This results in more diverse training data compared to InternVid-200M’s, thereby facilitating contrastive learning, enhancing video-text representation, and driving better outcomes for zero-shot action recognition and tasks like retrieval.", "category": "Experimental_Analysis"}, {"question": "What evidence is provided to demonstrate the effect of model scaling on ViCLIP's performance?", "answer": "The authors provide a comparison between two versions of ViCLIP, -L (300M) and -B (80M), in Tables B, C, and D, which show consistent improvements in zero-shot and fine-tuned video retrieval performance, as well as zero-shot action recognition, when scaling from the base to the large model. \n\nTable B. Scaling model in fine-tuned video retrieval on MSR-VTT, LSMDC, DiDeMo, MSVD, and ActivityNet\n| Method         | Data                 | MSR-VTT |  | LSMDC |  | DiDeMo |  | MSVD |  | ANet |  |\n|----------------|----------------------|---------|---------|-------|-------|--------|--------|------|------|------|------|\n|                |                      | T2V     | V2T     | T2V   | V2T   | T2V    | V2T    | T2V  | V2T  | T2V  | V2T  |\n| ViCLIP-B       | InternVid-200M       | 50.7    | 49.4    | 25.3  | 25.4  | 41.1   | 40.8   | 69.0 | 69.3 | 37.7 | 35.8 |\n| ViCLIP-L       | InternVid-200M       | 53.7    | 53.4    | 29.3  | 29.3  | 51.1   | 50.8   | 79.9 | 78.4 | 52.8 | 51.1 |\n| ViCLIP-B       | InternVid-10M-FLT    | 49.0    | 49.2    | 24.4  | 23.7  | 40.0   | 41.4   | 72.2 | 73.7 | 38.3 | 37.0 |\n| ViCLIP-L       | InternVid-10M-FLT    | 52.5    | 51.8    | 33.0  | 33.0  | 49.4   | 50.2   | 77.2 | 79.0 | 49.8 | 48.1 |\n\n\nTable C. Scaling model in zero-shot video retrieval on MSR-VTT, LSMDC, DiDeMo, MSVD, and ActivityNet.\n| Method         | Data                 | MSR-VTT |  | LSMDC |  | DiDeMo |  | MSVD |  | ANet |  |\n|----------------|----------------------|---------|---------|-------|-------|--------|--------|------|------|------|------|\n|                |                      | T2V     | V2T     | T2V   | V2T   | T2V    | V2T    | T2V  | V2T  | T2V  | V2T  |\n| ViCLIP-B        | +InternVid-200M    | 37.4        | 36.1        | 16.5      | 15.2      | 16.6       | 22.6       | 44.3     | 67.0     | 13.3     | 21.7     |\n| ViCLIP-L        | +InternVid-200M    | 39.3        | 39.5        | 18.3      | 16.6      | 17.1       | 25.5       | 47.3     | 70.0     | 13.7     | 21.6     |\n| ViCLIP-B        | +InternVid-10M-FLT | 38.6        | 38.5        | 18.5      | 17.0      | 16.3       | 25.0       | 44.8     | 67.2     | 13.0     | 21.8     |\n| ViCLIP-L        | +InternVid-10M-FLT | 42.4        | 41.3        | 20.1      | 16.9      | 18.4       | 27.9       | 49.1     | 75.1     | 15.1     | 24.0     |\n\nTable D. Scaling model in zero-shot action recognition results on Kinetics 400/600/700.\n| Method       | Training Data       | K400 top-1 | K400 AVG | K600 top-1 | K600 AVG | K700 top-1 | K700 AVG |\n|--------------|---------------------|------------|----------|------------|----------|------------|----------|\n| ViCLIP-B     | +InternVid-200M     | 56.58      | 69.20    | 53.57      | 66.20    | 45.82      | 58.28    |\n| ViCLIP-L     | +InternVid-200M     | 59.80      | 71.09    | 57.80      | 69.34    | 49.30      | 61.25    |\n| ViCLIP-B     | +InternVid-10M-FLT  | 58.52      | 71.11    | 55.37      | 68.27    | 47.09      | 59.98    |\n| ViCLIP-L     | +InternVid-10M-FLT  | 64.80      | 75.70    | 62.20      | 73.53    | 54.30      | 66.38    |\n", "category": "Experimental_Exposition"}, {"question": "What are the linear evaluation performance results of ViCLIP on the Kinetics-400 benchmark, and how do they compare to other approaches?", "answer": "The linear action recognition results on Kinetics-400 are presented in Table E. ViCLIP, trained on InternVid-10M-FLT/-200M, achieves a much higher top-1 accuracy compared to when trained on WebVid-10M, with a more than 10-point increase. ViCLIP-L performs close to TVTSv2-H/-B, which include extra learnable parameters for spatiotemporal modeling, and significantly outperforms VideoMAEv2-H. This is because MAE-based methodologies generally underperform in linear evaluations.\n\nTable E. Linear action recognition results on Kinetics-400.\n| Method         | Pretraining Data                                                                 | Kinetics-400 |          |\n|----------------|----------------------------------------------------------------------------------|--------------|----------|\n|                |                    | top-1        | top-5    |\n| VideoMAE-B [R1]| Kinetics-400       | 20.4         | -        |\n| VideoMAEv2-H [R2]| Kinetics+SthSth+AVA+WebVid2M | 25.8         | -        |\n| TVTS-B [R3]    | +YT-Temporal-180M   | 60.8         | -        |\n| TVTSv2-B [R4]  | +YT-Temporal-180M+WebVid-2M | 70.1         | -        |\n| TVTSv2-H [R4]  | +YT-Temporal-180M+WebVid-2M | 73.1         | -        |\n| ViCLIP-L       | +WebVid-10M         | 60.0         | 82.9     |\n| ViCLIP-L       | +InternVid-10M-FLT  | 71.1         | 90.4     |\n| ViCLIP-L       | +InternVid-200M     | 71.7         | 90.9     |\n\n[R1] Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS. 2022.\n\n[R2] Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR. 2023.\n\n[R3] Learning Transferable Spatiotemporal Representations from Natural Script Knowledge. In CVPR. 2023\n\n[R4] TVTSv2: Learning Out-of-the-box Spatiotemporal Visual Representations at Scale. In arXiv. 2023", "category": "Method_Comparison"}, {"question": "What steps were taken to investigate potential biases in the InternVid dataset, and what were the findings regarding age, gender, and race distributions?", "answer": "The authors investigated potential biases in the InternVid dataset by focusing on age, gender, and race distributions. The methodology involved counting keywords related to these categories in the video captions. The results are as follows: 1. Age distribution: 30.71% of captions contained age-related descriptions, with 84.59% about adults, 15.31% about children, and 0.08% about senior citizens. 2. Gender distribution: 33.7% of captions contained gender-related text, with 64.27% pertaining to men and 35.73% to women. 3. Race distribution: Only 2.51% of captions contained race-related descriptions, likely due to limitations in the captioning pipeline. Further analysis using a dedicated race recognition model is needed for more accurate statistics. ", "category": "Method_Disambiguation"}, {"question": "Does InternVID improve VideoCrafter and VideoFusion in terms of IS, FID, FVD, and CLIPSIM metrics, and why was InternVID-Aes-18M chosen over the full InternVID-200M dataset?", "answer": "The lack of full training code and details from VideoCrafter and VideoFusion makes it challenging to reproduce their results or train them further with additional data. The choice of InternVID-Aes-18M over the full InternVID-200M dataset is based on studies [1], which suggest that image aesthetic quality has a greater impact on text-to-image generation than the quantity of images. This principle is hypothesized to apply to video generation as well. To curate the subset, video aesthetics were computed by calculating the image aesthetic scores of four randomly sampled frames from each video clip using the LAION aesthetic predictor. The highest score among the frames was used to represent the video's aesthetic score, ensuring the model was trained on aesthetically pleasing videos, potentially improving the overall quality of the generated outputs. However, this approach has not been explicitly tested on VideoCrafter or VideoFusion, so it remains uncertain whether InternVID would improve IS, FID, FVD, and CLIPSIM metrics for these models. \n\n[1]Emu: Enhancing Image Generation Models Using Photogenic Needles in a Haystack. In arxiv. 2023.", "category": "Motivation_Analysis"}, {"question": "What are the quantitative and qualitative improvements in VideoChat-ViCLIP compared to vanilla VideoChat after replacing the previous visual encoder with ViCLIP, as discussed in Section 5.2 and Figures 13-17?", "answer": "Quantitatively, VideoChat-ViCLIP significantly outperforms the vanilla VideoChat (using Eva-g as the vision encoder) and other systems across all evaluation aspects of the quantitative video conversation evaluation framework in [1]. Specifically,as shown in Table D, improvements are observed in correctness of information (from 2.23 to 2.86), contextual understanding (from 2.53 to 3.08), and temporal understanding (from 1.94 to 2.36), with the average score increasing from 2.29 to 2.64, showing an overall performance gain. Qualitatively, the system with ViCLIP demonstrates better contextual appropriateness, correct information, and accurate temporal dynamics, leading to more coherent and meaningful dialogue exchanges. A qualitative comparison is provided in the [anonymous link][2]\n\nTable D. Performance benchmarking of text generation models.\n\n| Evaluation Aspect         | Correctness of Information | Detail Orientation | Contextual Understanding | Temporal Understanding | Consistency | Avg  |\n|---------------------------|----------------------------|--------------------|---------------------------|------------------------|-------------|------|\n| VideoChat (Eva-g)         | 2.23                       | 2.5                | 2.53                      | 1.94                   | 2.24        | 2.29 |\n| LLaMA Adapter             | 2.03                       | 2.32               | 2.3                       | 1.98                   | 2.15        | 2.16 |\n| Video LLaMA               | 1.96                       | 2.18               | 2.16                      | 1.82                   | 1.79        | 1.98 |\n| Video-ChatGPT             | 2.4                        | 2.52               | 2.62                      | 1.98                   | 2.37        | 2.38 |\n| VideoChat-ViCLIP-L        | 2.86                       | 2.52               | 3.08                      | 2.36                   | 2.4         | 2.64 |\n\n\n[1]Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models. In arxiv. 2023.\n\n[2]https://s1.imagehub.cc/images/2023/11/17/b164d37b037b2fcf26a2bc5d6947e3ae.jpeg", "category": "Method_Comparison"}, {"question": "What is the definition of UMT-SIM, and how is it used to compute the similarity score between a video clip and accompanying text?", "answer": "UMT-SIM refers to the use of Unmasked Teacher (UMT)[1] to compute the similarity score between a given video clip and the accompanying text. UMT is a video-language model that performs similarly to CLIP in terms of video-text similarity computation, but is trained in different data and task settings. Specifically, the paper extracts the normalized video embedding of the given clip using UMT's video encoder and the normalized text embedding of the text using UMT's text encoder, and then computes their cosine similarity.\n\n[1]Unmasked teacher: Towards training-efficient video foundation models. In ICCV. 2023.", "category": "Concept_Understanding"}, {"question": "What are the filtering strategies applied to create the InternVid-10M-FLT dataset, and how is diversity sampling (DIV) implemented?", "answer": "For InternVid-10M-FLT, the paper applies filtering strategies to video data in addition to DIV sampling. The filtering strategies for InternVid-10M-FLT include: a) Removing video clips shorter than 1s (approximately 23.15% of the total) or longer than 120s (around 0.84% of the total). b) Computing CLIPScore for each video clip using a randomly sampled frame from the clip with OpenAI’s CLIP-ViT-L/14, and then selecting clips within the top 30% of CLIPScores. c) Sampling 10M out of the remaining clips using DIV sampling. For DIV (diversity sampling), the frequencies of long videos in the segmented clip pool are counted, and clips are sampled with probabilities inverse to these frequencies to maximize data diversity.", "category": "Method_Mechanics"}, {"question": "What is the purpose of introducing the three interleaving video clip schemes in Figure 5, and why is no comparison or further analysis provided?", "answer": "The three interleaving video clip schemes in Figure 5 were introduced to illustrate potential applications of the dataset. They are not claimed as contributions, and no comparison or further analysis is provided because the intention was solely to demonstrate possible uses of the dataset.", "category": "Concept_Understanding"}, {"question": "Why was CLIP chosen for comparison in action recognition tasks, and what alternative models were used to ensure a fairer comparison?", "answer": "CLIP was initially chosen for comparison, but the authors acknowledge that it might not be the most suitable model for action recognition. To ensure a fairer comparison, the method was also compared with other specialized video models for action recognition, such as ViCLIP, CLIP-ViP, and TVTSv2. Notably, ViCLIP trained only using contrastive loss and without additional learnable parameters on InternVid outperforms both CLIP-ViP and TVTSv2, demonstrating the effectiveness of the dataset.\n\n| Method          | Training Data |  K400  |        | K600  |         | K700  |         |\n|-----------------|---------|---------|---------|---------|-------|-------|-------|\n|                 |                | top-1 | AVG   | top-1 | AVG   | top-1 | AVG   |\n| CLIP-ViP-B [2] | +HD-VILA-100M(sub+cap)+Img-text data             | 37.6  | -     | -     | -     | -     | -     |\n| TVTSv2-B [3]   | +YT-Temporal-180M+WebVid-2M                      | 54.3  | -     | -     | -     | -     | -     |\n| ViCLIP-B        | +InternVid-200M                                  | 56.58 | 69.20 | 53.57 | 66.20 | 45.82 | 58.28 |\n| ViCLIP-B        | +InternVid-10M-FLT                               | 58.52 | 71.11 | 55.37 | 68.27 | 47.09 | 59.98 |\n| ViCLIP-L        | +InternVid-200M                                  | 59.80 | 71.09 | 57.80 | 69.34 | 49.30 | 61.25 |\n| ViCLIP-L        | +InternVid-10M-FLT                               | 64.80 | 75.70 | 62.20 | 73.53 | 54.30 | 66.38 |\n| TVTSv2-H [R3]   | +YT-Temporal-180M+WebVid-2M                      | 59.6  | -     | -     | -     | -     | -     |\n\n\n[2] CLIP-ViP: Adapting Pre-trained Image-Text Model to Video-Language Representation Alignment. In ICLR. 2023.\n\n[3] TVTSv2: Learning Out-of-the-box Spatiotemporal Visual Representations at Scale. In arXiv. 2023.\n\n", "category": "Experimental_Exposition"}, {"question": "What is the novelty of the proposed method for generating the video-text dataset, and how does it address the data bottleneck in video-text pretraining? Additionally, how does ViCLIP demonstrate the effectiveness of the dataset despite not being entirely novel?", "answer": "The novelty of the proposed method lies in its scalable approach for constructing an efficient video-text dataset using Language Models (LM), addressing a well-known data bottleneck in video-text pretraining. In comparison to image-text pretraining, video-text pretraining is less explored due to the dearth of high-quality, large-scale video-language datasets. The paper’s approach addresses this gap by providing a means to generate such datasets efficiently and at scale. The paper’s method utilizes both BLIP2 and tag2text for frame captioning to ensure effective video description generation—BLIP2 focuses on the main objects and actions, while tag2text captures finer details and attributes. While ViCLIP is not entirely novel, it itself serves as a demonstration of the effectiveness of the dataset, which has shown to achieve state-of-the-art zero-shot action recognition performance on Kinetics 400/600/700. Furthermore, ViCLIP pretrained on InternVid-10M-FLT/DIV significantly outperforms the same model trained on WebVid10M which uses alt-text as captions, indicating the superior quality of the dataset. As for further architectural improvements and training techniques, there is room for exploration indeed. In this work, however, the paper’s focus is primarily on introducing and demonstrating the effectiveness of the dataset the paper has produced.", "category": "Method_Disambiguation"}, {"question": "What were the criteria used to select the image captioning models (BLIP2 and tag2text) and the language model (T5-summary) in the research?", "answer": "The image captioning models (BLIP2 and tag2text) were selected based on their exceptional performance in captioning tasks, validated on NoCaps and COCO datasets, with a focus on their robustness in generating accurate and meaningful captions. The T5-summary model was chosen after qualitative evaluations showed it outperformed existing LLMs like Llama2-7b and Llama2-13b. The selection criteria for T5 included: 1. Speed: T5 completed summaries in 14.96ms compared to Llama2-7b's 81.51ms on an A100-80G. The T5 model showcased significantly faster processing speed compared to the Llama models. 2. Absence of hallucinations: T5 did not produce hallucinations, enhancing its reliability.", "category": "Method_Disambiguation"}, {"question": "Were all generated descriptions at the finer level used for training the final model?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the methodology used for bias analysis in the InternVid dataset based on counting keywords in video captions?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was CLIPScore used to select video clips within the top 30% for the FLT process?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the T5-summary model the primary pre-trained language model used for processing captions?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the gender-related captions in the InternVid dataset predominantly about men?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "farT6XXntP", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models", "authors": ["Haoran Xu", "Young Jin Kim", "Amr Sharaf", "Hany Hassan Awadalla"], "abstract": "Generative Large Language Models (LLMs) have achieved remarkable advancements in various NLP tasks. However, these advances have not been reflected in the translation task, especially those with moderate model sizes (i.e., 7B or 13B parameters), which still lag behind conventional supervised encoder-decoder translation models. Previous studies have attempted to improve the translation capabilities of these LLMs, but their gains have been limited. In this study, we propose a novel fine-tuning approach for LLMs that is specifically designed for the translation task, eliminating the need for the abundant parallel data that traditional translation models usually depend on.\nOur approach consists of two fine-tuning stages: initial fine-tuning on monolingual data followed by subsequent fine-tuning on a small set of high-quality parallel data.  We introduce the LLM  developed through this strategy as **A**dvanced **L**anguage **M**odel-based tr**A**nslator (**ALMA**). Based on LLaMA-2 as our underlying model, our results show that the model can achieve an average improvement of more than 12 BLEU and 12 COMET over its zero-shot performance across 10 translation directions from the WMT'21 (2 directions) and WMT'22 (8 directions) test datasets. The performance is significantly better than all prior work and even superior to the NLLB-54B model \\citep{nllb} and GPT-3.5-text-davinci-003, with only 7B or 13B parameters. This method establishes the foundation for a novel training paradigm in machine translation.", "keywords": "Machine Translation;Language Language Models;Multilingual", "qa_pairs": [{"question": "How do the fine-tuned ALMA models compare with the dedicated encoder-decoder based translation models, such as NLLB-54B, in terms of performance in both high-resource and low-resource language settings?", "answer": "Most online translation models are generally smaller, often adhering to the transformer-big architecture [1], which typically results in models having around 300 million parameters. In the paper, the NLLB-200 model serves as a benchmark. It is one of the strongest encoder-decoder based translation models available, and its size is comparable to the paper’s ALMA models. The NLLB-54B's base model has 3 billion parameters, with additional parameters derived from a Mixture of Experts (MoE) approach. The paper's research demonstrates that for high-resource languages such as Chinese, German, Russian, Czech, and the low-resource language Icelandic, these ALMA models either match or outperform the dedicated encoder-decoder model NLLB-54B. However, for very low-resource languages, the monolingual data is limited.\n\nReference: [1] Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser Ł, Polosukhin I. Attention is all you need. Advances in neural information processing systems. 2017;30.", "category": "Method_Comparison"}, {"question": "What are the outcomes of using parallel data as monolingual data during the pre-training phase compared to using authentic monolingual data, such as the Oscar dataset?", "answer": "At an early stage in the research, treating parallel data as monolingual data was experimented with during the pre-training phase, which used 20 million English-Russian parallel sentences, amounting to a total of 1.8 billion tokens. The initial approach involved fine-tuning the LLaMA-2-7B model on these 1.8 billion tokens, mimicking the pre-training process, followed by further fine-tuning on English to Russian translation using the Flores development and test datasets (comprising approximately 2,000 sentences in total). To ensure a fair comparison, the authors also conducted an experiment where we fine-tuned LLaMA-2-7B on 1.8 billion tokens from the Oscar dataset, consisting of both English and Russian data, and which was also followed by fine-tuning on the same English-Russian parallel data. The outcomes of these comparative studies are presented below:\n\n|                              | BLEU  | Comet22 |\n|------------------------------|-------|---------|\n| Parallel data as monolingual data | 25.26 |  85.29  |\n| Oscar as monolingual data    | **26.43** |  **86.36**  |\n\nThe results indicate that treating parallel data as monolingual data is beneficial, but it does not yield as favorable outcomes as using authentic monolingual data, such as that from the Oscar dataset. It is hypothesized that the advantage of using Oscar’s monolingual data lies in its document-level context, which enables ALMA models to learn more natural expressions in Russian. In contrast, treating parallel data as monolingual data limits learning to sentence-level information, which may not provide the same depth of linguistic understanding and fluency.", "category": "Experimental_Exposition"}, {"question": "What are the key factors for achieving the highest translation performance in LLM-based translation systems, and how does the paper's approach compare to dedicated encoder-decoder models like NLLB-54B?", "answer": "The key to effective LLM-based translation lies in acquiring general language knowledge and establishing good bilingual alignment. A proficient translation model should first thoroughly understand the target language. However, many LLMs, primarily trained in English, often fall short in comprehending other languages. This deficiency is a primary factor behind the subpar translation performance of moderate-sized LLMs. Once a robust grasp of the target language is achieved, the next critical step is to ensure proper (bilingual) alignment. This alignment steers the model towards generating the desired translation output. The paper's supervised fine-tuning with high-quality data is a viable approach to achieve this alignment, but there could potentially be more effective methods to further enhance this alignment process.\n\nTo illustrate the differences in translation quality between a dedicated encoder-decoder model (NLLB-54B, trained on sentence-level parallel data) and the paper's ALMA-13B-LoRA model (which learns more general language information + bilingual alignment), consider an example involving specific Chinese internet slang. The terms “八卦” (bā guà) and “实锤” (shí chuí) literally translate to “eight trigrams\" and \"solid hammer\" respectively, but in slang, they mean “gossip\" and \"solid truth.\" When translating the phrase “这个八卦实锤了”, most translation systems, including NLLB-54B, face challenges. However, the paper's ALMA-13B-LoRA model exhibits a more natural and accurate understanding in its translation:\n\n- Source Chinese sentence: \"这个八卦实锤了\" \n- NLLB-54B translation: \"This is a real shame.\"\n- ALMA-13B-LoRA translation: \"The gossip is true.\"", "category": "Method_Comparison"}, {"question": "What are the results of evaluating the ALMA model on out-of-domain test sets, specifically comparing fine-tuning with only Flores data versus fine-tuning with Flores+WMT data?", "answer": "The paper compares the ALMA model fine-tuning only Flores to the one fine-tuned by Flores+WMT data. Since the test data is WMT’22 (even though it has 4 difference domains), The WMT training data used is ‘in-domain’ data, whereas the Flores data is ‘out-of-domain’ data. The results show that training with out-of-domain data may lead the lexical match drop (BLEU score down), while COMET scores are similar or even better. Appendix H of this paper contains more detailed experiments concerning the evaluation of out-of-domain test sets.", "category": "Experimental_Exposition"}, {"question": "How does the application of decoder-only models to machine translation represent a significant innovation and paradigm shift, given that the general concept of 'pre-training + fine-tuning' and the use of decoder-only models are not novel in themselves?", "answer": "'The new paradigm' specifically addresses machine translation in a new method, rather than proposing a general methodology. While the general concept of 'pre-training + fine-tuning' is indeed widely used in NLP tasks, it represents a broad strategy rather than a specific methodology. Although the concepts of 'modeling' (using a decoder-only model) and 'training' (through causal language modeling) are not novel in themselves, applying these techniques to machine translation is a significant innovation.The paper discusses three 'training' methods for translation: causal language modeling, prefix language modeling, and mixture of denoisers in Appendix A, but the simplest CLM performs the best. Traditionally, machine translation has primarily relied on processing vast quantities (millions) of parallel sentences using an encoder-decoder architecture. The application of decoder-only models to this field is basically unexplored and has rarely outperformed the conventional encoder-decoder approach. The remarkable performance of models like ChatGPT has sparked interest in decoder-only large language models (LLMs) for various applications. There must be potential for machine translation. However, most recent studies on machine translation for LLMs fail to achieve comparable performance to conventional encoder-decoder models. This area remains largely untapped, posing critical questions:\n\n - How should we train decoder-only models for translation tasks which were traditionally training on encoder-decoder architecture? \n- Given that LLMs are already extensively trained, is there still a need for large volumes of parallel data for additional training? Could this potentially impair the model's effectiveness? \n- Considering that LLMs are predominantly trained in English, how can we enhance their proficiency in other languages to improve translation quality? \n\nThe paper seeks to address these questions, proposing a novel and efficient training methodology for machine translation. The paper is the first to utilize 7B/13B LLMs for machine translation, and achieve performance comparable to state-of-the-art encoder-decoder models. This breakthrough lays the groundwork for future research in integrating LLMs with machine translation, marking an exciting new direction in the field. Although previous studies, such as BigTranslate, have implemented similar fine-tuning approaches, there are significant differences between their methods and the paper's method, particularly in acknowledging the negative impacts of excessive parallel data and the importance of data quality. Their approach resulted in performances that did not meet the paper's expectations. The paper's goal is to provide an accurate and effective training 'recipe' for decoder-only LLMs in machine translation, ensuring high performance. ", "category": "Method_Disambiguation"}, {"question": "Why are the few-shot prompting results considered undervalued, and how does the performance of the fine-tuned LLaMA-2-13B model compare to ALMA-13B-LoRA models in English-to-other language translations?", "answer": "The model evaluated for few-shot learning is not the standard LLaMA-2-13B, but rather LLaMA-2-13B after fine-tuning with 12 billion monolingual data tokens. This means that both the paper's supervised fine-tuning and few-shot learning methods are built upon a model that has extensively pre-trained on a large monolingual corpus. **Contexts where monolingual data fine-tuning is not applied are presented** in Table 6 (Appendix E). In this table, it is evident that a 5-shot fine-tuning approach applied solely on LLaMA-2-13B falls behind the ALMA-13B-LoRA model by 12.18 COMET points on the **en → xx** direction. Few-shot learning performance can be significantly influenced by the quality of the prompts used. Nonetheless, supervised fine-tuning still **offers several advantages** under these conditions:\n\n *   **Performance Edge**: Supervised fine-tuning **outperforms** few-shot learning, albeit marginally, especially in English-to-other language (en → xx) translations. This observation is consistent with findings from several previous studies ([1], [2]).\n *   **Stability of Results**: The performance of few-shot learning tends to vary considerably, which is heavily dependent on the quality of examples (shots) included in the prompt, as demonstrated in Table 11.\n*   **Resource Efficiency**: Few-shot learning typically requires a longer prefix, consuming more memory. If a hypothetical scenario is considered where each input token incurs a cost, supervised fine-tuning would be more economical in terms of resource utilization.\n\nReference: [1] Liu H, Tam D, Muqeeth M, Mohta J, Huang T, Bansal M, Raffel CA. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems. 2022 Dec 6;35:1950-65. \n\n[2] Mosbach M, Pimentel T, Ravfogel S, Klakow D, Elazar Y. Few-shot Fine-tuning vs. In-context Learning: A Fair Comparison and Evaluation. arXiv preprint arXiv:2305.16938. 2023 May 26.", "category": "Method_Disambiguation"}, {"question": "How does the proposed fine-tuning approach generalize to other large language models (LLMs) and languages, particularly those not well-represented in the pre-training data?", "answer": "The effectiveness of downstream tasks is closely linked to the pre-training stage. This understanding actually underscores the importance of non-English monolingual fine-tuning for these models. LLMs mainly trained in English are likely to exhibit performance trends similar to the MPT model, especially in languages like Russian, which they have not been exposed to during training. In such cases, any improvements in these models are solely reliant on the Russian knowledge extracted from parallel data. However, learning from parallel data can wash out the knowledge learned during the pre-training phase — to the extent that a model initialized randomly could achieve comparable performance to MPT after being trained with 20 million parallel data sentences. The primary aim of monolingual fine-tuning is to equip the model with a general understanding of a new language. Following this, parallel data fine-tuning is employed to guide the model in accurately utilizing this knowledge for effective translation. The paper's approach is designed to be universally applicable across all LLMs and languages. It involves fine-tuning the model on monolingual data of the target language, regardless of its resource availability, and subsequently fine-tuning on the corresponding parallel data. This method ensures that the model not only learns the new language but also becomes proficient in applying this knowledge for translation tasks. To demonstrate the generalization of the paper's methods to other large language models (LLMs), the authors extended the paper's experiments to Mistral-7B, another well-known but recently released model. The authors fine-tuned Mistral-7B using 18 billion tokens of monolingual data, which included a mix of five new languages as detailed in the paper, then evaluated its performance on the same multilingual test set used for the paper's previous experiments. The results, as presented below, show that ALMA models, whether based on LLaMA-2-7B or Mistral-7B, are capable of achieving comparable levels of performance. This indicates the potential of the paper's fine-tuning approach to be effectively generalized across different LLM architectures.\n\n|                | Avg. en>xx |       | Avg. xx>en |       |\n|----------------|------------|-------|------------|-------|\n|                | BLEU       | Comet22 | BLEU       | Comet22 |\n| ALMA-LLaMA-7B-LoRA | 29.78 | 86.37 | 34.31 | 84.12 |\n| ALMA-Mistral-7B-LoRA | 30.78 | 86.48 | 34.07 | 84.09 |\n\n\nTo demonstrate the adaptability of the paper's methods to low-resource languages, specifically Gujarati (gu), the authors initially fine-tuned LLaMA-2-7B using 0.1 billion Gujarati tokens and subsequently fine-tuned it on the Flores-200 development data. The model was then evaluated using the Flores-200 devtest dataset. The results of this experiment are detailed below. The performance of the paper’s fine-tuned model was compared against the zero-shot performance of LLaMA-2-7B and the outputs from BigTranslate. These comparisons revealed that while the other models produced some nonsensical outputs, the paper's fine-tuned model significantly outperformed them, showcasing the effectiveness of the paper's method. This was a preliminary experiment, constrained by time limitations, and does not represent the peak potential performance for Gujarati. Significant performance enhancement is anticipated with access to more monolingual data, higher-quality data, and more refined training practices. This experiment serves as evidence that the paper's method can be successfully extended to low-resource languages like Gujarati.\n\n|                | en>gu       |       | gu>en       |       |\n|----------------|-------------|-------|-------------|-------|\n|                | BLEU        | COMET22 | BLEU        | COMET22 |\n| LLAMA-2-7B, zero-shot | 0.48      | 38.88 | 0.18      | 39.33  |\n| BigTranslate   | 0.23      | 43.80 | 1.87      | 51.09  |\n| ALMA-7B-GU     | **11.31**     | **79.42** | **15.05**     | **66.60** |\n", "category": "Experimental_Exposition"}, {"question": "How does the application of decoder-only models and causal language modeling to machine translation justify the term 'New paradigm' in the title, given that these techniques are not novel in themselves?", "answer": " 'The new paradigm' specifically addresses machine translation in a new method, rather than proposing a general methodology. Although the concepts of 'modeling' (using a decoder-only model) and 'training' (through causal language modeling) are not novel in themselves, applying these techniques to machine translation is a significant innovation. The paper discusses three 'training' methods for translation: causal language modeling, prefix language modeling, and mixture of denoisers in Appendix A, but the simplest CLM performs the best. Traditionally, machine translation has primarily relied on processing vast quantities (millions) of parallel sentences using an encoder-decoder architecture. The application of decoder-only models to this field is basically unexplored and has rarely outperformed the conventional encoder-decoder approach. The remarkable performance of models like ChatGPT has sparked interest in decoder-only large language models (LLMs) for various applications. There must be potential for machine translation. However, most recent study on machine translation for LLMs fails to achieve comparative performance to conventional encoder-decoder models. This area remains largely untapped, posing critical questions: \n\n- How should we train decoder-only models for translation tasks which were traditionally training on encoder-decoder architecture? \n- Given that LLMs are already extensively trained, is there still a need for large volumes of parallel data for additional training? Could this potentially impair the model's effectiveness? \n- Considering that LLMs are predominantly trained in English, how can we enhance their proficiency in other languages to improve translation quality? \n\nThe paper seeks to address these questions, proposing a novel and efficient training methodology for machine translation. The paper is the first to utilize 7B/13B LLMs for machine translation, and achieve performance comparable to state-of-the-art encoder-decoder models. This breakthrough lays the groundwork for future research in integrating LLMs with machine translation, marking an exciting new direction in the field.", "category": "Method_Disambiguation"}, {"question": "What are the specific contributions of monolingual and parallel training data to the performance improvements observed in the ablation study presented in Table 3?", "answer": "The contribution of monolingual data is evident when the same parallel data fine-tuning is applied or when no parallel data training is used, as models fine-tuned with monolingual data show notable improvements, particularly in English to other language translations (e.g., the COMET score for en→xx translation jumps from 76.52 to 86.49). The contribution of parallel data is observed when models are trained with higher-quality parallel data, with the monolingual data training remaining constant or absent, leading to progressive performance enhancements (e.g., COMET score for en→xx translation increases from 73.89 to 76.52).", "category": "Concept_Understanding"}, {"question": "Does the good performance of the trained ALMA models stem from the reduced number of non-English languages (5) compared to other models like NLLB, GPT-3.5, and GPT-4?", "answer": "Although there is a phenomenon where training a model with an increasing number of languages can lead to diminished performance in high-resource languages [1], the primary focus of the paper is on developing robust translation models based on decoder-only large language models (LLMs), rather than exploring how the inclusion of multiple languages affects the performance of the ALMA translation model. As noted in the paper, comparisons with models like GPT-4 or NLLB are not meant to be direct or fair benchmarks but rather serve as a context to gauge the performance of the paper's models in relation to these well-known systems. It's important to highlight that the training data for models like GPT3.5/4 and other moderate-sized LLMs, such as LLaMA and MPT, often lack detailed language labeling. For instance, LLaMA-2's training data is composed of 90% English content and approximately 8% in unidentified languages (as detailed in Table 10 of their paper [2]). This lack of precise language identification complicates any analysis regarding the impact of the number of languages trained on these models, and makes it challenging to study the curse of multilinguality in this context. While comparisons between ALMA and models like NLLB or GPT3.5/4 are not entirely equitable due to differences in model size and training data, there is potential interest in investigating how the number of languages trained might affect ALMA’s performance.\n\n[1] Conneau A, Khandelwal K, Goyal N, Chaudhary V, Wenzek G, Guzmán F, Grave E, Ott M, Zettlemoyer L, Stoyanov V. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. 2019 Nov 5.\n\n[2] Touvron H, Martin L, Stone K, Albert P, Almahairi A, Babaei Y, Bashlykov N, Batra S, Bhargava P, Bhosale S, Bikel D. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288. 2023 Jul 18.", "category": "Method_Disambiguation"}, {"question": "Would the improvement over LLaMA-7B with fine-tuning hold if few-shot learning were used instead, and what are the results of such experiments?", "answer": "The authors conducted experiments on 5-shot learning for LLaMA-2-13B and ALMA-13B-LoRA, as detailed in Table 6 and Table 11 of the paper. The results show that higher-quality shots improve performance, but there is a notable performance gap between LLaMA-2-13B and ALMA-13B-LoRA. For example, the average COMET score for English to various languages (en→xx) is 74.82 for LLaMA-2-13B, significantly lower than the 87.00 achieved by ALMA-13B-LoRA. 5-Shot Learning for LLaMA-2-13B After Monolingual Fine-Tuning: After undergoing monolingual fine-tuning with 12 billion tokens, the results for 5-shot learning are detailed in Table 11, where it still does not match the performance of fully supervised learning. For example, in the case of en→xx, the average COMET score for 5-shot learning is 85.24, which is still lower than the 86.37 score achieved through supervised learning. These findings highlight the limitations of few-shot learning in certain contexts.", "category": "Experimental_Exposition"}, {"question": "What does the caption 'The average performance of ALMA-7B at the completion of each 1B-token fine-tuning' in Figure 5 mean in terms of the fine-tuning stages, and how do these performance numbers compare to those in Tables 1 and 2?", "answer": "The caption refers to the ALMA model's performance after each phase of fine-tuning with 1 billion monolingual tokens, followed by fine-tuning with human-written parallel data. For example, '3B' in Figure 5 indicates fine-tuning on 3 billion monolingual tokens plus parallel data fine-tuning. The results in Tables 1 and 2 are for models fine-tuned with 20 billion tokens followed by parallel data fine-tuning. The performance numbers in Figure 5 are consistent with those in Tables 1 and 2, as shown by the average (86.49 + 84.08) / 2 = 85.28 for ALMA-7B.", "category": "Concept_Understanding"}, {"question": "How does the performance of the proposed method vary with different prompts, and what steps were taken to verify the stability of the method?", "answer": "The idea that different prompts can influence translation performance is evident, and the potential that enriching these prompts could further improve results is also recognized. The paper's primary focus, however, is to demonstrate that a properly formulated prompt enables moderate-sized large language models (LLMs) to achieve performance comparable to state-of-the-art (SOTA) encoder-decoder translation models or GPT3.5/4. The prompts the paper has employed are widely recognized and validated in the field of machine translation, as evidenced by previous research ([1], [2], [3]). The authors re-train ALMA-7B-LoRA model with **two new prompts** as suggested in [4]: \n\n - the paper's prompt is: Translate this from <src> to <tgt>:\\n<src>: <input>\\n <tgt>: \n\n- New Prompt 1: <src>: <input>\\n<tgt>: \n\n- New Prompt 2: <input> Translate this from <src> to <tgt>: \n\n|                | Avg. en>xx |       | Avg. xx>en |       |\n|----------------|------------|-------|------------|-------|\n| Prompt (for ALMA-7B-LoRA) | BLEU  | Comet22 | BLEU  | Comet22 |\n| The paper's prompt               | **29.78** | **86.80**   | **34.34** | **84.24**   |\n| New prompt 1             | 27.63 | 85.44   | 33.434| 83.60   |\n| New prompt 2             | 27.44 | 85.31   | 33.314| 83.85   |\n\nThe authors conducted evaluations of the two new prompts using the WMT’22 test set and report the average scores for both translation directions. The performance of both new prompts is similar, though they slightly underperform compared to the paper's originally proposed prompt.\n\n[1] Hendy A, Abdelrehim M, Sharaf A, Raunak V, Gabr M, Matsushita H, Kim YJ, Afify M, Awadalla HH. How good are gpt models at machine translation? a comprehensive evaluation. arXiv preprint arXiv:2302.09210. 2023 Feb 18.\n\n[2] Chen Y, Liu Y, Meng F, Chen Y, Xu J, Zhou J. Improving Translation Faithfulness of Large Language Models via Augmenting Instructions. arXiv preprint arXiv:2308.12674. 2023 Aug 24.\n\n[3] Zeng J, Meng F, Yin Y, Zhou J. Tim: Teaching large language models to translate with comparison. arXiv preprint arXiv:2307.04408. 2023 Jul 10.\n\n[4] Zhang B, Haddow B, Birch A. Prompting large language model for machine translation: A case study. arXiv preprint arXiv:2301.07069. 2023 Jan 17.", "category": "Experimental_Exposition"}, {"question": "How does the fine-tuning process in this study, particularly the use of Low-Rank Adaptation (LoRA) with bilingual corpora, prevent the impairment of other NLP capabilities such as document summarization and logical question answering?", "answer": "The fine-tuning process involves two stages: monolingual fine-tuning to enhance the large language model's (LLM) proficiency in general non-English languages. The first stage does not significantly impact the model's performance in English across various tasks. For the second stage, Low-Rank Adaptation (LoRA) is used as a lightweight, plug-and-play module during fine-tuning with bilingual corpora. This approach avoids negative impacts on other NLP capabilities by allowing the model to extend its application to tasks like multilingual summarization or question-answering through task-specific LoRA modules, preserving overall proficiency in diverse NLP tasks.", "category": "Method_Mechanics"}, {"question": "Why was the evaluation of the InstructScore metric limited to translations from Chinese to English, and how does the performance of ALMA-13B-LoRA compare to GPT-3.5-D using this metric?", "answer": "The evaluation of the InstructScore metric was limited to translations from Chinese to English because the prompt for this metric is specifically tailored for this language pair.   The results show that ALMA-13B-LoRA outperforms GPT-3.5-D, with ALMA-13B-LoRA achieving a score of -4.5856 compared to GPT-3.5-D's score of -5.2944, where higher scores indicate better performance.\n\n|             | zh->en (Instructscore) |  \n|-------------|------------------------|  \n| GPT-3.5-D   | -5.2944                |  \n| ALMA-13B-LoRA | **-4.5856**              |  \n\nTo provide a more thorough assessment using recently released evaluation metrics, additional evaluation metrics used include wmt23-cometkiwi-da-xxl and Unbabel/XCOMET-XXL, both of which are 10 billion parameter models supporting all languages.The results, as presented in the table below, indicate that the paper's method continues to outperform GPT-3.5-D when evaluated with these latest metrics.\n\n|                 | Avg. xx->en |            |            |\n|-----------------|-------------|------------|------------|\n|                 | Comet22     | wmt23-cometkiwi-da-xxl | XCOMET-XXL |\n| GPT-3.5-D       | 83.90       | 80.06              | 84.59      |\n| ALMA-13B-LoRA   | **84.59**   | **81.50**          | **86.74**  |\n\n|                 | Avg. en->xx |            |            |\n|-----------------|-------------|------------|------------|\n|                 | Comet22     | wmt23-cometkiwi-da-xxl | XCOMET-XXL |\n| GPT-3.5-D       | 84.59       | 75.99              | 88.10      |\n| ALMA-13B-LoRA   | **87.00**   | **82.66**          | **92.76**  |\n\n\n", "category": "Method_Comparison"}, {"question": "What is the explanation for the generalized translation capabilities of large language models (LLMs)?, and does the proposed ALMA method fundamentally assist LLMs in learning deep bilingual text alignment?", "answer": "LLMs acquire extensive general knowledge from the massive volumes of monolingual data they process, as seen in the paper’s first stage of fine-tuning. This is a significant advantage over traditional encoder-decoder models, which rely solely on parallel data for text alignment learning. LLMs inherently possess translation capabilities; the paper’s task in the second stage is to guide the model to effectively utilize this information and produce accurate outputs. This second stage aligns more closely with AI alignment goals, directing LLM systems to achieve the specific intentions of the designers—in this case, translation. The ability of LLMs to generalize translation capabilities is attributed to their exposure to a vast array of monolingual data. For example, consider an example involving specific Chinese internet slang. The terms ‘八卦’ (bā guà) and ‘实锤’ (shí chuí) literally translate to ‘eight trigrams’ and ‘solid hammer,’ respectively, but in slang, they mean ‘gossip’ and ‘solid truth.’ When translating the phrase ‘这个八卦实锤了,’ most translation systems, including NLLB-54B, face challenges. However, the paper’s ALMA-13B-LoRA model exhibits a more natural and accurate understanding in its translation: - Source Chinese sentence: ‘这个八卦实锤了’ - NLLB-54B translation: ‘This is a real shame.’ - ALMA-13B-LoRA translation: ‘The gossip is true.’ Moreover, the results in Table 10 show that even when the training data is limited to out-of-domain Flores data, the COMET score remains high when tested on the cross-domain WMT’22 dataset, which includes conversational, news, e-commerce, and social content. This further validates the generalization capability of ALMA models. As for the question of whether the paper’s method fundamentally assists LLMs in learning deep bilingual text alignment, the authors wouldn’t claim it to be ‘fundamental’ at this stage. It is believed that simple supervised fine-tuning on high-quality parallel data is a viable method for achieving strong text alignment.", "category": "Method_Disambiguation"}, {"question": "Is the proposed recipe universally applicable across all LLMs and languages as demonstrated by experiments on multiple models and languages?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the performance of the fine-tuned model on low-resource language Gujarati significantly better than zero-shot LLaMA-2-7B and BigTranslate outputs?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the experiments include fine-tuning and evaluation of the Mistral-7B model on a mix of five new languages to demonstrate generalization?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "y886UXPEZ0", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Adapting Large Language Models via Reading Comprehension", "authors": ["Daixuan Cheng", "Shaohan Huang", "Furu Wei"], "abstract": "We explore how continued pre-training on domain-specific corpora influences large language models, revealing that training on the raw corpora endows the model with domain knowledge, but drastically hurts its prompting ability for question answering. Taken inspiration from human learning via reading comprehension--practice after reading improves the ability to answer questions based on the learned knowledge--we propose a simple method for transforming raw corpora into reading comprehension texts. Each raw text is enriched with a series of tasks related to its content. Our method, highly scalable and applicable to any pre-training corpora, consistently enhances performance across various tasks in three different domains: biomedicine, finance, and law. Notably, our 7B language model achieves competitive performance with domain-specific models of much larger scales, such as BloombergGPT-50B. Furthermore, we demonstrate that domain-specific reading comprehension texts can improve the model's performance even on general benchmarks, showing the potential to develop a general model across even more domains. Our model, code, and data are available at https://github.com/microsoft/LMOps.", "keywords": "Domain Adaption;Continual Pre-training;Large Language Model", "qa_pairs": [{"question": "How does the proposed method of incorporating reading comprehension tasks during the pre-training phase differ from the approach used in RECKONING: Reasoning through Dynamic Knowledge Encoding?", "answer": "RECKONING enhances reasoning ability by incorporating contextual knowledge into the model's parameters before exposing it to a question at inference time. Viewing the model’s learning from contextual knowledge as the ‘reading’ process and the model’s predicting on the question as the ‘comprehension’ process allows the overall RECKONING process to be considered a form of reading comprehension. However, the paper's reading comprehension differs from RECKONING in many aspects:\n - The most obvious difference lies in the goal: RECKONING enhances reasoning ability, while the paper improves domain expertise. \n- Another difference is the timing of model parameter updates: RECKONING performs these updates at inference time, whereas the paper performs updates during the pre-training stage. \n- Moreover, RECKONING learns contextual knowledge specifically for ad-hoc single or multiple related questions (the number of knowledge-related questions is set up to 18 in the experiment), while the paper learns domain-specific knowledge for various downstream tasks.", "category": "Method_Comparison"}, {"question": "Do the performance improvements shown in Table 4 (3% in biomedicine, 4.8% in finance, and 4.3% in law) justify the effort required to transform corpora into reading comprehension texts, considering the diversity of tasks within each domain?", "answer": "The performance improvements are justified as they demonstrate consistent gains across multiple tasks within each domain. In biomedicine, MedAlpaca-7B shows significant improvements on certain tasks (e.g., 7.6% on ChemProt and 2.2% on USMLE). However, the improvements are not consistent across other tasks, resulting in an average improvement of 0.9%, whereas the paper's method achieves 3.1%. Similarly, in LLaMA-13B, MedAlpaca-13B outperforms LLaMA-13B by 0.8% on average, while the paper's method achieves a 5.1% absolute gain. In the law domain, LexGPT outperforms its base model GPT-J on one task but does not exhibit the same trend on other tasks, resulting in less favorable average performance. In finance, the paper's average improvement enables it to compete with BloombergGPT-50B.\n\n|                                      | LLaMA-7B    |           |          | LLaMA-13B   |           |          |\n|--------------------------------------|-------------|-----------|----------|-------------|-----------|----------|\n|                                      | General LLM-7B | MedAlpaca-7B | AdaptLLM-7B | General LLM-13B | MedAlpaca13B |AdaptLLM-13B* |\n| PubMedQA | 59.6        | 58.6      | 63.3     | 59.6        | 60.7      | 62.5     |\n| ChemProt | 31.4        | 39.0      | 35.2     | 42.8        | 38.4      | 44.4     |\n| MQP      | 50.7        | 50.7      | 54.4     | 49.3        | 57.4      | 72.8     |\n| RCT      | 45.1        | 40.8      | 50.4     | 56.7        | 51.3      | 54.4     |\n| USMLE    | 34.5        | 36.7      | 33.1     | 34.7        | 39.0      | 34.6     |\n| AVERAGE                              | 44.2        | 45.1      | 47.3     | 48.6        | 49.4      | 53.7     |\n| **Absolute gain w.r.t. General LLM** | /           | +0.9      | **+3.1** | /           | +0.8      | **+5.1** |\n\n(* AdaptLLM-13B is trained from LLaMA-13B using the paper's method)\n", "category": "Experimental_Analysis"}, {"question": "Did the paper manually verify that its transformation process maintains the integrity, readability, and coherence of the text?", "answer": "Yes, the authors manually checked the transformed texts before the training process. They based their mining strategy on a verified method[1] and made modifications to ensure data quality. They ensured data quality by requiring longer sentence lengths and avoiding the newline character `\\n` in the mining stage, and manually verified each transformation template in the transformation stage. Additionally, they sampled 200 reading comprehension texts from each domain and manually evaluated their readability and coherence, achieving high satisfaction scores: 95.8 for BioMed, 89.0 for Finance, and 86.5 for Law.\n\n[1] Mozes van de Kar et al. Don’t prompt, search! mining-based zero-shot learning with language models. In EMNLP, Association for Computational Linguistics, 2022.", "category": "Experimental_Exposition"}, {"question": "How does the diversity of verbalizers affect the performance of the model, and what evidence supports this claim?", "answer": "The paper’s choice of verbalizers for the Natural Language Inference (NLI) task type aligns with [1], where the effectiveness has been verified. It is observed that the entailment relationship in NLI can be extended to other task types like Commonsense Reasoning and Paraphrase Detection. In Table 3 in the paper, the verbalizers for NLI are also applied to Commonsense Reasoning and Paraphrase Detection. Finally, the authors complement verbalizers for the remaining task types based on commonsense; for example, a verbalizer like ‘is defined as’ can be commonly accepted to connect a concept with its definition.\n\n**Verbalizer Effect:** Intuitively, the paper's hypothesis is that **a greater diversity of verbalizers leads to a broader range of comprehension tasks, ultimately enhancing performance**. To test this hypothesis, the authors conduct an ablation by preserving only one verbalizer for each task type. Specifically, only the first verbalizer in each row in Table 3 is preserved, and the reading comprehension data is remade to observe the effect.\n\nFirst, **the most direct impact is that the number of mined task examples decreases**. When using all the verbalizers (All Verbal), the authors mine an average of 2.1 examples per text, while using one verbalizer (Single Verbal) results in 1.4 examples per text.\n\n|                                | All Verbal | Single Verbal |\n|--------------------------------|--------------------------|-----------------------------|\n| # Mined examples per text | **2.1**                     | 1.4                         |\n\nNext, the authors conduct an experiment to train the general model using the new data, and the task results are as follows: \n\n|          | All Verbal | Single Verbal |\n|----------|:------------------------:|:---------------------------:|\n| Average Score  | **44.3**                 | 42.7                        |\n\nThese results demonstrate that with fewer verbalizers, the downstream task performance declines, verifying that **including more verbalizers can contribute to better performance**.\n\n[1] Mozes van de Kar et al. Don’t prompt, search! mining- based zero-shot learning with language models. In EMNLP, Association for Computational Linguistics, 2022.", "category": "Method_Disambiguation"}, {"question": "How does the performance of AdaptLLM compare to publicly available baselines such as GPT-J-6B and LexGPT-6B, and what are the results of these comparisons?", "answer": "AdaptLLM-6B shows a positive average score compared to GPT-J-6B and LexGPT-6B. For example, AdaptLLM-6B achieves an average score of 37.7 compared to GPT-J-6B's 35.9 and LexGPT-6B's 32.1. Additionally, AdaptLLM-13B outperforms LLaMA-13B and MedAlpaca-13B, with an average score of 53.7 compared to 48.6 and 49.4, respectively.\n\n|                 | GPT-J-6B | LexGPT-6B | AdaptLLM-6B |\n|-----------------|:--------:|:---------:|:-----------:|\n| SCOTUS-mic-F1   | 15.9     | 16.9      | 18.8        |\n| SCOTUS-mac-F1   | 13.6     | 7.7       | 20.1        |\n| CaseHOLD-mic-F1 | 34.9     | 27.0      | 34.7        |\n| CaseHOLD-mac-F1 | 34.9     | 27.0      | 34.7        |\n| UNFAIR-ToS      | 79.8     | 81.9      | 80.0        |\n| AVERAGE         | 35.9     | 32.1      | **37.7**    |\n\n(LexGPT-6B is trained from GPT-J-6B)\n\n|          | LLaMA-13B | MedAlpaca-13B | AdaptLLM-13B |\n|----------|:---------:|:-------------:|:------------:|\n| PubMedQA |   59.6    |     60.7      |   62.5   |\n| ChemProt |   42.8    |     38.4      |   44.4   |\n| MQP      |   49.3    |     57.4      |   72.8   |\n| RCT      |  56.7 |     51.3      |     54.4     |\n| USMLE    |   34.7    |    39.0   |     34.6     |\n| AVERAGE  |   48.6    |     49.4      |   **53.7**   |\n\n(MedAlpaca-13B is trained from LLaMA-13B)", "category": "Method_Comparison"}, {"question": "What was the rationale behind the selection of verbalizers for different task types, and how were they extended or complemented?", "answer": "The verbalizers for the Natural Language Inference (NLI) task type were chosen based on prior work [1], where their effectiveness was verified. The entailment relationship in NLI was extended to other task types like Commonsense Reasoning and Paraphrase Detection, as shown in Table 3. For the remaining task types, verbalizers were complemented based on commonsense, such as using 'is defined as' to connect a concept with its definition.\n\n[1] Mozes van de Kar et al. Don’t prompt, search! mining- based zero-shot learning with language models. In EMNLP, Association for Computational Linguistics, 2022.", "category": "Method_Mechanics"}, {"question": "What is the reason for the model's least performance improvement on the 'law' domain compared to other domains?", "answer": "- Biomedicine---PubMed Abstract [1]: Each text is an abstract for a research paper, which means the text should be highly clear, concise and readable. A randomly sampled example is as follows:\n\n```text\nPattern-recognition receptors: signaling pathways and dysregulation in canine chronic enteropathies-brief review.\nPattern-recognition receptors (PRRs) are expressed by innate immune cells and recognize pathogen-associated molecular patterns (PAMPs) as well as endogenous damage-associated molecular pattern (DAMP) molecules. With a large potential for synergism or convergence between their signaling pathways, PRRs orchestrate a complex interplay of cellular mediators and transcription factors, and thus play a central role in homeostasis and host defense. Aberrant activation of PRR signaling, mutations of the receptors and/or their downstream signaling molecules, and/or DAMP/PAMP complex-mediated receptor signaling can potentially lead to chronic auto-inflammatory diseases or development of cancer. PRR signaling pathways appear to also present an interesting new avenue for the modulation of inflammatory responses and to serve as potential novel therapeutic targets. Evidence for a dysregulation of the PRR toll-like receptor (TLR)2, TLR4, TLR5, and TLR9, nucleotide-binding oligomerization domain-containing protein (NOD)2, and the receptor of advanced glycation end products (RAGE) exists in dogs with chronic enteropathies. We describe the TLR, NOD2, and RAGE signaling pathways and evaluate the current veterinary literature-in comparison to human medicine-to determine the role of TLRs, NOD2, and RAGE in canine chronic enteropathies.\n```\n\n- Finance---Stock News: The paper collects stock news using the codebase of FinGPT [2], which could ensure the data quality. A randomly sampled example is as follows:\n\n```text\nChevron, Johnson & Johnson share gains lead Dow's 100-point jump\nPowered by strong returns for shares of Chevron and Johnson & Johnson, the Dow Jones Industrial Average is up Wednesday morning. Shares of Chevron CVX, -2.35% and Johnson & Johnson JNJ, -0.13% are contributing to the blue-chip gauge's intraday rally, as the Dow DJIA, -1.01% is trading 105 points higher (0.3%). Chevron's shares have climbed $3.84, or 2.4%, while those of Johnson & Johnson are up $2.85, or 1.8%, combining for an approximately 44-point boost for the Dow. Merck MRK, -0.25%, Apple Inc. AAPL,, and Intel INTC, -1.95% are also contributing significantly to the gain. A $1 move in any one of the 30 components of the benchmark equates to a 6.59-point swing. Editor's Note: This story was auto-generated by Automated Insights, an automation technology provider, using data from Dow Jones and FactSet. See our market data terms of use.\n```\n\n- Law—--FreeLaw Opinions [1]: Each text is a legal opinion from federal and state courts, documenting many oral snippets. In contrast, the corpora in the other domains mainly consist of written passages. It is inferred that the unique feature of incorporating spoken language may affect the overall readability of law corpora. A randomly sampled example is as follows (with certain portions omitted and presented as (…)). It can be observed that the readability of law corpora appears to be lower than that of the other two domains.\n\n```text\n92 P.3d 294 (2004)                                                                                                                                 \n2004 WY 71                                                                                                                                         \nSTATE of Wyoming, ex rel., WYOMING WORKERS' SAFETY AND COMPENSATION DIVISION, Appellant (Petitioner),                                              \nv.                                                                                                                                                 \nAnthony N. SAVICKI, Appellee (Respondent).                                                                                                         \n(...)\nWe have previously entertained the Division's dissatisfaction with our failure to consider subsequent wage increases based upon merit and said this:\n(...)\n[¶ 16] Conner rejected the notion that an employee may be penalized by wage increases earned for merit at a later time. We reiterate that the pivotal question is whether the employee's injury has caused a loss of earning capacity and, if so, whether a comparison of the pre-injury wage to the post-injury wage offered is compensable under the applicable statute. In this case, Savicki's post-injury wage is compensable, and the order awarding benefits is affirmed.\n```\n\n[1] Leo Gao et all. The pile: An 800gb dataset of diverse text for language modeling. CoRR, abs/2101.00027, 2021.\n\n[2] Hongyang Yang et all. Fingpt: Open-source financial large language models. CoRR, abs/2306.06031, 2023.", "category": "Experimental_Analysis"}, {"question": "Does the proposed approach benefit other types of models, such as encoder-decoder models or other decoder-only models?", "answer": "Yes, the proposed approach benefits other types of models. Experiments were conducted on Pythia[1], a decoder-only model, and the results demonstrate positive effects of the method.\n\n|          | **Pythia-70M**   |          | **LLaMA-13B**   |          | **LLaMA-2-Chat-7B** |          |\n|----------|-------------|----------|-------------|----------|--------------------|----------|\n|          | General LLM | AdaptLLM | General LLM | AdaptLLM |     General LLM    | AdaptLLM |\n| PubMedQA | 34.4        | **50.3** |    59.6     | **62.5** |        55.2        | **62.4** |\n| ChemProt | 7.6         | **10.4** |    42.8     | **44.4** |        33.8        | **36.8** |\n|    MQP   | **52.1**    |   50.5   |    49.3     | **72.8** |        59.3        | **63.9** |\n|    RCT   | 29.6        | **30.6** |   **56.7**  |   54.4   |        53.4        | **58.2** |\n|   USMLE  | 24.0        | **24.4** |   **34.7**  |   34.6   |        29.1        | **34.6** |\n|  AVERAGE | 29.5        | **33.2** |    48.6     | **53.7** |        46.2        | **51.2** |\n\n[1] Stella Biderman et al. Pythia: A suite for analyzing large language models across training and scaling. In ICML, volume 202 of Proceedings of Machine Learning Research, PMLR, 2023.", "category": "Experimental_Exposition"}, {"question": "Do the experimental results of the proposed method on the 7B LLaMA model consistently apply to larger-scale models and models that have undergone RLHF?", "answer": "Yes, the proposed method is effective on larger-scale models and models that have undergone RLHF. Experiments were conducted on Pythia-70M[1](Smaller and Different) LLaMA-13B[2] (larger model) and LLaMA-2-Chat-7B (after RLHF)[3], showing consistent improvements across various models in the biomedicine domain. Results demonstrate significant gains, as detailed in the table.\n\n|          | **Pythia-70M**   |          | **LLaMA-13B**   |          | **LLaMA-2-Chat-7B** |          |\n|----------|------------------|----------|------------------|----------|----------------------|----------|\n|          | General LLM      | AdaptLLM | General LLM      | AdaptLLM | General LLM          | AdaptLLM |\n| PubMedQA | 34.4             | **50.3** | 59.6             | **62.5** | 55.2                 | **62.4** |\n| ChemProt | 7.6              | **10.4** | 42.8             | **44.4** | 33.8                 | **36.8** |\n| MQP      | **52.1**         | 50.5     | 49.3             | **72.8** | 59.3                 | **63.9** |\n| RCT      | 29.6             | **30.6** | **56.7**         | 54.4     | 53.4                 | **58.2** |\n| USMLE    | 24.0             | **24.4** | **34.7**         | 34.6     | 29.1                 | **34.6** |\n| AVERAGE  | 29.5             | **33.2** | 48.6             | **53.7** | 46.2                 | **51.2** |\n\n[1] Mozes van de Kar et al. Don’t prompt, search! mining- based zero-shot learning with language models. In EMNLP, Association for Computational Linguistics, 2022.\n\n[2] Leo Gao et all. The pile: An 800gb dataset of diverse text for language modeling. CoRR, abs/2101.00027, 2021.\n\n[3] Hongyang Yang et all. Fingpt: Open-source financial large language models. CoRR, abs/2306.06031, 2023.\n", "category": "Experimental_Analysis"}, {"question": "Why does pre-training on domain-specific knowledge lead to a decrease in the prompting ability of large language models (LLMs)?", "answer": "It is not the domain-specific knowledge that hurts the prompting ability for question answering; instead, it is the limitation of data patterns within the domain-specific raw corpora that impacts this ability. Domain-Specific Corpora Have Fewer QAs than General Corpora. A case study on the question answering pattern is conducted. Specifically, searches are conducted over all three domain-specific corpora and a representative corpus in the general domain—RedPajama—to estimate how many sentence pairs follow the question-answering pattern by utilizing a specific mining pattern: ```text {What|Who|When|Where|Why|Which|How}{SENT1}? {SENT2} ``` \n\nThe following table lists the number of tokens belonging to the mined question-answering pairs per million tokens. Compared with the general corpora, domain-specific corpora have much lower rates of data following the question-answering pattern, which could lead to the observed decrease in question-answering ability. \n\n| Pre-train Corpora       | BioMed | Law  | Finance | General (RedPajama) |\n|-------------------------|--------|------|---------|---------------------|\n| # QA tokens            | 22.5   | 0.8 | 236.8   | **490.26**          |", "category": "Method_Disambiguation"}, {"question": "Were transformation templates manually checked for each domain during the transformation stage?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper conduct a manual evaluation of 200 reading comprehension texts per domain to assess readability and coherence?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper manually verify the effectiveness of their text transformation without compromising text integrity?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the reliability of the mining strategy used in the transformation process previously verified by other paper[1]?\n\n[1] Mozes van de Kar et al. Don’t prompt, search! mining-based zero-shot learning with language models. In EMNLP, Association for Computational Linguistics, 2022.", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the term 'input-output patterns' refer to limited data patterns in the domain-specific corpora?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the manual evaluation of transformed texts show high satisfaction with the quality of transformation across BioMed, Finance, and Law domains?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "OEL4FJMg1b", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models", "authors": ["Chong Mou", "Xintao Wang", "Jiechong Song", "Ying Shan", "Jian Zhang"], "abstract": "Despite the ability of text-to-image (T2I) diffusion models to generate high-quality images, transferring this ability to accurate image editing remains a challenge. In this paper, we propose a novel image editing method, DragonDiffusion, enabling Drag-style manipulation on Diffusion models. Specifically, we treat image editing as the change of feature correspondence in a pre-trained diffusion model. By leveraging feature correspondence, we develop energy functions that align with the editing target, transforming image editing operations into gradient guidance. Based on this guidance approach, we also construct multi-scale guidance that considers both semantic and geometric alignment. Furthermore, we incorporate a visual cross-attention strategy based on a memory bank design to ensure consistency between the edited result and original image. Benefiting from these efficient designs, all content editing and consistency operations come from the feature correspondence without extra model fine-tuning. Extensive experiments demonstrate that our method has promising performance on various image editing tasks, including within a single image (e.g., object moving, resizing, and content dragging) or across images (e.g., appearance replacing and object pasting). Code is available at https://github.com/MC-E/DragonDiffusion.", "keywords": "Diffusion model;Image editing;Image generation", "qa_pairs": [{"question": "How does the paper’s method compare with Self-Guidance, Paint-by-example, and other text-guided methods?", "answer": "The paper demonstrates its method against Self-Guidance and Paint-by-example in tasks such as object moving, appearance replacing, and object pasting.\n\n- Compared with Self-Guidance, which is a text-guided editing method, it can only achieve object-level editing due to the coarse-grained correspondence between text and image content. It is difficult to achieve editing objectives in some complex scenes or multi-object scenarios. Additionally, Self-Guidance lacks consistency constraints, leading to deviations between the editing results and the original image. In contrast, the paper’s method minimizes the impact of text. The editing results produced by different texts, whether relevant or irrelevant, are similar and meet the editing requirements.\n\n- Compared with Paint-by-example, a learnable object pasting method, it is true that such methods can produce more natural pasting effects. However, encoding the object image into multiple tokens can disrupt the object’s identity, leading to changes in the identity of the editing results. In contrast, the paper’s method adds guidance at each diffusion step, resulting in better identity consistency in the editing results.\n\n- Compared with other text-guided methods (e.g., Null-text inversion, InstructPix2Pix), although these methods can also achieve object modulation, the editing results lack reference and depend randomly on text description. Additionally, similar to Self-Guidance, text struggles to achieve a good correspondence with images in complex scenes and multi-object scenarios, leading to editing failures.", "category": "Method_Comparison"}, {"question": "How does this paper compare to previous (and contemporary) work such as DragGAN and DragDiffusion?", "answer": "The generalization ability of DragGAN is limited by the GANs. Each trained model can only edit a specific category of images, such as face, elephant, and car. Moreover, the input image cannot deviate (e.g., unaligned faces, multiple faces or cars) from the training distribution of the corresponding GAN model. Otherwise, obvious degradation will occur. The paper's method performs better than DragDiffusion without the need for training.The differences in guidance methods between the paper's approach (score-based guidance) and DragDiffusion (latent optimization): Treating $z_t$ as a learnable variable in DragDiffusion can affect the original iteration distribution of the diffusion model. The most important is that both DragGAN and DragDiffusion are designed specifically for content dragging. All in all, the paper's approach is a general and fine-grained image editing framework.", "category": "Method_Comparison"}, {"question": "What is the reason for the flickering in no-change areas, such as clouds in the sun example and the background in the apple example, and how does the balance between the energy functions $E_{edit}$ and $E_{content}$ affect this issue?", "answer": "The flickering in no-change areas is due to a trade-off between the energy functions $E_{edit}$ and $E_{content}$, which are designed for editing and content consistency, respectively. An ablation study shows that $E_{edit}$ can achieve editing with a small weight ($w_e$), and the influence of $w_c$ on $E_{edit}$ is relatively small. Therefore, a larger weight is applied to $E_{content}$ to better maintain content consistency while achieving editing goals. The flickering may also be amplified in the demo video due to directly connecting results of different editing operations.", "category": "Experimental_Analysis"}, {"question": "What is the purpose and significance of the memory bank in the proposed method?", "answer": "The memory bank is a container designed to store intermediate information ($z_t$, $K_t$, $V_t$) during the DDIM inversion process. These outputs are reused to guide the subsequent editing process. ", "category": "Method_Disambiguation"}, {"question": "How does the energy design in the proposed method work, and what numerical studies or evidence support its effectiveness?", "answer": "Energy guidance combines the generation target y with the diffusion process as q(y|$x_t$). It has been rigorously proven in [1] for generation control in diffusion models. The energy function measures the distance between the current state $x_t$ and the editing target y. The first design controls the category of the generation result by training a classifier for $x_t$, but this requires additional training and is challenging due to noise. The paper proposes taking advantage of the strong semantics of the intermediate features in pre-trained stable diffusion to transform the editing target y into the feature similarity between the editing starting point and the target point. The paper's approach eliminates training cost while achieving good editing results. Fig. 4 investigates how to make the energy guidance work better. Fig. 5 demonstrates how to combine different energy functions to achieve editing goals. The gradients generated by the content consistency ($E_{content}$) and editing ($E_{edit}$) energy functions at different time steps present a gradually converging process, and the gradient generated by each energy function mainly focuses on their respective target areas.\n\n[1] Diffusion models beat GANs on image synthesis", "category": "Method_Mechanics"}, {"question": "Why is the inference time of the proposed method slower than DragGAN, and how does the generative capability of the diffusion model compare to GANs?", "answer": "The higher inference complexity of the proposed method is due to its base model, Stable Diffusion, which requires multiple iterations compared to the single inference of GAN-based models like DragGAN. However, the generative capability of the diffusion models is significantly better than that of the GANs. For example, in terms of robustness, DragGAN requires a specific trained model when editing different categories of images, and the input image cannot deviate from the distribution of GAN's training set. DragGAN's editing results have severe degradation without image alignment. The GAN-based editing method fails in complex scenarios. In contrast, the paper's method is based on the diffusion model with high robustness to edit general images. It is worth noting that the complexity of the paper\ns method in the preparation stage is significantly lower than that of DragGAN (3.62s vs 52.40s in Tab.1). Most importantly, the paper's DragonDiffusion is a general image editing method that can accomplish a range of image editing tasks, e.g., object moving, resizing, pasting, content dragging, and appearance replacing. In contrast, DragGAN is limited to content dragging.", "category": "Method_Comparison"}, {"question": "What is the specific problem targeted by this paper, and why were diffusion feature matching, drag-style editing, memory bank, and visual cross-attention strategy introduced?", "answer": "The paper aims to build a general framework for drag-style image editing, including tasks like object moving, resizing, pasting, content dragging, and appearance replacing, which has not been achieved before.The paper chose energy guidance in diffusion models to achieve image editing. Energy guidance combines the generation target y with the diffusion process as q(y|$x_t$). Therefore, the task becomes designing a suitable energy function [1] to measure the distance between the current state x_t and the editing target y. The first generation control design in energy guidance is to control the category of the generation result. It trains a classifier for $x_t$ to measure the distance between the current state $x_t$ and the target category y. However, this type of method requires additional training, and due to the influence of noise, the classification of $x_t$ is challenging. The paper proposes taking advantage of the strong semantics in the intermediate features of pre-trained stable diffusion, which transforms the editing target y into the feature matching between the editing starting point and the target point. The paper's approach eliminates training cost while achieving good editing results. Although feature matching can achieve editing operations, the final result can not maintain consistency with the original image, such as in unedited areas. To maintain consistency, the paper uses the memory bank to collect intermediate information during the image inversion and inject it into the editing process through visual cross-attention. This allows the editing process to fully capture the original image information. Visual cross-attention  is important in maintaining content consistency. Therefore, these components are indispensable for the paper's editing framework.\n\n[1] Diffusion models beat GANs on image synthesis", "category": "Method_Disambiguation"}, {"question": "Why is the FID score higher for the proposed method compared to DragGAN?", "answer": "The FID score difference is due to the base models used: DragGAN is based on pre-trained StyleGAN, while the paper's method uses pre-trained Stable Diffusion. Stable Diffusion has significantly better generative capabilities than StyleGAN. GAN models, including StyleGAN, have limited generative capabilities, leading to severe degradation in unaligned images and poor background generation even in aligned images. ", "category": "Experimental_Analysis"}, {"question": "How were the weights ($w_e$, $w_c$) for the image editing ($E_{edit}$) and content consistency ($E_{content}$) energy functions determined, and what were the findings from the ablation study on these weights?", "answer": "An ablation study was conducted on the weights ($w_e$, $w_c$) of the image editing ($E_{edit}$) and content consistency ($E_{content}$) energy functions. The results showed that $E_{edit}$ can complete the editing operation with a small $w_e$, and the influence of $w_c$ on $E_{edit}$ is relatively small. Therefore, a larger weight was applied to $E_{content}$ to better maintain content consistency while achieving the editing goals.", "category": "Experimental_Analysis"}, {"question": "What experimental findings were obtained regarding the impact of introducing gradient guidance at different diffusion times?", "answer": "The ablation study indicates that introducing gradient guidance in the early sampling stages is crucial, as its effectiveness diminishes with progressing iterations. Additionally, using guidance information across more time steps enhances texture details. To optimize performance and complexity, guidance is introduced in the first 30 steps.", "category": "Experimental_Exposition"}, {"question": "Why does DragonDiffusion exhibit superior FID scores compared to DragDiffusion, despite DragDiffusion incorporating LoRA parameters, and how does the guidance method in DragonDiffusion ensure compatibility with the diffusion process?", "answer": "DragonDiffusion exhibits superior FID scores compared to DragDiffusion due to differences in their guidance methods. DragDiffusion optimizes a latent $z_t$ in the diffusion process through multiple iterations, but this method only considers the editing target and does not guarantee compatibility with the current time step t. In contrast, DragonDiffusion uses score-based gradient guidance to solve $p(z_{t-1}|z_{t}, y)$, where $y$ is the editing target. This method ensures compatibility with the diffusion process, as it has been proven that $p(z_{t-1}|z_{t}, y)\\sim N(\\mu+\\eta g, \\Sigma)$, where $\\eta g$ is the gradient guidance derived from an energy function $p(y|z_t)$. The energy function in DragonDiffusion is also accurate and robust in measuring feature correspondence, contributing to its superior performance. It is believed that the difference in guidance methods is one of the reasons for the performance difference. When using these two guidance methods without their content consistency design (LORA and visual cross-attention) and retaining only DDIM inversion as the generation prior of the source image, the editing effects demonstrate that the latent optimization in DragDiffusion has a deviation impact on diffusion sampling.", "category": "Experimental_Analysis"}, {"question": "Is ICLR a suitable venue for this paper, given that it involves studying feature and latent representations of diffusion models for image editing ?", "answer": "In the paper's method design, it studies the feature representation learned by the pre-trained diffusion model to design energy functions. During the editing process, the method injects the gradient guidance generated by the energy function into the discrete latent representation of the pre-trained diffusion model, achieving a general image editing framework. Therefore, the paper's method is also a study of the feature representation and latent representation of diffusion models. In addition, the Subject Areas of ICLR include applications in vision, audio, speech, language, music, etc. Therefore, the paper's work is suitable for ICLR.", "category": "Method_Disambiguation"}, {"question": "What are the key differences between DragonDiffusion and Self-Guidance[1] in terms of their approaches to image editing, particularly in how they use energy functions and modify attention layers? Additionally, how does DragonDiffusion compare to Self-Guidance in tasks like resizing and moving objects?\n\n[1] Epstein et al., Diffusion Self-Guidance for Controllable Image Generation, NeurIPS, 2023.", "answer": "The key differences lie in how editing operations are modeled. Self-Guidance calculates the similarity between text (corresponding to the object to be edited) and image features, selecting regions with high similarity for editing. This approach is coarse because:\n(1) Text can only locate content at the object level, and more accurate local content cannot be expressed in text, making it impossible to be edited with self-guidance. \n(2) In multi-object scenes, a single text corresponds to multiple objects, making it impossible to use self-guidance to edit a specific object. \n(3) Text can not provide effective content consistency, leading to deviations between the editing results and the original image. In contrast, the paper's method models editing operations based on the correspondence of intermediate features in the pre-trained diffusion model, which can achieve pixel-to-pixel correspondence, and the editing area is manually defined. Therefore, the paper's method can achieve fine-grained image editing, such as content dragging, which is not supported by self-guidance. In the paper's method, attention is used to inject content consistency guidance, resulting in better consistency between the edited results and the original image.", "category": "Method_Comparison"}, {"question": "Is the paper's primary focus on 'representation learning' as traditionally defined?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "Sx038qxjek", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing", "authors": ["Zhibin Gou", "Zhihong Shao", "Yeyun Gong", "yelong shen", "Yujiu Yang", "Nan Duan", "Weizhu Chen"], "abstract": "Recent developments in large language models (LLMs) have been impressive. However, these models sometimes show inconsistencies and problematic behavior, such as hallucinating facts, generating flawed code, or creating offensive and toxic content. Unlike these models, humans typically utilize external tools to cross-check and refine their initial content, like using a search engine for fact-checking, or a code interpreter for debugging. Inspired by this observation, we introduce a framework called CRITIC that allows LLMs, which are essentially “black boxes” to validate and progressively amend their own outputs in a manner similar to human interaction with tools. More specifically, starting with an initial output, CRITIC interacts with appropriate tools to evaluate certain aspects of the text, and then revises the output based on the feedback obtained during this validation process. Comprehensive evaluations involving free-form question answering, mathematical program synthesis, and toxicity reduction demonstrate that CRITIC consistently enhances the performance of LLMs. Meanwhile, our research highlights the crucial importance of external feedback in promoting the ongoing self-improvement of LLMs.", "keywords": "Large Language Models;In-context Learning;Self-Verification;Self-Correction;Truthfulness;Tool-use;Interaction", "qa_pairs": [{"question": " What are the novel contributions of CRITIC compared to existing methods like Self-Correct ([1], using external APIs), Self-Ask ([2], employing a search engine), and Self-Debug ([3], via a Python interpreter), which also use external feedback to correct LLM outputs?", "answer": "Before CRITIC, Self-Correct had already implemented 'tool-driven correction' in a toxicity task (more precisely, by training additional correctors). Specifically, although it did not introduce tools in math or constrained generation tasks and relied solely on intrinsic feedback from the model, it indeed used the Perspective API to train a GPT-2 corrector in the toxicity task and also used the API at inference time. In fact, when the authors proposed CRITIC, they indeed took Self-Correct as an important reference work and used it as a baseline for analysis and comparison in the paper's experiments (Sec. 4.3). Based on these consensuses, the multi-faceted contributions of CRITIC as follows: 1. Firstly, the paper proposes a unified CRITIC framework that integrates different tools and tasks into one framework, and design a series of new prompting methods that teach black-box LLMs to self-verify and self-correct through interaction with tools. This enables LLMs to have reliable self-correction capabilities through in-context learning, without relying on extra training. 2. Secondly, the paper conducts comprehensive experiments across distinct tasks, demonstrating significant performance improvements offered by CRITIC across different LLMs and model sizes. 3. Furthermore, the paper is the first to discover the unreliability of LLMs in self-correction through systematic experiments. The paper delves deeper into the main reason for the instability of Self-Correction from the perspective of uncertainty estimation, finding that LLMs struggle to 'know what they know' without external feedback (Appendix D.1). Therefore, feedback from external tool interaction is crucial for the consistent self-improvement of LLMs.\n\n[1] Welleck, Sean, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. \"Generating Sequences by Learning to Self-Correct.\" In The Eleventh International Conference on Learning Representations. 2022.\n\n[2] Press, Ofir, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. \"Measuring and narrowing the compositionality gap in language models.\" arXiv preprint arXiv:2210.03350 (2022).\n\n[3] Chen, Xinyun, Maxwell Lin, Nathanael Schärli, and Denny Zhou. \"Teaching large language models to self-debug.\" arXiv preprint arXiv:2304.05128 (2023).", "category": "Method_Disambiguation"}, {"question": "What types of errors does the interpreter feedback address in the GSM task's non-oracle setting, and what is the distribution and correction rate of these errors?", "answer": "The feedback from the interpreter is not solely about syntactic correctness. As mentioned in Sec. 4.2 and demonstrated in the cases in Appendix E.2, CRITIC leverages both 'error messages and execution results,' enabling it to fix syntax errors, correct runtime errors (e.g., timeout error in Listing 9 in Appendix E.2) and self-reflect on unreasonable reasoning steps and outputs (e.g., Listing 7 in Appendix E.2). Notably, in non-oracle settings, the paper does not only apply correction on answers where the interpreter fails, but by default, the paper allows LLMs to verify and then correct if necessary and 'stop if the executed result remains unchanged for two consecutive revisions,' as detailed in Sec. 4.2. Moreover, to offer readers a more comprehensive understanding of the specific corrections made by CRITIC and the specific benefits derived from tool feedback, the paper carried out a manual statistical analysis of the types of corrections made by CRITIC on the GSM8k full test set (1319 samples).Specifically, four different categories of initial program errors were identified: syntax errors, runtime errors, unreasonable outputs (such as irrational negative values), and other intrinsic reasoning errors. The accuracy of the initial PoT (Init) and CRITIC for each type of error was calculated. The settings for corrections were consistent with the non-oracle setting in the original paper, with up to four rounds of correction. The statistics are presented in the following table:\n\n| Error Type          | Init (Count) | Init (Acc) | CRITIC (Count) | CRITIC (Acc) |\n|---------------------|-------------|----------|---------------|----------|\n| Intrinsic Error     | 281 (77.4%) | 0.0      | 206 (83.7%)   | 26.7     |\n| Unreasonable Output | 61 (16.8%)  | 0.0      | 26 (10.6%)    | 57.4     |\n| Syntax Error        | 17 (4.7%)   | 0.0      | 11 (4.5%)     | 35.3     |\n| Runtime Error       | 4 (1.1%)    | 0.0      | 3  (1.2%)     | 25.0     |\n| All Init Errors     | 363         | 0.0      | 246           | 32.2     |\n\nAs shown in the above table:\n- (1) The majority of error types in the initial PoT responses are intrinsic reasoning errors (77.4%), such as misunderstanding the question or omitting conditions. The initial responses also exhibit a relatively high proportion (16.8%) of unreasonable output errors, while syntax and runtime errors are less frequent but not absent (5.8%). \n\n- (2) CRITIC has a high success rate in correcting unreasonable output and syntax errors (57.4% and 35.3% respectively). However, the correction rate for intrinsic errors, for which reliable feedback cannot be obtained, is relatively low (26.7%). Overall, CRITIC reduces errors in the initial erroneous samples by 32.2% in a non-oracle setting.", "category": "Method_Disambiguation"}, {"question": "What are the costs associated with calling external tools in the experiments, and how are these costs managed?", "answer": "All external tools used in the experiments are free of charge. For QA tasks, a Google Web Crawler with a caching mechanism was built, storing 9GB of search results from January to April 2023. The code[1] for this is open-sourced.The results of the Search Engine in the paper are all obtained using this code.For mathematical program synthesis tasks, a local code interpreter is used, which is free. For toxicity reduction tasks, the PERSPECTIVE API provided by Google is used, which is also free.\n\n[1]https://anonymous.4open.science/r/llm-agent-web-tools", "category": "Experimental_Setup"}, {"question": "What specific error types were identified in the initial Program of Thoughts (PoT) responses, and how effective was CRITIC in correcting these errors on the GSM8k full test set?", "answer": "The paper carried out a manual statistical analysis of the types of corrections made by CRITIC on the GSM8k full test set (1319 samples). Specifically, the paper identified four different categories of initial program errors: syntax errors, runtime errors, unreasonable outputs (such as irrational negative values), and other intrinsic reasoning errors. The paper calculated the accuracy of the initial PoT (Init), and CRITIC for each type of error. The settings for corrections are consistent with the non-oracle setting, with up to four rounds of correction. The statistics are presented in the following table:\n\n| Error Type          | Init (Count) | Init (Acc) | CRITIC (Count)| CRITIC (Acc)|   \n|---------------------|-------------|----------|-------------|------|  \n| Intrinsic Error     | 281 (77.4%) | 0.0      | 206 (71.8%) | 26.7 |  \n| Unreasonable Output | 61 (16.8%)  | 0.0      | 26 (9.1%)   | 57.4 |  \n| Syntax Error        | 17 (4.7%)   | 0.0      | 11 (3.8%)   | 35.3 |  \n| Runtime Error       | 4 (1.1%)    | 0.0      | 3  (1.0%)   | 25.0 |  \n| All Init Errors     | 363         | 0.0      | 246 (85.7%) | 32.2 |\n| Wrong Correction    | -           | 100.0    | 41 (14.3%)  | 95.7 |\n\nAs can be seen in the table:\n\n- (1) The majority of error types in the initial PoT responses are intrinsic reasoning errors (77.4%), such as misunderstanding the question or omitting conditions. The initial responses also exhibit a relatively high proportion (16.8%) of unreasonable output errors, while syntax and runtime errors are less frequent but not absent (5.8%).\n- (2) CRITIC has a high success rate in correcting unreasonable output and syntax errors (57.4% and 35.3% respectively). However, the correction rate for intrinsic errors, for which reliable feedback cannot be obtained, is relatively low (26.7%). Overall, CRITIC reduces errors in the initial erroneous samples by 32.2% in a non-oracle setting.\n- (3) Notably, while CRITIC has corrected a substantial number of errors in the initial PoT, as can be seen from the last row of the table above, there is a decrease of -4.3% in the accuracy of CRITIC on originally correct outputs. This results in the error modes after tool feedback also including 14.3% wrong corrections.", "category": "Experimental_Exposition"}, {"question": "How does the importance of different tools vary across different tasks, and what insights does this provide into the sources of improvement in the CRITIC paper?", "answer": "In knowledge-intensive tasks like commonsense QA (AmbigNQ and TriviaQA) and multi-hop knowledge reasoning (HotpotQA), web tools such as Wikipedia page browsing and Google snippet play key roles, as shown in Appendix E.1. For mathematical program synthesis tasks, external knowledge is less critical, and a code interpreter serves the function of a calculator, with feedback derived from error messages and execution results, as detailed in Appendix E.2.", "category": "Method_Mechanics"}, {"question": "Why was a sampling configuration of p=0.9 used for the toxicity reduction task in Section 4.3, while p=0.5 was used for other tasks in Sections 4.1 and 4.2?", "answer": "A sampling configuration of p=0.9 was used for the toxicity reduction task in Section 4.3 because it aligns with the standard set in the RealToxicityPrompts benchmark [1]. This ensures consistency with the methodology employed across all baseline studies [2], enabling a fair comparison.\n\n[1] Gehman, Samuel, et al. \"Realtoxicityprompts: Evaluating neural toxic degeneration in language models.\" arXiv preprint arXiv:2009.11462 (2020).\n\n[2] Welleck, Sean, et al. \"Generating sequences by learning to self-correct.\" arXiv preprint arXiv:2211.00053 (2022).", "category": "Motivation_Analysis"}, {"question": "How does the CRITIC framework compare with other frameworks that endow language models with self-correction and tool use, given that the ideas of natural language feedback[1, 2, 3, 4] and tool use[5,6] are not new?\n\n[1] https://arxiv.org/abs/2204.14146\n\n[2] https://arxiv.org/abs/2303.11366\n\n[3] https://arxiv.org/abs/2212.08073\n\n[4] https://arxiv.org/abs/2303.16749\n\n[5] https://arxiv.org/abs/2207.14502\n\n[6] https://arxiv.org/abs/2302.04761\n", "answer": "CRITIC is novel, and more than a trivial blend of existing works on self-correction [1, 2, 3] and tool use [4,5]. While these are crucial fast-moving areas of LLMs, CRITIC offers unique findings, integrated frameworks, and valuable insights, making CRITIC a distinct contribution. The relevant papers in these two fields actually have very different research perspectives and stances from CRITIC:\n\n **Intrinsic Self-Correct with NL feedback**: Works include Self-Critique [6], CAI [7], Reflexion[1], Self-Refine [2], and others [8-10], they prompt or train language models to correct their results. In contrast, the paper's study is the first to demonstrate that such a 'Self-Verification and Self-Correction' approach has proven to be remarkably unreliable across diverse tasks and various LLMs (Sec. 4 and Appendix D.1). Specifically, modest improvements or even deterioration are observed universally using self-correct without external feedback. Consequently, CRITIC emphasizes the importance of feedback from external interactions for the consistent self-improvement of LLMs. The proposed framework is general and has proven effective across multiple tasks. These profound reflections on unreliable self-correction of LLMs and crucial findings on the importance of external feedback can be important learnings for the community. \n\n**The Unreliability of Self-correction**: Moreover, CRITIC further delves into the core reason behind the unreliability of self-verification from the perspective of uncertainty estimation, as shown in Appendix D.1. Essentially, the models are incapable of accurately identifying 'what they know' (i.e., LLMs don't know what they know) [8] without relying on external tools. Therefore, without the aid of oracle verification (employed in many contemporary works such as Reflexion [1] and Self-Refine [2]), self-correction might surprisingly deteriorate performance for many tasks, even worsening the initial answer (as demonstrated in Table 2, 3 under CRITIC w/o Tool, and in Table 8 under Self-Refine). A nice recent follow-up work to CRITIC, titled 'LLMs cannot self-correct reasoning yet' [9] extends the study of CRITIC (which cites CRITIC numerous times). It further validates and expands the paper's findings on the unreliability of Self-Verify and Self-Correct in CRITIC using GPT-4 in more settings on many reasoning tasks. \n\n**Tool-Use**: Another category of related work is Tools Augmented LMs, such as ReAct [4], PoT [5], and Toolformer [10]. These works significantly differ from CRITIC as they focus on tool learning. None of them consider using tool-interaction as faithful feedback to iteratively improve the model, which is the key component of CRITIC. In conclusion, CRITIC is the first to unveil the unreliability of self-verification and self-correction across diverse tasks and LLMs of various families and sizes. By initially highlighting the challenges LLMs encounter in self-verification [8], self-correction [6,7], the paper's goal is to rectify any potential overestimations of these LLM's abilities within the research community [1-3]. Feedback from external tool interaction is crucial for consistent self-improvement of LLMs.\n\n[1] Shinn, Noah, et al. \"Reflexion: Language agents with verbal reinforcement learning.\" Thirty-seventh Conference on Neural Information Processing Systems. 2023.\n\n[2] Madaan, Aman, et al. \"Self-refine: Iterative refinement with self-feedback.\" Thirty-seventh Conference on Neural Information Processing Systems. 2023.\n\n[3] Chen, Xinyun, et al. \"Teaching large language models to self-debug.\" arXiv preprint arXiv:2304.05128 (2023).\n\n[4] Yao, Shunyu, et al. \"React: Synergizing reasoning and acting in language models.\" arXiv preprint arXiv:2210.03629 (2022).\n\n[5] Chen, Wenhu, et al. \"Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks.\" arXiv preprint arXiv:2211.12588 (2022).\n\n[6] Saunders, William, et al. \"Self-critiquing models for assisting human evaluators.\" arXiv preprint arXiv:2206.05802 (2022).\n\n[7] Bai, Yuntao, et al. \"Constitutional ai: Harmlessness from ai feedback.\" arXiv preprint arXiv:2212.08073 (2022).\n\n[8]Kadavath, Saurav, et al. \"Language models (mostly) know what they know.\" arXiv preprint arXiv:2207.05221 (2022).\n\n[9]Huang, Jie, et al. \"Large language models cannot self-correct reasoning yet.\" arXiv preprint arXiv:2310.01798 (2023).\n\n[10]Schick, Timo, et al. \"Toolformer: Language models can teach themselves to use tools.\" arXiv preprint arXiv:2302.04761 (2023).", "category": "Method_Comparison"}, {"question": "Does the statistical analysis of corrections made by CRITIC on the GSM8k test set show that the majority of initial program errors are syntax errors?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does CRITIC correct only syntax errors and runtime errors in the non-oracle setting, as suggested by the improvements?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the manual statistical analysis reveal that CRITIC has a higher success rate in correcting intrinsic reasoning errors compared to syntax and runtime errors?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the feedback from the interpreter in the GSM task limited to syntactic correctness in a non-oracle setting?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "sllU8vvsFF", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "LRM: Large Reconstruction Model for Single Image to 3D", "authors": ["Yicong Hong", "Kai Zhang", "Jiuxiang Gu", "Sai Bi", "Yang Zhou", "Difan Liu", "Feng Liu", "Kalyan Sunkavalli", "Trung Bui", "Hao Tan"], "abstract": "We propose the first Large Reconstruction Model (LRM) that predicts the 3D model of an object from a single input image within just 5 seconds. In contrast to many previous methods that are trained on small-scale datasets such as ShapeNet in a category-specific fashion, LRM adopts a highly scalable transformer-based architecture with 500 million learnable parameters to directly predict a neural radiance field (NeRF) from the input image. We train our model in an end-to-end manner on massive multi-view data containing around 1 million objects, including both synthetic renderings from Objaverse and real captures from MVImgNet. This combination of a high-capacity model and large-scale training data empowers our model to be highly generalizable and produce high-quality 3D reconstructions from various testing inputs, including real-world in-the-wild captures and images created by generative models. Video demos and interactable 3D meshes can be found on our LRM project webpage: https://yiconghong.me/LRM.", "keywords": "3D Reconstruction;Large-Scale Training;Transformers", "qa_pairs": [{"question": "How does the LRM model address the challenge of the one-to-many mapping problem in single-image-to-3D tasks, and does it merely fit to the training distribution rather than learning a 3D distribution?", "answer": "LRM is a deterministic model that learns the distribution of 3D data during training and fits the reference image to this learned distribution during inference. The paper simplified the one-to-many mapping problem to a one-to-one mapping, which has been shown to be effective in many cases (evidenced in Figure 2, Figure 6, and Figure 7). Additionally, LRM bridges 2D and 3D representations, potentially enabling one-image-to-many-3D-shapes generation through training latent diffusion models on LRM-predicted triplanes. For instance, training latent diffusion models on triplanes predicted by LRM to generate new shapes or reconstruct 3D models from text-to-multiview generations.", "category": "Method_Disambiguation"}, {"question": "What are the specific implementation details regarding the number of attention heads and how the attention mechanism operates between image features [32x32, 768] and embeddings [32x32x3, 1024] in the cross-attention layer of LRM's image-to-triplane decoder?", "answer": "LRM's image-to-triplane decoder uses 16 attention heads in all multi-head attention. In the cross-attention layer, positional embeddings are applied as queries (Q), and image features are applied as keys (K) and values (V), all projected to the same dimension $d_{D}=1024$. Each of the 32x32x3=3072 positional embedding tokens queries all 32x32=1024 image feature tokens, resulting in a weight shape of [32x32x3, 32x32]. When multiplied by the values [32x32, 1024], it produces output tokens of dimension [32x32x3, 1024], which matches the input dimension of the next layer.", "category": "Method_Mechanics"}, {"question": "What computational resources are required for training and inference of the proposed LRM, and is there a smaller version available with reduced resource demands?", "answer": "Like many other recent visual/language models (e.g., CLIP [1], GPT-3 [2], and Stable Diffusion [3]), large-scale training is widely applied to stabilize the optimization processes of large models and achieve the best results at the cost of high computational demand. . The paper's best-performing LRM was trained on 128 NVIDIA (40G) A100 GPUs for 3 days, with inference requiring a single 11G GPU. Additionally, a smaller version of LRM, using the same architecture as the baseline model in Appendix C Table 1, was trained on 8 NVIDIA (40G) A100 GPUs for 4 days with batch size 64 and gradient accumulation of 4 steps, achieving performance between the baseline and the final models.\n\n[1] Learning Transferable Visual Models From Natural Language Supervision. Radford et al., 2021.\n\n[2] Language Models are Few-Shot Learners. Brown et al., 2020.\n\n[3] High-Resolution Image Synthesis with Latent Diffusion Models. Rombach et al., 2021.", "category": "Experimental_Setup"}, {"question": "What quantitative comparisons were made between the proposed method (LRM) and prior state-of-the-art methods (Point-E[1], Shap-E[2], and One-2-3-45[3]) in terms of novel view synthetic quality and geometric quality?\n\n[1] Point-E: A System for Generating 3D Point Clouds from Complex Prompts. Nichol et al., 2022.\n\n[2] Shap-E: Generating Conditional 3D Implicit Functions. Jun and Nichol, 2023.\n\n[3] One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization. Liu et al., 2023.", "answer": "Point-E trains an image-to-3D point cloud diffusion model, Shap-E encodes point clouds to latent representations and trains a diffusion model on the latents to generate parameters of a 3D implicit function, and One-2-3-45 reconstructs multiview images generated with a 2D diffusion model. Randomly selected 100 objects from the Google Scanned Objects (GSO) dataset, the novel view synthetic quality of 20 reference views (FID, CLIP-Similarity, PSNR, LPIPS) and the geometric quality (Chamfer Distance) were measured, as shown in the Table below. It can be seen that the LRM consistently outperforms previous approaches in all metrics.\n\n|    | FID \\(\\downarrow\\) | CLIP-Similarity \\(\\uparrow\\) | PSNR \\(\\uparrow\\) |  LPIPS \\(\\downarrow\\) |  Chamfer Distance \\(\\downarrow\\) |\n| :---                |   :----:       |    :----:   |      :----:   |     :----:   |    ---: |   \n| Point-E        | 123.70  | 0.741  | 15.60  |  0.308  |  0.099  |\n| Shap-E        | 97.05  | 0.805   | 14.36 |  0.289 | 0.085 |\n| One-2-3-45  |  139.24  |  0.713  |  12.42  |  0.448  | 0.123  |\n| LRM (ours)  | **31.44**  | **0.902**  |  **19.60**  |  **0.163**  |  **0.053**  |", "category": "Method_Comparison"}, {"question": "What causes the issue of averaging modes in the texture of occluded portions of shapes in the proposed method?\n\n", "answer": "The issue of averaging modes in the texture of occluded portions of shapes is caused by the joint effect of LRM's deterministic approach, an L2 loss, and the use of MVImgNet data for training, which covers only 180 degrees of view. A similar issue occurs in Masked Autoencoders[1], where the model reconstructs blurry contents when a large block of region is removed from an image. A potential solution is to post-optimize the LRM's output using generative prior and SDS loss, as proposed by DreamFusion[2], though this incurs extra processing time.\n\n[1] Masked Autoencoders Are Scalable Vision Learners. He et al., 2021.\n[2] DreamFusion: Text-to-3D using 2D Diffusion. Poole et al., 2022.", "category": "Experimental_Analysis"}, {"question": "Has the LRM model been extended to perform sparse view reconstruction, and if so, how does it handle sparse multi-view inputs?", "answer": "Yes, the LRM model has been extended to accept sparse multi-view inputs. This is achieved by passing encoded multi-view image features to the image-to-triplane decoder and using positional embeddings to query features of all images via cross-attention to construct the triplane. This approach results in high-fidelity shapes and reduces blurry occluded portions due to better 3D object coverage.", "category": "Method_Mechanics"}, {"question": "Is it reasonable to solve the inherently probabilistic task of novel view synthesis using a discriminative approach, given the method's ability to scale up?", "answer": "Similar to the classic PixelNerf [1] and Pixel2Mesh [2] and many other deterministic approaches, LRM considers a simplified assumption of one-to-one mapping between image and 3D. Intuitively, this line of work focuses on reconstructing the average shape given partial 2D information, just as humans can naturally infer a probable 3D model of an object from just a single view. It is believed that the definition of single-image-to-3D task is not necessarily generative but depends on the specific application. One important goal of the paper is to justify the plausibility of large-scale training for 3D, which is very under-explored in previous research. Moreover, the paper's LRM learns generic 3D prior, and it bridges 2D and 3D representations that can potentially facilitate 3D generation in the future. For instance, training latent diffusion models on triplanes predicted by LRM to generate new shapes or reconstruct 3D models from text-to-multiview generations.\n\n[1] PixelNeRF: Neural Radiance Fields from One or Few Images. Yu et al., 2021.\n\n[2] Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images. Wang et al., 2018.", "category": "Method_Disambiguation"}, {"question": "What distinguishes the proposed LRM method from being merely an engineering effort?", "answer": "The paper's proposed LRM is inspired and based on many previous successes in computer vision, such as image representation learning (DINO), inter-modal modeling (transformer-based cross-attention), 3D representation (triplane-NeRF), and the idea of large-scale training. But the effort is not mainly coming from engineering efforts, as there are enormous alternative design choices, while LRM was built from considerate research insights; the paper's method successfully integrates those components into a scalable and end-to-end trainable system, justifying the plausibility of large-scale 3D training. Distinctively, LRM is a feed-forward network that does not rely on any generative prior, and it considers 2D images and 3D shapes as bridgeable modalities and models their correspondence via cross-attention without explicitly defining any spatial alignment. LRM is more data-friendly and efficient than recent large-scale 3D models such as Shap-E [1] and Point-E [2], and it is the state-of-the-art single-image-to-3D approach that produces very high-fidelity results. The methods and ideas presented in this paper carry huge research value that can impact future 3D deep learning studies.\n\n[1] Point-E: A System for Generating 3D Point Clouds from Complex Prompts. Nichol et al., 2022.\n\n[2] Shap-E: Generating Conditional 3D Implicit Functions. Jun and Nichol, 2023.", "category": "Method_Disambiguation"}, {"question": "What quantitative evaluations were conducted to compare the proposed method (LRM) against state-of-the-art methods such as Point-E[1], Shap-E[2], and One-2-3-45[3] for novel view synthesis and 3D reconstruction?\n[1] Point-E: A System for Generating 3D Point Clouds from Complex Prompts. Nichol et al., 2022.\n\n[2] Shap-E: Generating Conditional 3D Implicit Functions. Jun and Nichol, 2023.\n\n[3] One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization. Liu et al., 2023.\n", "answer": "Point-E trains an image-to-3D point cloud diffusion model, Shap-E encodes point clouds to latent representations and trains a diffusion model on the latents to generate parameters of a 3D implicit function, and One-2-3-45 reconstructs multiview images generated with a 2D diffusion model. 100 objects were randomly selected from the Google Scanned Objects (GSO) dataset. The novel view synthetic quality of 20 reference views (measured by FID, CLIP-Similarity, PSNR, LPIPS) and the geometric quality (measured by Chamfer Distance) were evaluated, as shown in the Table below. The paper’s LRM consistently outperforms previous approaches in all metrics.\n\n|    | FID $\\downarrow$ | CLIP-Similarity $\\uparrow$ | PSNR $\\uparrow$ |  LPIPS $\\downarrow$ |  Chamfer Distance $\\downarrow$ |\n| :---                |   :----:       |    :----:   |      :----:   |     :----:   |    ---: |   \n| Point-E        | 123.70  | 0.741  | 15.60  |  0.308  |  0.099  |\n| Shap-E        | 97.05  | 0.805   | 14.36 |  0.289 | 0.085 |\n| One-2-3-45  |  139.24  |  0.713  |  12.42  |  0.448  | 0.123  |\n| LRM (the paper's)  | **31.44**  | **0.902**  |  **19.60**  |  **0.163**  |  **0.053**  |\n", "category": "Method_Comparison"}, {"question": "Why does LRM struggle to synthesize plausible appearances for objects with complex, asymmetric patterns on the occluded side(e.g., Figure 2. Giraffe, supp. website shoe example)?", "answer": "LRM struggles due to its deterministic approach, which produces averaged modes of the unseen, combined with the use of L2 loss and MVImgNet data that covers only 180 degrees of view. This issue is similar to that seen in Masked Autoencoders, where large region removal results in blurry reconstructions. A potential solution is to post-optimize the LRM's output (e.g., using generative prior and SDS loss proposed by DreamFusion [2]) at the cost of extra processing time.\n\nOn the other hand, a research found that LRM can be easily extended to accept sparse multi-view inputs by passing encoded multi-view image features to the image-to-triplane decoder, and making the positional embeddings query from features of all images via cross-attention to construct the triplane. This model can create very high-fidelity shapes where the issue of blurry occluded portions can be largely eliminated due to the higher coverage of the 3D object. \n\n[1] Masked Autoencoders Are Scalable Vision Learners. He et al., 2021.\n\n[2] DreamFusion: Text-to-3D using 2D Diffusion. Poole et al., 2022.", "category": "Experimental_Analysis"}, {"question": "How does the proposed LRM method quantitatively compare to state-of-the-art methods (Point-E[1], Shap-E[2], and One-2-3-45[3]) in terms of novel view synthetic quality and geometric quality?\n\n[1] Point-E: A System for Generating 3D Point Clouds from Complex Prompts. Nichol et al., 2022.\n\n[2] Shap-E: Generating Conditional 3D Implicit Functions. Jun and Nichol, 2023.\n\n[3] One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization. Liu et al., 2023.", "answer": "Point-E trains an image-to-3D point cloud diffusion model, Shap-E encodes point clouds to latent representations and trains a diffusion model on the latents to generate parameters of a 3D implicit function, and One-2-3-45 reconstructs multiview images generated with a 2D diffusion model. 100 objects were randomly selected from the Google Scanned Objects (GSO) dataset. The novel view synthetic quality of 20 reference views (measured by FID, CLIP-Similarity, PSNR, LPIPS) and the geometric quality (measured by Chamfer Distance) were evaluated, as shown in the Table below. The paper’s LRM consistently outperforms previous approaches in all metrics.\n\n|    | FID $\\downarrow$ | CLIP-Similarity $\\uparrow$ | PSNR $\\uparrow$ |  LPIPS $\\downarrow$ |  Chamfer Distance $\\downarrow$ |\n| :---                |   :----:       |    :----:   |      :----:   |     :----:   |    ---: |   \n| Point-E        | 123.70  | 0.741  | 15.60  |  0.308  |  0.099  |\n| Shap-E        | 97.05  | 0.805   | 14.36 |  0.289 | 0.085 |\n| One-2-3-45  |  139.24  |  0.713  |  12.42  |  0.448  | 0.123  |\n| LRM (the paper's)  | **31.44**  | **0.902**  |  **19.60**  |  **0.163**  |  **0.053**  |", "category": "Method_Comparison"}, {"question": "What sampling method is used for the 128 points in volume rendering: uniform sampling, a two-stage sampling strategy, or another method?", "answer": "The LRM uniformly samples 128 points for each ray in neural rendering. While uniform sampling with perturbation and two-stage coarse-to-fine sampling were also tested, no significant difference in results was observed. ", "category": "Experimental_Setup"}, {"question": "What causes the blurriness and inconsistent color tone/saturation on the back side of the model (e.g., figure 3 penguin), and what are the potential solutions to address this issue?", "answer": "The blurriness and inconsistent color tone/saturation on the back side of the model are caused by the deterministic nature of the LRM model, which produces averaged modes of the unseen regions, and the use of L2 loss along with MVImgNet data covering only 180 degrees of view. A similar issue is observed in Masked Autoencoders[1], where blurry reconstructions occur when large regions are removed. Potential solutions include post-optimizing the LRM's output using generative prior and SDS loss (as proposed by DreamFusion[2]), though this would require extra processing time. Additionally, research suggests that LRM can be easily extended to accept sparse multi-view inputs by passing encoded multi-view image features to the image-to-triplane decoder, and making the positional embeddings query from features of all images via cross-attention to construct the triplane. This model can create very high-fidelity shapes where the issue of blurry occluded portions can be largely eliminated due to the higher coverage of the 3D object.\n\n[1] Masked Autoencoders Are Scalable Vision Learners. He et al., 2021. \n[2] DreamFusion: Text-to-3D using 2D Diffusion. Poole et al., 2022.", "category": "Experimental_Analysis"}, {"question": "How does the model handle camera parameter normalization and object orientation in the training datasets (Objaverse and MvImgNet), and how does this compare to the general approach of assuming the world coordinate as the camera coordinate(extrinsic = np.eye(4))?", "answer": "The model relies on correct camera parameters and normalizes the cameras during training. The paper uses 3D data in Objaverse (3D models) and MvImgNet (videos of objects) for training; the majority of objects in Objaverse have a certain orientation, but objects in MvImgNet do not. As specified in Section 4.1, to render multiviews from objects in Objaverse, the 3D shape is normalized to the box $[-1,1]^{3}$ in world space, and 32 random views are rendered with the same camera pointing toward the world center at arbitrary poses. The camera poses are sampled from a ball of radius $[1.5, 3.0]$ and with height in range $[0.75, 1.60]$. During training, the camera is normalized before feeding an image to LRM (details in Section 4.2). Specifically, regardless of the initial position of the camera, its poses are normalized to position $[0,-2,0]$, with the camera's vertical axis aligned with the upward z-axis in the world frame (analogy to the case of extrinsic=[[1,0,0,0],[0,1,0,-2],[0,0,1,0],[0,0,0,1]]). Such normalization does not alter the input image, so it actually rotates and scales the object accordingly. This means that regardless of where the input image is taken from, it will be projected onto the triplane from exactly the same direction. As discussed in Section 3.2, this method greatly reduces the optimization space of triplane features and facilitates model convergence. Finally, at inference, LRM assumes the unknown camera parameters to be the normalized cameras that were applied to train the Objaverse data (Section 4.2).\n", "category": "Experimental_Setup"}, {"question": "How does the reliance on multiview data in the LRM approach impact the incorporation of diversity from single-view image sources?", "answer": "The LRM training requires multiview data, but the growing size and diversity of 3D datasets, like Objaverse-XL with 10.2 million 3D assets, and potential multiview supervision from videos, suggest ample multiview data for large-scale 3D learning. While single-view images are more numerous, recent single-view-based approaches like 3DGP[1] still struggle with consistency and completeness. \n\n[1] 3D Generation on ImageNet. Skorokhodov et al., 2023.", "category": "Method_Disambiguation"}, {"question": "What are the key factors contributing to the superior performance of LRM compared to Point-E[1] and Shap-E[2], and is this due to the larger datasets (Objaverse and MvImgNet) used by LRM or the design of its learnable positional encoding and other architectural features?\n\n[1] Point-E: A System for Generating 3D Point Clouds from Complex Prompts. Nichol et al., 2022. \n[2] Shap-E: Generating Conditional 3D Implicit Functions. Jun and Nichol, 2023.", "answer": "Point-E trains an image-to-3D point cloud diffusion model, and Shap-E encodes point clouds to latent representations and trains a diffusion model on the latents to generate parameters of a 3D implicit function. Both models contain hundreds of millions of learnable parameters and are trained with several million 3D assets (unknown data source and unknown computational cost). In terms of the network and dataset sizes, the paper's LRM has 500 million learnable parameters, and it is trained on 1 million 3D data, which does not show an advantage. In terms of the network architecture, Point-E, Shap-E, and LRM all use transformer-based models and apply cross-attention for inter-modality modeling (i.e., image-to-point cloud, point cloud+image-to-3D latents, and image-to-triplane). It is the choice of very compact and expressive triplane representation together with an end-to-end trainable framework that enables the scaling of the LRM and its adequate learning on large datasets (Objaverse and MvImgNet). Compared to the unstructured point cloud representation applied in Point-E and Shap-E, LRM applies the structured triplane representation that is aligned with the world frame, which naturally facilitates 2D-to-3D projection. It is also worth mentioning that Point-E uses 4K points (as tokens) and Shap-E uses 16K points (as tokens), but the paper's LRM only uses $3{\\times}32{\\times}32{=}3072$ triplane tokens, which largely reduces the modeling complexity. Additionally, compared to the two-stage approach in Shap-E, which attempts to generate latents that can produce the parameters of implicit 3D functions through a diffusion model, LRM directly maps 2D images to triplanes, which should be much more stable and efficient to learn. Overall, LRM is a more data-friendly and efficient model than Point-E and Shap-E.", "category": "Method_Comparison"}]}
{"id": "8Wuvhh0LYW", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models", "authors": ["Wenqi Shao", "Mengzhao Chen", "Zhaoyang Zhang", "Peng Xu", "Lirui Zhao", "Zhiqian Li", "Kaipeng Zhang", "Peng Gao", "Yu Qiao", "Ping Luo"], "abstract": "Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements. Although recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization. To tackle this issue, we introduce an Omnidirectionally calibrated Quantization ($\\textbf{OmniQuant}$) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters. OmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights. Operating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization. For instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples. Extensive experiments validate OmniQuant's superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices. Codes are available at \n\\url{https://github.com/OpenGVLab/OmniQuant}.", "keywords": "Large Language Model Compression;Differentiable Quantization", "qa_pairs": [{"question": "What is the individual and cumulative impact of LET (Layer-wise Error Tracking) on each linear layer, and how does it affect the model's performance across different layers?", "answer": "The authors excluded the LET of the second linear layer due to unstable gradients caused by high sparsity of features after the non-linear layer. They evaluated four LET pairs: [ln1, (q_proj, k_proj, v_proj)], [v_proj, out_proj], [Q, K], and [ln2, fc1]. By removing the LET from each layer and measuring the average perplexity and accuracy on 6 zero-shot tasks, they found that all four LETs improve performance, particularly the [ln1, (q_proj, k_proj, v_proj)] pair. These results indicate that activation outliers are more severe after layer normalization layers. The experiment is detailed in Table A5 of the paper.\n\n| LLaMa-1-7B                              | Average PPL | Average Acc. |\n| --------------------------------------- | ----------- | ------------ |\n| FP                                      | 6.38        | 64.09        |\n| W4A4 (ours)                             | 12.87       | 52.65        |\n| W4A4  - [ln1, (q_proj, k_proj, v_proj)] | 19.87       | 46.79        |\n| W4A4  -[v_proj, out_proj]               | 13.03       | 51.68        |\n| W4A4  -[Q, K]                           | 13.34       | 51.47        |\n| W4A4  -[ln2, fc1]                       | 14.47       | 51.04        |\n", "category": "Experimental_Exposition"}, {"question": "How does OmniQuant impact the reasoning ability of large language models (LLMs) as evaluated by the MMLU benchmark?", "answer": "OmniQuant consistently achieves comparable or superior performance in reasoning ability on the MMLU benchmark,  as shown in the following table.\n\n| LLaMa-1-7B (FP: 38.41 ) | W4A16g128 | W3A16g128 | W2A16g128 | W4A4  |\n| ----------------------- | --------- | --------- | --------- | ----- |\n| RTN                     | 37.37     | 33.43     | 22.55     | 23.31 |\n| GPTQ                    | 35.39     | 30.53     | 23.83     | -     |\n| AWQ                     | 37.71     | 35.43     | 22.58     | -     |\n| Outlier Suppression+    | -         | -         | -         | 25.72 |\n| OmniQuant               | 37.50     | 35.60     | 26.03     | 26.93 |", "category": "Experimental_Exposition"}, {"question": "How does OmniQuant perform when quantizing stronger chatbots like Vicuna-v1.5 and evaluating them on benchmarks such as MT-Bench?", "answer": "OmniQuant consistently outperforms RTN and AWQ when quantizing Vicuna-v1.5 and evaluating through MT-Bench, with a win rate of 83.4% against RTN and 51.8% against AWQ.\n\n| Vicuna-v1.5-7B-W3A16g128 | OmniQuant Win | Tie  | OmniQuant Lost | Win rate |\n| ------------------------ | ------------- | ---- | -------------- | -------- |\n| Omniquant v.s. RTN       | 101           | 239  | 20             | 83.4 %   |\n| Omniquant v.s. AWQ       | 70            | 225  | 65             | 51.8%    |", "category": "Experimental_Exposition"}, {"question": "How does OmniQuant compare to SpQR[1] and SqueezeLLM[2] in terms of performance and quantization support?\n\n[1] Dettmers, T., Svirschevski, R., Egiazarian, V., Kuznedelev, D., Frantar, E., Ashkboos, S., Borzunov, A., Hoefler, T. and Alistarh, D., 2023. SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression. arXiv preprint arXiv:2306.03078.\n[2] Kim, S., Hooper, C., Gholami, A., Dong, Z., Li, X., Shen, S., Mahoney, M.W. and Keutzer, K., 2023. SqueezeLLM: Dense-and-Sparse Quantization. arXiv preprint arXiv:2306.07629.", "answer": "OmniQuant has been compared to SpQR and SqueezeLLM in Section A7. While OmniQuant may perform slightly worse than SqueezeLLM in some scenarios, it focuses on uniform (INT) quantization, which offers simplicity, flexibility, and strong hardware support. Additionally, OmniQuant achieves competitive performance in both weight-only quantization and weight-activation quantization, whereas SpQR and SqueezeLLM only support weight-only quantization. This distinction provides valuable context to the comparison.", "category": "Method_Comparison"}, {"question": "What is the novelty of the learnable approach in OmniQuant compared to the gradient-independent learning scheme used in Outlier Suppression+?", "answer": " The novelty of OmniQuant's learnable approach lies in using differential gradient updates for various quantization parameters, which significantly reduces quantization error compared to the grid-searching method in Outlier Suppression+. Figure A3 illustrates this, and table 2 in the paper demonstrates the superiority of OmniQuant to OS+.", "category": "Method_Comparison"}, {"question": "What are the key differences between OmniQuant and Outlier Suppression+ (OS+) in terms of parameter optimization, block-wise optimization, equivalent transformation, and weight clipping, and what experimental results support these differences?", "answer": "OmniQuant differs from Outlier Suppression+ (OS+) in several key aspects: (1) Parameter Optimization: OmniQuant uses gradient updates to optimize both shifting and scaling parameters, while OS+ relies on pre-defined shifting strength and grid-searched scaling parameters. (2) Block-wise Optimization: OmniQuant leverages block-wise optimization while OS+ executes grid searching through a mix of single linear layer and multiple linear layers objective. Block-wise optimization considers interaction within a block and can produce better performance as demonstrated in BRECQ. However, it is a challenge for OS+ to expand to block-wise quantization, which would significantly enlarge the solution space and increase the time of grid searching. (3) Equivalent Transformation: OmniQuant extends equivalent transformation to attention operations, while OS+ only applies it to linear layers. (4) Weight Clipping: OmniQuant introduces learnable weight clipping (LWC), enabling low-bits (2,3) weight-only quantization, whereas OS+ supports only weight-activation quantization. These differences result in OmniQuant achieving higher average accuracy across various LLaMa models (7B, 13B, 30B, 65B) compared to OS+, as demonstrated in Table 2.An overview is as follows:\n\n| Model          | Method     | Average Acc. |\n|----------------|------------|--------------|\n| LLaMa-1-7B     | OS         | 48.43        |\n| LLaMa-1-7B     | OmniQuant  | 52.65        |\n| LLaMa-1-13B    | OS         | 49.86        |\n| LLaMa-1-13B    | OmniQuant  | 54.37        |\n| LLaMa-1-30B    | OS         | 52.62        |\n| LLaMa-1-30B    | OmniQuant  | 56.63        |\n| LLaMa-1-65B    | OS         | 52.52        |\n| LLaMa-1-65B    | OmniQuant  | 59.23        |\n", "category": "Method_Comparison"}, {"question": "How does OmniQuant address the concern of overfitting when optimized with limited data, and what evidence is provided to demonstrate its generalization ability?", "answer": "OmniQuant mitigates overfitting by focusing optimization on LET and LWC rather than all weights, thus constraining the solution space. Evidence of its generalization ability is shown in Table A11, where OmniQuant performs superiorly even with only 16 calibration samples, and through its performance on the MMLU benchmark.\n\n| LLaMa-1-7B (FP: 38.41 )       | W4A16g128 | W4A16g128 | W3A16g128 |  W4A4  |\n|-------------|----------------------|-----------|-----------|-----------|\n| RTN                | 37.37     | 33.43     | 22.55     | 23.31 |\n|  GPTQ                | 35.39     | 30.53     | 23.83     | -     |\n|  AWQ                 | 37.71     | 35.43     | 22.58     | -     |\n| Outlier Suppression+ | -         | -         | -         | 25.72 |\n| OmniQuant           | 37.50     | 35.60     | 26.03     | 26.93 |\n", "category": "Experimental_Analysis"}, {"question": "What is the motivation and effectiveness of the weight clipping method (LWC) in the proposed approach, and how does it address the challenges posed by varying weight ranges during training?", "answer": "LWC reduces the optimization difficulty in learning clipping thresholds because LWC reparameterizes the learnable clipping threshold into the product of a learnable 0~1 clipping strength and the maximum or minimum of weights. The benefit is that no matter how the maximum and minimum of weights change caused by LET with the training going, LWC would always learn a clipping strength under the reference of the current maximum and minimum of weights. Such a reparameterization technique has been utilized in various areas for better training properties. For example, batch normalization decomposes the neural activation into a variable approximately following normal distribution and learnable scale and shift parameters, which improves the training stability. Figure A5 in the appendix presents the dynamic change in weight ranges (maximum minus minimum) during training with different clipping-based methods. It can be seen that the weight range varies a lot as the training goes on (see the curve of Min-Max) when training the paper's OmniQuant, which is caused by the paper's proposed another component LET. It poses a challenge for traditional absolute clipping thresholds (PACT) or step size learning (LSQ). In contrast, LWC calculates relative clipping strength and enables it to adapt to varying weight ranges, as shown in Figure A5 (c).", "category": "Method_Disambiguation"}, {"question": "How does the proposed Eq. (2) address the optimization difficulty of learning clipping thresholds for outliers, as described in Section 3.2, and what is the advantage of this approach over traditional methods like PACT and LSQ?", "answer": "The proposed LWC reparameterizes the learnable clipping threshold into the product of a learnable clipping strength (0~1) and the maximum or minimum of weights. This reparameterization allows LWC to adapt to varying weight ranges caused by LET, reducing optimization difficulty. Unlike traditional methods like PACT and LSQ, which retain a small weight range and clip many regular weight values, LWC dynamically calculates relative clipping strength and adapts to the weight range, ensuring better optimization, as demonstrated in Figure A5.", "category": "Method_Mechanics"}, {"question": "How does the proposed LET method differ from Outlier Suppression+ (OS+) and AWQ in terms of equivalent transformation, execution position, and optimization strategy, and why is LET considered superior?", "answer": "The proposed LET method differs from Outlier Suppression+ (OS+) and AWQ in several ways. For equivalent transformation, AWQ only considers scaling operations, while OS+ and LET consider both scaling and shifting operations. In terms of execution position, AWQ and OS+ apply equivalent transformations only to linear layers, whereas LET also considers matrix multiplication within attention (Eq. (5) in the paper), which enlarges the solution space and improves performance. For optimization, OS+ uses pre-defined shifting, and both AWQ and OS+ find scaling factors through grid searching based on heuristic strategies. In contrast, LET optimizes all equivalent transformation parameters through end-to-end gradient descent, significantly improving performance. Comparative experiments in Table 2 demonstrate the superiority of LET over OS+.\n\n|  | Operation         | Position                 | Optimization                                         |\n| :---- | :---------------- | :----------------------- | :--------------------------------------------------- |\n| AWQ   | scale             | linear layer             | grid searching                                       |\n| OS+   | scale & shift     | linear layer             | grid searching for scaling & pre-defined for shifting |\n| LET   | scale & shift     | linear layer & attention | gradient descent                                     |\n\n", "category": "Method_Comparison"}, {"question": "How does the performance of OmniQuant compare to Outlier Suppression+ (OS+) in the context of W4A4 quantization and zero-shot tasks?", "answer": "OmniQuant outperforms OS+ in W4A4 quantization across six zero-shot tasks, with accuracy improvements ranging from 4.01% to 6.68%. For instance, LLaMa-1-7B achieves an average accuracy of 52.65% with OmniQuant, compared to 48.43% with OS+.\n\n\n| Model         | Method     | W4A4 Quantization Average Acc. |\n|---------------|------------|-------------------------------|\n| LLaMa-1-7B    | OS+        | 48.43                         |\n| LLaMa-1-7B    | OmniQuant  | 52.65                         |\n| LLaMa-1-13B   | OS+        | 49.86                         |\n| LLaMa-1-13B   | OmniQuant  | 54.37                         |\n| LLaMa-1-30B   | OS+        | 52.62                         |\n| LLaMa-1-30B   | OmniQuant  | 56.63                         |\n| LLaMa-1-65B   | OS+        | 52.52                         |\n| LLaMa-1-65B   | OmniQuant  | 59.23                         |\n", "category": "Method_Comparison"}, {"question": "How does the proposed method behave without careful initialization, particularly in asymmetric cases such as W4A8, W4A6, and W8A4, where SmoothQuant can perform poorly?", "answer": "To validate the impact of initialization, experiments were conducted on LLaMa-1-7B with W4A4 quantization, initializing scaling as 1 and shifting as 0. The results show that careful initialization of scaling and shifting improves final performance, with scaling initialization being more critical due to its role in alleviating outliers.\n\n| LLaMa-1-7B W4A4         | Average PPL | Average Acc. |\n|---------------------|--------------|-------------|\n| careful initialization | 12.87       | 52.65        |\n|  initialize scaling as 1 | 13.64       | 51.37        |\n|  initialize shifting as 0 | 12.95       | 52.22        |\n\n Additionally, OmniQuant demonstrates robust performance in asymmetric cases, as evidenced in Table A6 and Table A17 of the paper.\n\n|  LLaMa-1-7B    | Average PPL | Average Acc. |\n|------------|-------------|--------------|\n|  FP          | 6.38         | 64.09         |\n| W4A4        | 12.87        | 52.65         |\n|W4A8        | 6.60         | 62.40         |\n| W4A6        | 6.85         | 61.64         |\n| W8A4        | 11.52        | 53.46         |\n", "category": "Experimental_Exposition"}, {"question": "How does the paper address the challenge of combining LWC and LET, and what evidence supports the effectiveness of simultaneous training over iterative training?", "answer": "The paper addresses the challenge by training LWC and LET simultaneously. The authors have also explored an iterative training approach by iterations or epochs. The results clearly indicate that training LWC and LET simultaneously yields the best performance. LWC leverages the extremums after LET as the reference, allowing it to capture the dynamic changes in the weight range introduced by LET. This synergy between LET and LWC creates a progressive process, where both techniques reinforce each other rather than interfere. Additional experiments show that simultaneous training with 20 epochs achieves comparable performance to iterative training with 40 epochs, demonstrating the effectiveness and efficiency of training LWC and LET simultaneously.\n\n| LLaMa-1-7B/ W4A4/ Iterative LWC and LET          | Average PPL | Average Acc. |\n|--------------------------------------------------|------------|-------------|\n| simultaneous training (OmniQuant)               | 12.87      | 52.65       |\n| each iteration                                   | 13.56      | 50.91       |\n| each epoch                                       | 13.51      | 52.06       |\n| each epoch + double training epochs              | 12.80      | 52.50       |\n\n", "category": "Method_Mechanics"}, {"question": "What is the novelty of Learnable Equivalent Transformation (LET) compared to existing quantization kernels like LLM.int8, and how does it compare to PACT and LSQ in terms of performance?", "answer": "LET introduces gradient optimization into the equivalent transformation of large language models, which significantly advances low-bits quantization. The paper compares LET with state-of-the-art weight-only (GPTQ, AWQ) and weight-activation (LLM-QAT, Outlier Suppression +) quantization methods. Additionally, PACT and LSQ were transferred to weight quantization in Table A14, showing that LWC outperforms previous clipping-based methods. For activation quantization, MinMax quantization was used instead of LSQ, as LSQ led to poorer performance than MinMax quantization, as shown in Table 4 of the LLM-QAT paper.", "category": "Method_Comparison"}, {"question": "How does the runtime efficiency of OmniQuant compare to GPTQ and AWQ in terms of latency for deploying quantized models?", "answer": "OmniQuant, GPTQ, and AWQ are uniform quantization methods that allow their quantized models to be deployed with similar latency. The paper focuses on practical deployment using existing MLC-LLM, and the provided table shows that OmniQuant’s quantized models can be deployed using MLC-LLM, GPTQ kernel, or AWQ kernel, with respective latencies for LLaMa-7B and LLaMa-13B models.\n\n|                  | LLaMa-7B      | LLaMa-13B     |\n| ---------------- | ------------- | ------------- |\n| FP               | 69.2 token/s  | 52.5 token/s  |\n| 4bit-MLC-LLM     | 134.2 token/s | 91.3 token/s  |\n| 4bit-GPTQ kernel | 91.30 token/s | 63.2 token/s  |\n| 4bit-AWQ kernel  | 155.4 token/s | 94.86 token/s |", "category": "Method_Comparison"}, {"question": "How does the performance of LWC compare to PACT and LSQ in the context of the OmniQuant pipeline?", "answer": "OmniQuant firstly leverages LET to transfer magnitude from activation to weight, and then quantizes the weight after LET with LWC. LWC consistently outperforms PACT and LSQ in the OmniQuant pipeline. This is demonstrated in Table A14, where LWC is replaced with PACT and LSQ, and in Figure A5, which illustrates LWC's adaptability to changing weight ranges introduced by LET, addressing optimization difficulties not addressed by traditional methods.", "category": "Method_Comparison"}, {"question": "How does the proposed method OmniQuant compare with Outlier Suppression+ (OS+)[1] in terms of average accuracy for different LLaMa model sizes?\n\n[1]Xiuying Wei, Yunchen Zhang, Yuhang Li, Xiangguo Zhang, Ruihao Gong, Jinyang Guo, and Xianglong Liu. Outlier suppression+: Accurate quantization of large language models by equivalent and optimal shifting and scaling. arXiv preprint arXiv:2304.09145, 2023.", "answer": "The comparison between OmniQuant and OS+ in terms of average accuracy for different LLaMa model sizes is as follows:\n\n| W4A4        | Quantization | Average Acc. |\n| ----------- | ------------ | ------------ |\n| LLaMa-1-7B  | OS+          | 48.43        |\n| LLaMa-1-7B  | OmniQuant    | 52.65        |\n| LLaMa-1-13B | OS+          | 49.86        |\n| LLaMa-1-13B | OmniQuant    | 54.37        |\n| LLaMa-1-30B | OS+          | 52.62        |\n| LLaMa-1-30B | OmniQuant    | 56.63        |\n| LLaMa-1-65B | OS+          | 52.52        |\n| LLaMa-1-65B | OmniQuant    | 59.23        |", "category": "Method_Comparison"}, {"question": "What are the key differences between OmniQuant and Outlier Suppression+ (OS+) in terms of parameter optimization, block-wise optimization, equivalent transformation scope, and weight clipping?", "answer": "OmniQuant differs from Outlier Suppression+ (OS+) in several ways: (1) OmniQuant uses gradient optimization to obtain both shifting and scaling parameters, while OS+ relies on pre-defined shifting strength and grid-searched scaling parameters. (2) OmniQuant leverages block-wise optimization while OS+ executes grid searching through a mix of single linear layer and multiple linear layers objective . Block-wise optimization considers interaction within a block and can produce better performance as demonstrated in BRECQ. However, it is a challenge for OS+ to expand to block-wise quantization, which would significantly enlarge the solution space and increase the time of grid searching. (3) OmniQuant expands the equivalent transformation to attention operations, while OS+ only applies it to linear layers. (4) OmniQuant introduces learnable weight clipping (LWC), enabling low-bits weight-only quantization, whereas OS+ supports only weight-activation quantization. These distinctions have been added to Section A2 in the revision.", "category": "Method_Comparison"}, {"question": "What are the advantages of the proposed LWC method over PACT and LSQ, particularly when combined with LET in scenarios where weights are frequently changed?", "answer": "The advantages of LWC over PACT and LSQ are demonstrated in Figure A5 in the appendix, which illustrates the dynamic change in weight ranges during training with different clipping-based methods. LWC calculates relative clipping strength using extremums after LET as proxies, enabling it to adapt more naturally to changing weight ranges and address the optimization challenges introduced by LET, unlike traditional absolute clipping thresholds or step size learning methods.", "category": "Method_Comparison"}, {"question": "What is the impact of quantizing the Softmax output to 8-bit on the overall model performance and accuracy?", "answer": "Quantizing the Softmax output to 8-bit and 6-bit results in acceptable performance degeneration, as block-wise calibration can compensate for the loss. However, 4-bit Softmax quantization leads to significant performance loss, requiring further exploration and techniques like log2 quantization in RepQViT [1]. Detailed results are provided in Table A7 of the paper.\n\n| LLaMa-1-7B           | Average PPL | Average Acc. |\n| -------------------- | ----------- | ------------ |\n| FP                   | 6.38        | 64.09        |\n| W4A4 + Softmax 16bit | 12.87       | 52.65        |\n| W4A4 + Softmax 8bit  | 12.91       | 51.93        |\n| W4A4 + Softmax 6bit  | 13.20       | 51.70        |\n| W4A4 + Softmax 4bit  | 18.80       | 48.52        |\n\n[1] Repq-vit: Scale reparameterization for post-training quantization of vision transformers", "category": "Experimental_Exposition"}, {"question": "How does the performance of training LWC and LET simultaneously compare to an iterative training approach, and what evidence supports this comparison?", "answer": "The results, as presented in the following table, clearly indicate that training LWC and LET simultaneously yields the best performance. LWC leverages the extremums after LET as proxies, allowing it to capture the dynamic changes in the weight range introduced by LET. This synergy between LET and LWC creates a progressive process, where both techniques reinforce each other rather than interfere. To further support this statement, an additional experiment (last row in the table), training LWC and LET iteratively with double the epochs, shows that simultaneous training with 20 epochs achieves comparable performance to iterative training with 40 epochs. This demonstrates the effectiveness and efficiency of training LWC and LET simultaneously.\n\n| LLaMa-1-7B/ W4A4/ Iterative LWC and LET | Average PPL | Average Acc. |\n| --------------------------------------- | ----------- | ------------ |\n| simultaneous training (OmniQuant)       | 12.87       | 52.65        |\n| each iteration                          | 13.56       | 50.91        |\n| each epoch                              | 13.51       | 52.06        |\n| each epoch + double training epochs     | 12.80       | 52.50        |", "category": "Method_Comparison"}]}
{"id": "XUZ2S0JVJP", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance", "authors": ["Lingjiao Chen", "Matei Zaharia", "James Zou"], "abstract": "The rapid adoption of large language models (LLMs) has led to an growing number of companies offering generative LLMs as callable services at varying costs. We find that popular generative LLM APIs, such as GPT-4, ChatGPT, and J1-Jumbo, exhibit heterogeneous pricing structures, with fees that can differ by two orders of magnitude, and heterogeneous performance across tasks and input queries. This makes it challenging for users to decide which generative LLM APIs to utilize for their applications and budget. Motivated by these findings, we propose FrugalGPT, an algorithmic framework that adaptively selects which generative LLMs to use for different queries to reduce cost and improve accuracy. Our experiments demonstrate that, for a range of natural language tasks including news classification, reading comprehension, and scientific question answering, FrugalGPT can match the performance of the best individual generative LLM (e.g., GPT-4) with up to a 98% cost reduction or improve the accuracy over GPT-4 by 4% at the same cost. The ideas and findings presented in this paper lay a foundation for using LLMs sustainably and efficiently.", "keywords": "LLM; GPT-4; joint optimization of performance-cost; MLaaS", "qa_pairs": [{"question": "How does FrugalGPT perform when there is a distribution shift on the input text, and what experiments were conducted to evaluate this?", "answer": "New experiments were conducted to evaluate FrugalGPT's performance under distribution shift by rephrasing input text using ChatGPT without changing its meaning. On the HEADLINES dataset, FrugalGPT demonstrated robustness to such shifts, achieving the same accuracy as GPT-4 (84%) with only 33% of the cost.", "category": "Experimental_Exposition"}, {"question": "Why were evaluation datasets limited to tasks with short outputs such as classification or question answering, and how does this choice impact the objectivity of the evaluation?", "answer": "Tasks with short outputs have ground-truth labels that enable objective evaluations. Many LLM benchmarks and evaluations rely on short outputs or multiple choice questions. Evaluating long generations often involves LLMs as judges, introducing biases and subjectivity. Focusing on short outputs ensures the evaluation is objective and aligns with human judgments.", "category": "Motivation_Analysis"}, {"question": "What is the practical impact of search space pruning compared to ordering APIs in increasing order of cost?", "answer": "Pruning identifies the most effective set of LLMs to query and provides an additional 1-2% performance improvement over using a fixed ordering of APIs in increasing cost, as demonstrated in the ablation study comparing FrugalGPT with GPT-J, GPT-Neo, and GPT-4.\n\n| Cost (fraction of GPT-4) | FrugalGPT | FrugalGPT with fixed API order|\n| :----------------: | :------: | :------: | \n| 10%  | 86.8 |85.1 |\n| 20%  | 86.9 |85.6 |\n| 100%  | 87.2 | 85.8|", "category": "Method_Comparison"}, {"question": "What is the impact of changing the threshold T on the results, and how sensitive is the method to T?", "answer": "The threshold T affects performance by controlling when a query is routed to an API. However, practitioners do not need to manually tune T, as it is learned automatically during FrugalGPT’s training process.", "category": "Experimental_Exposition"}, {"question": "What are the specific benefits of FrugalGPT compared to fine-tuning methods?", "answer": "FrugalGPT requires significantly less data than fine-tuning; for instance, fine-tuning a ROBERTa model on COQA needs 127k data points, whereas FrugalGPT needs only a few hundred samples. Additionally, FrugalGPT can be used in conjunction with fine-tuned LLMs to further reduce costs.", "category": "Method_Comparison"}, {"question": "Why were QA and text classification tasks chosen for the experiments, and how do the selected datasets compare in terms of difficulty?", "answer": "These tasks were chosen because they are relatively easy to evaluate and commonly used in LLM research. However, several of the datasets, such as COQA and SCIQ, are quite challenging, with even GPT-4 achieving only 40%-70% accuracy on them.", "category": "Experimental_Analysis"}, {"question": "How does the training set size for FrugalGPT affect its accuracy, and why is the cost of tuning FrugalGPT on a few thousand examples justified in large-scale applications?", "answer": "FrugalGPT is designed for applications involving a large number of LLM queries. Consider, for example, Yabble, which analyzes customer posts for sentiment, or Jasper , which offers chatbots for varying domains. Here, as in many other industry applications of LLM, many tens of thousands of new queries per day are expected. In these cases, the cost of tuning FrugalGPT on a few thousand examples is well worth the reduction in cost when applying it to millions over time. An additional study on the effects of FrugalGPT training set size in the HEADLINES dataset varied the data used to train FrugalGPT from 500 to 5000. The study also measured the accuracy of the FrugalGPT LLM cascade on the test set with 5000 samples. As shown in the following table, FrugalGPT’s performance was consistently better than GPT-4 (85.5%). \n\n| Training set size            | Accuracy |\n| :----------------: | :------: | \n| 500        |   85.9   |\n| 1000           |  86.8  |\n| 2000   | 86.8 |\n| 3000   | 86.8 |\n| 4000   | 87.2 |\n| 5000   | 87.0 |", "category": "Experimental_Exposition"}, {"question": "Is FrugalGPT a task-specific or task-agnostic method, and are the parameters (permutation, judge model, threshold) trained separately for each task?", "answer": "FrugalGPT is a task-agnostic method, but it requires additional training for different tasks to achieve better performance. This is demonstrated by its cost savings in various tasks such as price prediction, reading comprehension, and news classification.", "category": "Method_Disambiguation"}, {"question": "What is the cost of estimating MPI, and does it require running all LLMs over all training examples?", "answer": "MPI is estimated based on each API's performance on the training partitions. For applications requiring numerous and continuous queries to LLMs, the cost of estimating MPI is typically less than 5% compared to the cost of processing queries during deployment.", "category": "Method_Disambiguation"}, {"question": "Can the DistilBERT judge be used directly for inference in classification tasks, and how does its performance compare to FrugalGPT?", "answer": "Yes, the DistilBERT judge can be used directly for inference by processing a query with each possible label and selecting the label with the highest score. However, its performance is significantly worse than FrugalGPT. For instance, on the HEADLINES dataset, the judge achieves an accuracy of 60%, compared to FrugalGPT's 87%. This discrepancy arises because judging is less complex than prediction; the judge may assign low scores to all labels, leading to poor predictions, whereas FrugalGPT can defer to a more capable LLM for higher-quality answers.", "category": "Method_Comparison"}, {"question": "What is the reason for the judge model using a two-dimensional output and selecting the maximum value from the final layer (normalized by softmax) within a regression problem?", "answer": "The judge model uses a 2-dimensional output and takes the maximum value of the last layer (normalized by softmax) because there are no numerical labels for quality, making the standard regression training paradigm unsuitable. Instead, the paper trains a model to predict correctness, which is binary and can be curated from the ground-truth labels. The model's confidence in predicting correctness is then used as the quality estimation. ", "category": "Method_Disambiguation"}, {"question": "Does sampling outputs from the same LLM multiple times and ensembling their answers lead to significant performance improvements, and how does this compare to the performance gains achieved by FrugalGPT?", "answer": "Sampling outputs from ChatGPT three times with different temperatures (0, 0.1, 0.2) and taking the mode of the generations as the final output results in marginal performance gains on the HEADLINES dataset. However, the performance gain is relatively smaller compared to FrugalGPT, which achieves an accuracy of 87% at the same cost. ", "category": "Method_Comparison"}, {"question": "Why does the leftmost dot of FrugalGPT in Figure 1(d) show better performance than the cheap models, despite being at a similar budget point where one would expect a single API call?", "answer": "The leftmost dot of FrugalGPT used slightly more budget than the cheapest model by routing 2% of queries to ChatGPT, resulting in an accuracy gain of 0.7% (from 78% to 78.7%).", "category": "Experimental_Analysis"}]}
{"id": "iSAgvYhZzg", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "You Only Look at Screens: Multimodal Chain-of-Action Agents", "authors": ["Zhuosheng Zhang", "Aston Zhang"], "abstract": "Autonomous user interface (UI) agents aim to facilitate task automation by interacting with the user interface without manual intervention. Recent studies have investigated eliciting the capabilities of large language models (LLMs) for effective engagement in diverse environments. To align with the input-output requirement of LLMs, existing approaches are developed under a sandbox setting where they rely on external tools and application-specific APIs to parse the environment into textual elements and interpret the predicted actions. Consequently, those approaches often grapple with inference inefficiency and error propagation risks. To mitigate the challenges, we introduce Auto-UI, a multimodal solution that directly interacts with the interface, bypassing the need for environment parsing or reliance on application-dependent APIs. Moreover, we propose a chain-of-action technique---leveraging a series of intermediate previous action histories and future action plans---to help the agent decide what action to execute. We evaluate our approach on a new device-control benchmark AITW with 30$K$ unique instructions, spanning multi-step tasks such as application operation, web searching, and web shopping. Experimental results show that Auto-UI achieves state-of-the-art performance with an action type prediction accuracy of 90\\% and an overall action success rate of 74\\%. Code is publicly available at Anonymous.", "keywords": "Autonomous Agent;Large Language Models;Natural Language Interface;Vision-language Model;Chain-of-Thought.", "qa_pairs": [{"question": "Why is it necessary to consider a multimodal version or a better spatial representation of HTML syntax in the proposed method, and how does the performance of the proposed approach compare to a strong multimodal ICL baseline (GPT-4V)?", "answer": "The necessity arises because parsing the visual environment into textual elements can lead to error propagation, information loss, and lengthy inputs, causing inference inefficiency. The paper's proposed approach, Auto-UI, outperforms the strong multimodal ICL baseline (GPT-4V) with significantly higher scores across various metrics: \n\n| Model                     | Overall | General  |  Install  | GoogleApps  | Single  |\n| ------------------------- | ------- | ------- | ------- | ------- | ------- |\n| GPT-4V  | 52.96   | 43.01   | 46.14   | 49.18      | 78.29  |\n| Auto-UI | 74.27   | 68.24   | 76.89   | 71.37      | 84.58  |", "category": "Motivation_Analysis"}, {"question": "How does the performance of the baseline version of a large language model fine-tuned for multimodal input compare to other models??", "answer": "Fine-tuned LLMs with multimodal inputs (LLaVA)[1] can be seen as a variant of the paper's framework by changing the backbone modules. LLaVA achieves slightly better performance gains compared to other baselines, and the proposed chain-of-action technique further enhances its performance.\n\n| Model                     | General  |\n| ------------------------- | ------- |\n| T5 backbone            | 68.24 |\n| LLaVA  backbone      | 68.70   |\n| w/o chain of actions | 58.40   |\n\n[1] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS 2023.", "category": "Method_Comparison"}, {"question": "It looks PaLM and ChatGPT are pretty low on this dataset, while they only take text input, and BC models and Auto-UI models take image screen as input, and get very high results. What are the primary sources of performance gain in the proposed model, and how do the absence of multimodal perception and chain-of-action technique affect its performance?", "answer": "The performance gain in the Auto-UI model primarily comes from two aspects: multimodal perception and the chain-of-action technique. Ablation experiments show that the paper's model out of first principles thinking can serve as a strong autonomous agent andmultimodal perception is critical, as removing the image encoder (w/o multimodal perception) results in a significant drop in performance (62.70 vs. 74.27). The chain-of-action technique further enhances the model's ability by leveraging previously executed actions and future action plans, improving memory and planning capabilities. Thus, while the image encoder is a significant contributor, the chain-of-action technique also plays a crucial role in the overall performance.\n\n| Model                     | Overall |\n| ------------------------- | ------- |\n| Auto-UI                   | 74.27   |\n| w/o multimodal perception | 62.70   |\n| w/o chain of actions      | 68.53   |\n\nIt is reasonable that the baseline results are low due to the lack of vision encoders. As a result, they may suffer from error propagation or information loss (as discussed in the introduction section). To have a deeper understanding of the performance, the authors have conducted additional analysis on the performance regarding the action type and action execution (i.e., clicking accurate positions and scrolling in the correct direction).\n\n| Model                     | Overall | Action Type  | Click        | Scroll       |\n| ------------------------- | ------- | ------------ | ------------ | ------------ |\n| ChatGPT                   | 5.93    | 41.72        | 8.50         | 4.00         |\n| Auto-UI                   | 68.24   | 87.03        | 58.34        | 82.74        |\n\nThe ICL method (ChatGPT) is quite accurate at predicting the action type (41.72%) but fails at lower-level executions, e.g., clicking positions (8.5%) and scrolling directions (4.0%).", "category": "Experimental_Analysis"}, {"question": "What is the detailed explanation of the two paradigms (a) and (b) presented in Figure 1?", "answer": "The two paradigms are explained as follows:\n\n(a) Sandbox paradigm: This paradigm relies on intermediate transformations between environments and agents. On the input side, it uses external tools like optical character recognition (OCR) and icon detectors to parse the environment into textual elements. On the output side, it requires accessing internal APIs to interact with the environment, such as using JavaScript element selection on a webpage or a Python interpreter to execute actions.\n\n(b) First principles thinking paradigm: This paradigm allows direct interactions on the screen without needing intermediate environment parsing or application-dependent APIs.", "category": "Concept_Understanding"}, {"question": "What are the Sandbox Paradigm and the First Principles Thinking Paradigm, and how do they differ in terms of their approaches to environment-agent interactions?", "answer": "The Sandbox Paradigm relies on intermediate transformations between environments and agents. On the input side, it uses external tools like optical character recognition (OCR) and icon detectors to parse the environment into textual elements. On the output side, it requires accessing internal APIs to interact with the environment, such as using JavaScript element selection on a webpage or a Python interpreter to execute actions. In contrast, the First Principles Thinking Paradigm allows direct interactions on the screen without needing intermediate environment parsing or application-dependent APIs.", "category": "Concept_Understanding"}, {"question": "What do the terms 'touch_point' and 'lift_point' refer to in the context of the study?", "answer": "Touch_point and lift_point refer to the axis of touching and lifting, respectively. These notations are widely accepted in related studies and are considered commonsense. Examples are provided in Section 3.3 during the discussion on coordinate normalization.", "category": "Concept_Understanding"}, {"question": "What are the training details of the proposed method, and how does it differ from traditional behavioral cloning (BC) approaches?", "answer": "The training details are provided in Section 4.4. The proposed method is a form of behavioral cloning (BC), but it introduces two key novelties: (i) a multimodal agent for autonomous UI control that directly interacts with screens, bypassing the need for environment parsing and application-specific API access, and (ii) a chain-of-action technique that uses previously executed actions and future action plans to guide the agent's decisions. These innovations represent a significant departure from prior studies. Additionally, Auto-UI achieves state-of-the-art performance with 90% action type prediction accuracy and 74% action success rate, with inference times of less than one second.", "category": "Method_Comparison"}, {"question": "How does the ICL method (ChatGPT) perform at the action plan level and in lower-level executions such as clicking positions and scrolling directions?", "answer": "The ICL method (ChatGPT) is quite accurate at predicting the action type, with a success rate of 41.72%, but performs poorly in lower-level executions, achieving only 8.5% accuracy in clicking positions and 4.0% in scrolling directions.\n\n| Model                     | Overall | Action Type  | Click        | Scroll       |\n| ------------------------- | ------- | ------- | ------- | ------- |\n| ChatGPT | 5.93    | 41.72       | 8.50  | 4.00   |\n| Auto-UI | 68.24   | 87.03       | 58.34 | 82.74  |", "category": "Experimental_Exposition"}, {"question": "What are the major errors identified in the error analysis conducted in Section 5.1, and how do these errors manifest in the model's predictions?", "answer": "The major errors identified in the error analysis lie within the click region and scroll direction predictions. The model tends to click the wrong place or scroll in the wrong direction, despite predicting the right action most of the time. This suggests a need for improving the model’s ability to understand screen layouts, potentially through the use of more advanced vision features.", "category": "Experimental_Analysis"}, {"question": "Is the precision of the language decoder's token predictions to four decimal places important, and could the same results be achieved by splitting image screens into patches and using their centers as coordinate representations?", "answer": "No, the actual precision of the language decoder's token predictions is not important. For clicking actions, as defined in [1], correctness is determined by whether the touch and lift points fall within 14% of the screen’s diagonal length from any of the gold gestures or within the same bounding box. For scrolling actions, a scroll action is considered correct if it has the same scroll axis as the gold gesture. The paper did not use patches but instead used the global (pooled) representation of vision features as described in Section 3.2 (Encoding).\n\n[1]Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Android in the wild: A large-scale dataset for Android device control. arXiv preprint arXiv:2307.10088, 2023.\n\n", "category": "Experimental_Analysis"}, {"question": "What are the main types of errors observed by the proposed framework, and does the framework provide insights into how to assign these errors to specific modules for improvement?", "answer": "The major errors observed by the framework lie within the click region and scroll direction predictions. While the model predicts the right action most of the time, it often clicks the wrong place or scrolls in the wrong direction. This suggests a need to enhance the model's understanding of screen layouts, potentially by incorporating more advanced vision features.", "category": "Experimental_Analysis"}, {"question": "How does the chain-of-action (CoA) technique enhance the decision-making process and contribute to the overall effectiveness of the Auto-UI approach?", "answer": "The chain-of-action (CoA) technique enhances the decision-making process by organizing a chain of previous action histories on the input side and a chain of future action plans on the output side. This bidirectional mechanism encourages the agent to leverage action history and make future plans before predicting the current action.The working mechanism can be seen as bidirectional. CoA improves overall effectiveness by increasing accuracy by +5.74%, with significant improvements in clicking and scrolling accuracy (9.96% and 6.91%, respectively). This is because CoA provides action history and future plans as context, enabling more accurate executions, especially in tasks like clicking and scrolling, where the model already performs well in predicting action types (over 81.69%).\n\n| Model                     | Overall | Action Type  | Click        | Scroll       | Text |\n| ------------------------- | ------- | ------- | ------- | ------- | ------- |\n| w/o Chain of Actions | 58.99 | 81.69 | 48.38 | 75.83 | 93.57 |\n| w/ Chain of Actions | 68.24  (+9.25) | 87.03  (+5.34) | 58.34 (+9.96) | 82.74  (+6.91)  | 93.99 (+0.42) |", "category": "Method_Mechanics"}, {"question": "What is the rationale behind selecting the specific baselines used in the evaluation, and how do they ensure a thorough comparison with existing approaches?", "answer": "The specialized UI agent was selected as it represents the previous state-of-the-art approach in existing studies. In-context learning LLMs were chosen because they reflect the widely used paradigm of developing LLM agents. Fine-tuned LLMs were included to investigate the potential of leveraging open-source LLMs for the problem. These baselines encompass both the in-context learning and fine-tuning paradigms, along with various backbone models of different sizes, ensuring a thorough evaluation of the work's performance against existing approaches.", "category": "Motivation_Analysis"}, {"question": "What is the performance of GPT-4V compared to other models such as PaLM 2-CoT, ChatGPT, and Auto-UI?", "answer": "GPT-4V outperforms PaLM 2-CoT and ChatGPT but is outperformed by Auto-UI, which has an Overall score of 74.27.\n\n| Model                     | Overall | General  |  Install  | GoogleApps  | Single  |\n| ------------------------- | ------- | ------- | ------- | ------- | ------- |\n| PaLM 2-CoT|  39.6 | -  |  - |  - |   -  | \n| ChatGPT  | 7.72  | 5.93  | 4.38  | 10.47  |  9.39  |  8.42  |\n| GPT-4V  | 52.96   | 43.01   | 46.14   | 49.18      | 78.29  |\n| Auto-UI | 74.27   | 68.24   | 76.89   | 71.37      | 84.58  |\n", "category": "Method_Comparison"}, {"question": "Were image screens split into patches and their centers used as coordinate representations for action prediction?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the ICL baseline implemented based on previous research [1]?\n\n[1] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS 2023. ", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the precision of the language decoder's token prediction to four decimal places crucial for the accuracy of clicking and scrolling actions?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "KOTsHW6mBI", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Advancing Beyond Identification: Multi-bit Watermark for Large Language Models", "authors": ["KiYoon Yoo", "Wonhyuk Ahn", "Nojun Kwak"], "abstract": "We propose a method to tackle misuses of large language models beyond the identification of machine-generated text. While existing methods focus on detection, some malicious misuses demand tracing the adversary user for counteracting them. To address this, we propose Multi-bit Watermark via Position Allocation, embedding traceable multi-bit information during language model generation. Leveraging the benefits of zero-bit watermarking, our method enables robust extraction of the watermark without any model access, embedding and extraction of long messages ($\\geq$ 32-bit) without finetuning, and maintaining text quality, while allowing zero-bit detection all at the same time. Moreover, our watermark is relatively robust under strong attacks like interleaving human texts and paraphrasing.", "keywords": "watermark;large language model;safety", "qa_pairs": [{"question": "How is the accuracy of list decoding computed, and is a match between a message in the list and the encoded message counted as a success?", "answer": "The accuracy of list decoding is computed by taking the 'closest message out of the candidates,' assuming the best-case scenario. This approach demonstrates the effectiveness of list decoding in identifying possible candidates but does not claim it is achievable without further means. In addition, the paper also notes in the end of Section 3.3 that 'list decoding technique is not unique to ours and can be applied to other methods as long as the decoding stage is computationally feasible.' Nonetheless, the ability to encode long messages without any added latency is unique to the paper's algorithm due to Position Allocation (compared to the concurrent works). In Figure 5c, the validity of the confidence measure used in list decoding is demonstrated by showing that it is inversely proportional to the error rate. In practice, list decoding is especially useful because provenance tracing via watermarking is far from finding an exact solution, but narrowing down the possible leakage points for a more detailed inspection that may be costly. For instance, when encoding the timestamp, list decoding outputs the range of time, for which the practitioners manually inspect and find suspicious activity. When encoding the exact user ID, list decoding will output a list of IDs with varying characters in certain positions. Then the practitioners will consider the IDs that actually exist in their systems and further inspect other details such as recent user activities.", "category": "Method_Mechanics"}, {"question": "What is the performance of the multi-bit encoding scheme on instruction-tuned models like LLaMA-2-7b and Mistral-7b, and does instruction-tuning limit the capacity of watermarking?", "answer": "The results for instruction-tuned models for LLaMA-2-7b (`meta-llama/Llama-2-7b-chat-hf`) and Mistral-7b (`mistralai/Mistral-7B-Instruct-v0.1`) are shown below. The results are on 8-bit and approximately 250 tokens for 100 samples. For the tuned models, the instruction was 'Complete the following news article:'. While the finetuned version on LLaMA shows noticeable falloff, this is not the case for Mistral (Jiang et al., 2023)[1]. \n\n| Model        | Bit Acc.   |\n| ------------ | ---------- |\n| LlaMA-2-7b   | .986 (.06)  |\n| + chat       | .922 (.13)  |\n| Mistral-7b   | .987 (.06)  |\n| +chat        | .977 (.08)  |\n\n The findings suggest that instruction-tuning itself does not limit the capacity of watermarking, but the dataset and amount of tuning may be the key factors. \n\n[1] Jiang, Albert Q., et al. \"Mistral 7B.\" arXiv preprint arXiv:2310.06825 (2023).", "category": "Experimental_Exposition"}, {"question": "How does the quality and bit accuracy of the multi-bit watermarking scheme presented here compare with the Kirchenbauer et al.[1] 0-bit scheme as delta increases?\n\n[1]John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. arXiv preprint arXiv:2301.10226, 2023a.", "answer": "The table and Figure 9 show the quality and bit accuracy tradeoff for increasing $\\delta$ in the 8-bit message embedding. The P-SP metric measures similarity between watermarked and human-written sentences. Although zero-bit watermark results are lacking, Figure 3 indicates that quality remains equivalent regardless of $\\delta$ magnitude.\n\n| Delta |   0.5   |   1    |   2    |   3   |   5   |  10   |\n|-------|---------|--------|--------|-------|-------|-------|\n| Bit Acc. | 0.618 (.18) | 0.751 (.19) | 0.917 (.13) | 0.979 (.08) | 0.997 (.02) | 0.994 (.04) |\n| P-SP  |   0.384  |   0.385  |   0.376  |  0.356  |  0.325  |  0.244  |\n", "category": "Experimental_Exposition"}, {"question": "What are the key contributions of the paper ?", "answer": "The key contributions are: 1. Proposing and validating Position Allocation to extend zero-bit watermark to multi-bit, with Colorlisting enhancing performance. 2. Demonstrating the feasibility of directly embedding long multi-bit information into language model outputs. This allows a robust watermark that can withstand realistic attacks, such as altering the start and end positions of the machine text (copy-paste attack) or changing entire sentence structures (paraphrasing), offering advantages over previous multi-bit watermark works in this regard. 3. Discussing the limitations of multi-bit watermarking, specifically the tradeoff between watermark capacity and detection rate.", "category": "Method_Disambiguation"}, {"question": "What evidence supports the claim that the proposed multi-bit watermarking approach has sufficient bandwidth in practice, especially for LLMs that have undergone RLHF/SFT?", "answer": "Certain tasks such as code generation have shown to be difficult for watermarking due to the inherently low entropy distribution. Nonetheless, many other open-ended generation tasks certainly have enough capacity for multi-bit watermark. Generation of disinformation falls into this case as numerous different passages can be created given a certain topic or propaganda. The paper's main experiments simulate this scenario by completing a news-like subset of C4. As shown in the experiments, the method can extract an 8-bit message from 125 tokens with approximately 95% accuracy.For reference, the above paragraph contains 123 tokens. The results for instruction-tuned models for LLaMA-2-7b (`meta-llama/Llama-2-7b-chat-hf`) and Mistral-7b (`mistralai/Mistral-7B-Instruct-v0.1`)  are shown below . The results are on 8-bit and approximately 250 tokens for 100 samples. For the tuned models, the instruction was 'Complete the following news article:'. While the finetuned version on LLaMA shows noticeable falloff, this is not the case for Mistral (Jiang et al., 2023)[1]. \n\n| Model         | Bit Acc.  |\n| ------------- | --------- |\n| LlaMA-2-7b    | .986 (.06)|\n| + chat        | .922 (.13)|\n| Mistral-7b    | .987 (.06)|\n| +chat         | .977 (.08)|\n\n This hints at the possibility that instruction-tuning itself does not limit the capacity of watermarking, but rather the dataset and amount of tuning may be the key. \n\n[1]Jiang, Albert Q., et al. \"Mistral 7B.\" arXiv preprint arXiv:2310.06825 (2023).", "category": "Experimental_Analysis"}, {"question": "What are the results of the ablation experiments comparing list decoding and random bit flips in the most likely decoded message, and how do they demonstrate the effectiveness of each method?", "answer": "The experiments, conducted with an effective message length of 16, which was tested on $b=32$, $r=4$, show the absolute accuracy improvement over the best message (87.0%) for different list sizes. The results are as follows:\n\n| List Size: | 2     | 4     | 8     | 16    |\n|------------|-------|-------|-------|-------|\n| Random flip| 0.001 | 0.002 | 0.002 | 0.003 |\n| Random flip from best | 0.012 | 0.020 | 0.031 | 0.043 |\n| List decoding by confidence | 0.025 | 0.035 | 0.047 | 0.057 |\n\n. These results indicate that applying random flips from the best message provides a good list, and choosing positions by confidence consistently improves over random flips.", "category": "Experimental_Exposition"}, {"question": " What are the limitations of the multi-bit watermarking method presented by Yoo et al. (2023)[1], and how does the proposed MPAC algorithm address these limitations?\n\n[1]KiYoon Yoo, Wonhyuk Ahn, Jiho Jang, and Nojun Kwak. Robust multi-bit natural language watermarking through invariant features. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 2092–2115, 2023.", "answer": "The most robust multi-bit watermarking method among post-processing methods (Abdelnabi & Fritz, 2021[2], Yang et al., 2022[3]) is Yoo et al. (2023)[1]. However, this work only considers corruption at word-levels such as deletion, insertion, and substitution. Methodologically, one clear limitation is the sentence-level method whereby embedding and extraction are done at each sentence level. As an illustration, consider the example below:\n\n```\nSentence 1 -> extracted message : “000”\nSentence 2 -> extracted message : “000”\nSentence 3 -> extracted message : “111”\nSentence 4 -> extracted message : “111”\nFinal extracted message : “000000111111”\n```\n\nIf any human text were to be added into the first part of the writing (instance of copy-paste attack), the message is irrecoverable as it has no way to identify _when_ the machine text starts or ends. The problem persists when a single sentence in the middle is removed as well. In contrast, the proposed algorithm MPAC need not know this and achieves relative robustness in various kinds of attacks. In addition, the paper's method enables zero-bit detection (machine-text detection) and multi-bit watermarking all at the same time.\n\nThe concurrent works (Wang et al., 2023[4]; Fernandez et al., 2023[5]), which are detailed above in the general comments and Related Works section of the paper, also rely on the hashing function similar to the paper. Theoretically, the paper's MPAC has an advantage in embedding long messages without any added latency during generation (Figure 3).\n\n[2]Sahar Abdelnabi and Mario Fritz. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pp. 121– 140. IEEE, 2021.\n[3]Xi Yang, Jie Zhang, Kejiang Chen, Weiming Zhang, Zehua Ma, Feng Wang, and Nenghai Yu. Tracing text provenance via context-aware lexical substitution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp. 11613–11621, 2022.\n[4]Lean Wang, Wenkai Yang, Deli Chen, Hao Zhou, Yankai Lin, Fandong Meng, Jie Zhou, and Xu Sun.nTowards codable text watermarking for large language models. arXiv preprint arXiv:2307.15992,n2023a.\n[5]Pierre Fernandez, Antoine Chaffin, Karim Tit, Vivien Chappelier, and Teddy Furon. Three bricks to consolidate watermarks for large language models. arXiv preprint arXiv:2308.00113, 2023a.", "category": "Method_Comparison"}, {"question": "How do the inherent low entropy of certain text distributions, such as code, impact the embedding capacity, and how does this relate to the generation of disinformation in the paper's experiments?", "answer": "Some of the text distributions inherently have low entropy, preventing them from embedding large, if any, amount of information. Nonetheless, malicious use cases of language model also involves open-ended generation rather than a fixed solution. Generation of disinformation falls into this case as numerous different passages can be created given a certain topic or propaganda. In the paer's experiments, this scenario is simulated using a news-like subset of C4, successfully embedding an 8-bit message into 125 tokens with approximately 95% accuracy. The referenced paragraph contains 123 tokens.", "category": "Experimental_Analysis"}, {"question": "What results were obtained for instruction-tuned models including LLaMA-2-7b, Mistral-7b, and OPT-1.3b, and what does this suggest about the impact of instruction-tuning on watermarking capacity?", "answer": "The results for instruction-tuned models for LLaMA-2-7b (`meta-llama/Llama-2-7b-chat-hf`), Mistral-7b (`mistralai/Mistral-7B-Instruct-v0.1`), and a relatively small model (`OPT-1.3b`) are shown below.\n\nThe results are on 8-bit and $\\approx$ 250 tokens for 100 samples. For the tuned models, the instruction was 'Complete the following news article:'. While the finetuned version on LLaMA shows noticeable falloff, this is _not_ the case for Mistral (Jiang et al., 2023)[1].\n\n| Model | Bit Acc. |\n| ---------| ---------| \n| LlaMA-2-7b | .986 (.06) | \n| + chat | .922 (.13) |  \n| Mistral-7b | .987 (.06)| \n| +chat | .977 (.08) |\n|OPT-1.3b|.982 (.07)|\n\nThese results suggest that instruction-tuning itself does not limit the capacity of watermarking, but rather the dataset and amount of tuning may be the key factors.\n\n[1]Jiang, Albert Q., et al. \"Mistral 7B.\" arXiv preprint arXiv:2310.06825 (2023).", "category": "Experimental_Exposition"}, {"question": "Why does the model's performance decrease when embedding longer messages at fixed bits per token, as shown in Figure 3?", "answer": "The decrease in performance when embedding longer messages is due to the complexity of the medium, quantified as bits per token (BPT). This limitation is inherent to watermarking methods but does not diminish the main contribution of demonstrating the feasibility of embedding a robust multi-bit watermark on language model outputs. The embedding of ≥ 32-bit messages is intended to show the feasibility of long message embedding without latency, which is computationally infeasible in concurrent works(Fernandez et al., 2023a; Wang et al., 2023). Additionally, Section 3.3 and Figure 6b discuss how SNR is a more reliable factor in estimating performance, while Section 5 and Appendix A.8 analyze the tradeoff between watermark capacity and detection rate.\n\n[1]Pierre Fernandez, Antoine Chaffin, Karim Tit, Vivien Chappelier, and Teddy Furon. Three bricks to consolidate watermarks for large language models. arXiv preprint arXiv:2308.00113, 2023a.\n[2]Lean Wang, Wenkai Yang, Deli Chen, Hao Zhou, Yankai Lin, Fandong Meng, Jie Zhou, and Xu Sun. Towards codable text watermarking for large language models. arXiv preprint arXiv:2307.15992, 2023a.", "category": "Experimental_Analysis"}, {"question": "How does the paper address the potential limitations of relying solely on quantitative metrics like AUC and bit-accuracy, and how does it illustrate the real-world applicability and user experience of the proposed multi-bit watermarking method?", "answer": "The paper provides an example scenario of tracing back to the provenance after identifying machine-text as done in Fernandez et al. (2023)[1], demonstrating that with an 8-bit message (256 users) embedded onto 250 tokens and a false positive rate of 1e-3, 98.3% of machine text is correctly identified, with a 92.6% exact match for the correct user. Additionally, the MPAC method maintains the exact text quality of zero-bit watermarking (KGW).The effect of watermarking on the text quality compared to no watermarking shows promising results in human evaluations when sufficiently large models are used for open-ended generation by Kirchenbauer et al. 2023b[2](Appendix A.2 and A.9).\n\n[1]Pierre Fernandez, Antoine Chaffin, Karim Tit, Vivien Chappelier, and Teddy Furon. Three bricks to consolidate watermarks for large language models. arXiv preprint arXiv:2308.00113, 2023a.\nPierre Fernandez, Guillaume Couairon, Herve J ´ egou, Matthijs Douze, and Teddy Furon. The stable ´ signature: Rooting watermarks in latent diffusion models. arXiv preprint arXiv:2303.15435, 2023b.\n[2]John Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu, Khalid Saifullah, Kezhi Kong, Kasun Fernando, Aniruddha Saha, Micah Goldblum, and Tom Goldstein. On the reliability of watermarks for large language models. arXiv preprint arXiv:2306.04634, 2023b.", "category": "Experimental_Analysis"}, {"question": "How does the proposed system address potential real-world challenges such as sophisticated adversarial attacks and heavy noise, beyond the evaluated human-induced copy-paste attacks?", "answer": "The paper's work covers potential evasion attempts by experimenting on paraphrasing with Chat-GPT and other language models. While not exhaustive, these tests are relevant and automatable. In addition, list decoding is one solution for mitigating real-world challenges. This is especially useful in practice because provenance tracing via watermarking is far from finding an exact solution, but narrowing down the possible leakage points for a more detailed inspection that may be costly. This is because the decoded message is not error-free in the real-world and thus, one cannot finalize the decision by watermarking alone. For instance, when encoding the timestamp, list decoding outputs the range of time, for which the practitioners can manually inspect and find suspicious activity. When encoding the exact user ID, list decoding will output a list of IDs with varying characters in certain positions. Then the practitioners will consider the IDs that actually exist in their systems and further inspect other details such as recent user activities. ", "category": "Method_Mechanics"}, {"question": "How does the method address the vulnerability of watermark tampering or removal if the pseudo-random generator function is reverse-engineered or predicted?", "answer": "Watermark spoofing has been mentioned as a possible attack to 'fake' a watermark to harm the reputation of certain models (Sadasivan et al., 2023). More realistically, removing the watermark may be possible if one knows the tokens in the colorlist. An important parameter that trades off privacy (difficulty of reverse engineering) and robustness (to corruptions) is the context width $h$. Assuming the use of LLaMA-2 tokenizer, the brute-force attack takes $|\\mathcal{V}|^{h+1}= 32000^2 \\approx 1e^{9}$, which is a considerable amount of effort even for the easiest case ($h$=1). All this presumes that the API call to extracting the watermark is accessible and that it shows the total number of colorlisted tokens. Simply displaying the extracted multi-bit message and the p-value under certain thresholds, e.g. {1e-6, 1e-5, 1e-4, 1e-3, 1e-2}, will make reverse engineering extremely difficult.\n\n[1]Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can AI-generated text be reliably detected?, 2023. arXiv preprint arxiv:2303. 11156 (2023).", "category": "Method_Mechanics"}, {"question": "What are the known methods or strategies that might weaken or remove the watermark effectively, given its robustness against attacks like interleaving human texts and paraphrasing?", "answer": "A strong corruption method is to use a strong paraphraser. In the experiments, Chat-GPT was found to be a strong candidate in breaking the watermarks.An example of paraphrased text is shown below:\n\nWatermarked: Furthermore, the county has committed to providing Amazon with $20 million in cash grants to help increase the amount of affordable housing in the area and improve bus services. To fund this, the county will implement a new tax on office space valuations. Additionally, Amazon will receive $7 million in electricity rebates over 15 years, with an average annual discount of $60,000. These rebates are set to begin in 2020.\n\nParaphrased: The county also has pledged to provide Amazon with $20 million in cash grants, to increase the amount of affordable housing in the area and to upgrade bus service. The county would receive the money through a new tax on office space valuations. In addition, the county plans to provide Amazon with a 15-year payment schedule for about $7 million in electricity rebates. This would give the company an average annual discount of $60,000, according to the agreement. The electricity rebates would start in 2020.", "category": "Experimental_Analysis"}, {"question": "What are the key factors that enable the MPAC method to integrate easily with existing large language models and platforms, and how can potential challenges such as discriminating between message bits be addressed?", "answer": "The MPAC method is a simple extension of KGW’s zero-bit watermarking, involving a straightforward step of sampling the position before selecting the token subset, which results in zero overhead. It is also theoretically extendable to other zero-bit watermarking methods, as demonstrated by the works of Aaronson and Kirchner (2023)[1] and Kuditipudi et al. (2023)[2]. The adaptation of the Position Allocation scheme is detailed in Appendix A.10. The primary challenge of discriminating between message bits (e.g., $\textbf{m}[p]=0$ and $\textbf{m}[p]=1$) can be addressed by inverting $\\mathbf{r}$ using Eq. 8, since exponential minimum sampling favors tokens with large $\\mathbf{r}_v$. \n\n[1]Scott Aaronson and Hendrik Kirchner. Watermarking gpt outputs. https://www.\nscottaaronson.com/talks/watermark.ppt, 2023. Accessed: 2023-09-14.\n[2]Kuditipudi, Rohith, et al. \"Robust distortion-free watermarks for language models.\" arXiv preprint arXiv:2307.15593 (2023).", "category": "Method_Mechanics"}, {"question": "What are the practical benefits of studying watermarking methods for close-sourced language models, given the challenges with open-source LLMs and the difficulty in achieving a low FPR?", "answer": "Studying watermarking methods for close-sourced language models is important due to their prevalence, as evidenced by large user bases like ChatGPT and Bard. The largest API provider, ChatGPT, had over 206 million monthly unique visitors in April, while Bard had over 140 million monthly unique visitors in May. Moreover, companies in Korea and China have also launched servicing their close-sourced models as API or apps (Baek, 2023[1]; David, 2023[2]). As of now, the most capable models are arguably close-sourced and if this trend were to continue, it is likely that the number of users will not decline. Watermarking is one of the solutions that can mitigate malicious uses of closed-sourced language models. Having a reference point in academia strongly incentivizes the owners of closed-sourced models to apply them. Another point to take into consideration is that from a protection standpoint, enforcing the adversary to take evading measures is itself valuable if it is costly. The best — arguably — paraphrasing model, ChatGPT is behind paywall when called via API.\n\n[1]Byung-yeul Baek, Naver to launch HyperCLOVA X AI service to rival ChatGPT, https://www.koreatimes.co.kr/www/tech/2023/11/129_350568.html, Accessed: 2023-11-10\n\n[2]Emillia David, Baidu launches Ernie chatbot after Chinese government approval, https://www.theverge.com/2023/8/31/23853878/baidu-launch-ernie-ai-chatbot-china, Accessed: 2023-11-10", "category": "Method_Disambiguation"}, {"question": "Why was a fixed-length list $|\\mathbb{L}|$ used for list decoding instead of a confidence threshold approach?", "answer": "The authors chose a fixed-sized list for simplicity because the predictions were uncalibrated, making it difficult to realize a confidence threshold. The results were intended to demonstrate the effectiveness of list decoding in identifying possible candidates, not to claim that it was achievable without further means.In Figure 5c, the validity of the confidence measure used in list decoding is demonstrated by showing that it is inversely proportional to the error rate.", "category": "Motivation_Analysis"}, {"question": "How is list decoding relevant and justified in the context of bit accuracy metric, and how can practitioners utilize the 16 less-confident messages in an LLM detection/attribution setting?", "answer": "List decoding is included because provenance tracing via watermarking is not error-free and does not provide an exact solution. It helps narrow down possible leakage points for detailed inspection, which can be costly. For example, when encoding a timestamp, list decoding outputs a range of time for practitioners to manually inspect for suspicious activity. When encoding a user ID, it outputs a list of IDs with varying characters in certain positions, allowing practitioners to consider existing IDs and inspect further details like recent user activities.", "category": "Method_Mechanics"}, {"question": "What is the true positive rate (TPR) at different false positive rate (FPR) thresholds for the proposed watermarking method, and how does it vary with bit-width?", "answer": "The true positive rate (TPR) at FPR=1e-3 is as follows:\n\n| Bit-width | @FPR=1e-3 |\n|-----------|-----------|\n| 0         | 0.983     |\n| 8         | 0.975     |\n| 16        | 0.950     |\n| 24        | 0.931     |\n| 32        | 0.907     |\n\n As bit-width increases, the detection performance decreases, which is expected because the zero-bit setting is a special case of embedding a single-bit message of $\\mathbf{m}=$ '0' in all samples.", "category": "Experimental_Exposition"}, {"question": "How does the efficiency of the proposed multi-bit watermarking method compare to the zero-bit watermarking method by Kirchenbauer et al. [2023a][1] in detecting machine-generated texts?\n\n[1]John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. arXiv preprint arXiv:2301.10226, 2023a.", "answer": "The TPR at different FPR thresholds is shown below. This uses 500 positive (watermarked) samples and 1000 negative samples at 250 tokens.\n\n| Bit-width | @FPR=$1e^-3$ |\n| :-------- | :-------- |\n| 0         | .983      |\n| 8         | .975      |\n| 16        | .950      |\n| 24        | .931     |\n| 32        | .907      |\n\nAs bit-width increases, the detection performance decreases. This is expected as the zero-bit setting is a special case of embedded a single-bit message of$\\mathbf{m}$= \"0\" in all the samples. To give an apples-to-apples comparison, the authors embed the same 8-bit message ($\\mathbf{m}$=\"00000000\"). This is essentially the same setting with the zero-bit watermarking, but applying only Position Allocation on top of it. To see the true positive rate at lower FPR, the authors sampled 10,000 negative samples and computed the true positive rate.  The results for zero-bit and 8-bit are shown below.\n\n\n| Bit-width               | @FPR=1e-3 | @FPR=1e-4 | @FPR=1e-5 |\n|-------------------------|-----------|-----------|-----------|\n| 0                       | 0.997     | 0.997     | 0.997    |\n| 8 (single message)      | 0.994     | 0.992     | 0.992     |\n| 8                       | 0.974     | 0.965     | 0.965     |\n\n\nAs expected, embedding the same message using 8-bit shows nearly identical results. Compared to this, actually embedding the unique 8-bit message does lead to performance degradation, but still maintains a fairly high TPR even at FPR=1e-5. The reason behind this is that the detection metric of non-watermarked (human) text is over-estimated as bit-width is increased.", "category": "Method_Comparison"}, {"question": "How does the proposed multi-bit watermark methodology differ from the red-green list approach in [1], and what are the implications for robustness?\n\n[1] A Watermark for Large Language Models. John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, Tom Goldstein. ICML (2023)", "answer": "The paper's work does not claim to 'enhance the robustness intrinsic to' Kirchenbauer et al., (2023a). Rather, the paper proposes a new research direction for extending this to multi-bit. Comparing the performance directly with KGW is not an apples-to-apples comparison. In Section 3.2 Message Encoding, the paper shows that the zero-bit detection problem is equivalent to embedding a single-bit message. As watermarking (Vleeschouwer et al., 2002[2]) involves a tradeoff between load capacity (how much information) and robustness (how much information is preserved), the paper tackles an inherently more challenging task. The paper discusses in detail the tradeoff between message length and detection in Section 5 and presents FPR comparison with zero-bit detection in Table 7. Comparing the watermark detection performance against Kirchenbauer et al., (2023a) to quantify the degradation as bit-width increases is expected, as the zero-bit setting is a special case of embedding a single-bit message of $\\mathbf{m}=$'0' in all the samples. To give an apples-to-apples comparison, the authors embed the same 8-bit message ($\\mathbf{m}$='00000000'). This is essentially the same setting as the zero-bit watermarking, but with Position Allocation applied on top of it. To determine the true positive rate at lower FPR, the authors sampled 10,000 negative samples and computed the true positive rate. The results for zero-bit and 8-bit are shown below. \n\n| Bit-width          | @FPR=1e-3 | @FPR=1e-4 | @FPR=1e-5 |\n|--------------------|-----------|-----------|-----------|\n| 0                  | 0.997     | 0.997     | 0.997     |\n| 8 (single message) | 0.994     | 0.992     | 0.992     |\n| 8                  | 0.974     | 0.965     | 0.965     |\n\n \nAs expected, embedding the same message using 8-bit shows nearly identical results. Compared to this, actually embedding the unique 8-bit message does lead to performance degradation, but still maintains a fairly high TPR even at FPR=1e-5. In Appendix A.8, the paper discusses in depth the reason behind this: the detection metric of non-watermarked (human) text is over-estimated as bit-width is increased. \n\n[2]C. De Vleeschouwer, JF Delaigle and B. Macq, \"Invisibility and application functionalities in perceptual watermarking an overview,\" in Proceedings of the IEEE, vol. 90, no. 1, pp. 64-77, Jan. 2002, doi: 10.1109/5.982406.\n", "category": "Method_Comparison"}, {"question": "Why is there no comparative robustness analysis between the proposed method and the method established in [1], and how does the proposed MPAC method achieve robustness against various attacks?\n\n[1] A Watermark for Large Language Models. John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, Tom Goldstein. ICML (2023)", "answer": "While the mentioned work[1] discusses a variety of possible attacks, which is in itself valuable, the only empirical result is Figure 6 which experiments with paraphrasing using T5. The paper follows the experimental setting of Kirchenbauer et al., (2023b)[1], which focuses on the robustness of the zero-bit watermark. The paper shows the results on three realistic and strong attacks: - Copy-pasting human text (Figure 4a) - Paraphrasing using Chat-GPT (Figure 4b) - Paraphrasing using a specialized model (Table 4). The paper contains thorough experiments with multiple sets of attacks, and one of the strengths of the paper is its demonstration of the model's performance under copy-paste attacks. It is believed that the paper would be strengthened by including an additional comparison with other multi-bit methods. The most robust multi-bit watermarking method among post-processing methods, as cited in Abdelnabi & Fritz (2021)[1] and Yang et al. (2022)[2], is Yoo et al. (2023)[3]. However, this work only considers corruption at word-levels such as deletion, insertion, and substitution. Methodologically, one clear limitation is the sentence-level method, whereby embedding and extraction are done at each sentence level. As an illustration, consider the example below: ``` Sentence 1 -> extracted message : '000' Sentence 2 -> extracted message : '000' Sentence 3 -> extracted message : '111' Sentence 4 -> extracted message : '111' Final extracted message : '000000111111' ``` If any human text were to be added into the first part of the writing (instance of copy-paste attack), the message is irrecoverable as it has no way to identify when the machine text starts or ends. The problem persists when a single sentence in the middle is removed as well. In contrast, the proposed algorithm MPAC need not know this and achieves relative robustness in various kinds of attacks. The concurrent works released around July 2023, which are listed above in the general comments and Related Works section of the paper, also rely on the hashing function similar to the paper. It is expected that the concurrent works will have better robustness than the post-processing methods. Theoretically, the paper's MPAC has an advantage in embedding long messages without any added latency during generation.\n\n[1]John Kirchenbauer, Jonas Geiping, Yuxin Wen, Manli Shu, Khalid Saifullah, Kezhi Kong, Kasun Fernando, Aniruddha Saha, Micah Goldblum, and Tom Goldstein. On the reliability of watermarks for large language models. arXiv preprint arXiv:2306.04634, 2023b.\n\n[2]Sahar Abdelnabi and Mario Fritz. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pp. 121– 140. IEEE, 2021.\n\n[3]Xi Yang, Jie Zhang, Kejiang Chen, Weiming Zhang, Zehua Ma, Feng Wang, and Nenghai Yu. Tracing text provenance via context-aware lexical substitution. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp. 11613–11621, 2022.\n\n[4]KiYoon Yoo, Wonhyuk Ahn, Jiho Jang, and Nojun Kwak. Robust multi-bit natural language watermarking through invariant features. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 2092–2115, 2023.\n\n", "category": "Method_Comparison"}, {"question": "Does the correlation between increased bit-width and diminished robustness in Figure 4(a) suggest that the zero-bit watermark proposed by Kirchenbauer et al. (2023a)[1] might be a more robust solution than the proposed multi-bit watermark?\n\n[1]John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. arXiv preprint arXiv:2301.10226, 2023a.\n", "answer": "While increasing bit-width does affect performance, it does not imply that Kirchenbauer et al., (2023a)'s zero-bit watermark is more robust.  Embedding multi-bit information opens up a wide array of applications previously infeasible by zero-bit watermark alone. Moreover, the paper discusses and analyze in depth the limitations of multi-bit watermarking in (1) Section 3.3 how SNR is a more reliable factor in estimating the performance and (2) the tradeoff between the capacity of the watermark and its detection rate (Section 5 and Appendix A.8).", "category": "Method_Disambiguation"}, {"question": "Do the results indicate that certain instruction-tuned models, like Mistral-7b, maintain high accuracy in watermarking despite finetuning?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the experiments show that instruction-tuning significantly limits the capacity for watermarking in all models?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "82A2EfMu3e", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Efficient Discrete Physics-informed Neural Networks for Solving Evolutionary Partial Differential Equations", "authors": ["Ye Li", "Siqi Chen", "Bin Shan"], "abstract": "Physics-informed neural networks (PINNs) have shown promising potential for solving partial differential equations (PDEs) using deep learning. \nHowever, PINNs face training difficulties for evolutionary PDEs, particularly for dynamical systems whose solutions exhibit multi-scale or turbulent behavior over time.\nThe reason is that PINNs may violate the temporal causality property since all the temporal features in the PINNs loss are trained simultaneously. \nThis paper proposes to use implicit time differencing schemes to enforce temporal causality, and use transfer learning to sequentially update the PINNs in space as surrogates for PDE solutions in different time frames.\nThe evolving PINNs are better able to capture the varying complexities of the evolutionary equations, while only requiring minor updates between adjacent time frames.\nOur method is theoretically proven to be convergent if the time step is small and each PINN in different time frames is well-trained.\nIn addition, we provide state-of-the-art (SOTA) numerical results for a variety of benchmarks for which existing PINNs formulations may fail or be inefficient.\nWe demonstrate that the proposed method improves the accuracy of PINNs approximation for evolutionary PDEs and improves efficiency by a factor of 4–40x.\nAll code and data can be found in the supplemental materials.", "keywords": "Neural networks;Partial differential equation;Physics-informed machine learning", "qa_pairs": [{"question": "What is the maximum allowable increase in the time-step parameter ($\\Delta t$) before it negatively impacts the fine-tuning of the last layer, and why?", "answer": "The time-step parameter $\\Delta t$ should remain small to ensure small prediction errors, as indicated by the error bound in Eq.(6). Increasing $\\Delta t$ excessively leads to large prediction errors, rendering training ineffective. The chosen value of $\\Delta t$ aligns with traditional finite difference methods, and Appendix A.4.2 demonstrates that the total parameters and computation for discrete PINNs are comparable to a single continuous PINN.", "category": "Concept_Understanding"}, {"question": "What is the meaning of 'equations being dissipative' in the theoretical result, and does this imply that dissipative/diffusive terms are always required? Additionally, is the proposed method applicable to general hyperbolic equations, including those with external forces like high Reynolds regime flows?", "answer": "The assumption that the equation must be dissipative is only used to prove the error bound Eq.(6). Experimental results demonstrate that the method also performs well for non-dissipative equations, such as the chaotic KS and NS equations.", "category": "Concept_Understanding"}, {"question": "Why does using a single multi-head neural network lead to slower convergence compared to using separate neural networks for each hidden stage?", "answer": "Theoretically, a single multi-head neural network to represent the $q$ hidden stage and $u^{n+1}$ could be beneficial in terms of sharing parameters and reducing the computational effort.However, experimentally requires more epochs and time to train to the same loss compared to using $q+1$ neural networks separately.", "category": "Experimental_Analysis"}, {"question": "What is the rationale behind using the tanh activation function in the MLP network instead of the commonly used relu activation function?", "answer": "The relu activation function is unsuitable for PINN approximation because its second-order derivative is zero and does not contribute to the high-order term in the residual loss of partial differential equations. The tanh activation function is a common choice in the PINN community due to its nonlinear properties.", "category": "Method_Disambiguation"}, {"question": "Was the numerical accuracy of the prediction tested, and does the Crank-Nicolson scheme achieve second-order accuracy in time?", "answer": "Theoretical proof in error bound Eq.(6) confirms second-order accuracy in time for the Crank-Nicolson scheme, represented by the term $\\tau^2$. If the loss is trained to be small at every timestamp and the collation points are large enough, the lead term of the error bound will be $\\tau^2$. Numerical verification was performed by computing the RD equation for $\\Delta t=1/100$ and $1/200$, resulting in errors of $7.01e-05$ and $1.82e-05$ respectively.", "category": "Method_Disambiguation"}, {"question": "Was the paper’s method evaluated on out-of-domain data, and was its applicability for long-term predictions such as weather forecasting explored?", "answer": "The paper's discrete PINNs can predict at trained discrete timestamps in time and arbitrary continuous physical points in space. For long-term predictions, the method can be trained iteratively in time to predict.", "category": "Experimental_Setup"}, {"question": "How does the training time of the proposed method vary with an increasing number of time steps, and what factors influence this variation?", "answer": "When the time steps are increased from 200 to 400, 600, 800, and 1000, the training time increases slightly from 1463s to 1984s, 2339s, 2743s, and 2977s, respectively. However, the hyperparameter threshold value ϵ in Algorithm 1, which varies for different time steps, significantly affects both the training time and accuracy.", "category": "Experimental_Exposition"}, {"question": "What is the required range for $\\Delta t$ to ensure small prediction errors in the proposed method, and does this requirement vary for different Partial Differential Equations (PDEs)?", "answer": "The $\\Delta t$ should be small to maintain a small prediction error, as indicated by the error bound in Eq.(6). A large $\\Delta t$ results in significant prediction errors, rendering training ineffective. The value of $\\Delta t$ is selected comparably to traditional finite difference methods, indicating a consistent approach across different PDEs.", "category": "Concept_Understanding"}, {"question": "What are the implications of assumption 1, and how does the method perform with non-dissipative equations?", "answer": "The dissipative assumption is used solely to prove the error bound Eq.(6). However, the error bound may still hold for non-dissipative equations. Experimental results indicate that the method is effective for non-dissipative equations, including chaotic KS and NS equations.", "category": "Concept_Understanding"}, {"question": "What are the limitations of the proposed method in solving chaotic PDEs, particularly in large-time predictions, and how do these limitations compare to traditional spectral methods?", "answer": "The proposed method can successfully solve chaotic PDEs in a limited time region but fails for large-time predictions due to the practical accuracy limit of neural network approximation to the spatial PDE (Eq.(4)), which is $10^{-6}~10^{-8}$. In contrast, traditional spectral methods achieve higher accuracy ($10^{-12}~10^{-16}$). Small errors in previous timestamps can lead to significant errors later due to the chaotic nature of the problem.", "category": "Method_Disambiguation"}, {"question": "What are the specific novel contributions of the paper beyond the application of implicit time discretization to PDEs?", "answer": "The paper contributes to the PINN community by rediscovering the good performance of discrete PINNs for solving evolutionary PDEs, both theoretically and numerically. It reduces computational time using transfer learning techniques and provides an error estimate result. For comparison to related works like Runge-Kutta,  classical numerical methods such as explicit Runge-Kutta suffer from stability constraints in time step and space step sizes, and implicit Runge-Kutta methods suffer from solving nonlinear equations many times, while the paper's method is stable and easy to approximate in space by neural networks in every time stamp.", "category": "Method_Disambiguation"}, {"question": "How are the weights (\\lambda values) chosen in the training process, and are they updated adaptively or kept static?", "answer": "The weights (\\lambda values) are chosen as constant values (e.g., 1, 100) throughout the training process, as the loss form for other baselines differs, making adaptive updating unsuitable.", "category": "Experimental_Setup"}, {"question": "What is the novelty of the paper in terms of using a time-discrete PINN type formulation with a standard time-stepping scheme, and how does it differ from prior work like Raissi et al.[1] or Mishra and Molinaro[2]?\n\n[1]Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686–707, 2019.\n\n[2]Siddhartha Mishra and Roberto Molinaro. Estimates on the generalization error of physics-informed neural networks for approximating PDEs. IMA Journal of Numerical Analysis, 43(1):1–43, 01 2023. ISSN 0272-4979. doi: 10.1093/imanum/drab093. URL https://doi.org/10.1093/ imanum/drab093.", "answer": "The novelty of the paper lies in rediscovering the good performance of discrete PINNs for solving evolutionary PDEs, both theoretically and numerically. Discrete PINNs are typically considered time-consuming and are rarely used in the PINNs literature. The paper has reduced computational time using transfer learning techniques and provided an error estimate result. Compared to related works like Runge-Kutta, classical numerical methods like explicit Runge-Kutta suffer from stability constraints on time and space step sizes, and implicit Runge-Kutta methods require solving nonlinear equations multiple times. In contrast, the proposed method is stable and can easily approximate the spatial domain using neural networks at every time step", "category": "Method_Comparison"}, {"question": "Is it sufficient to solely update the weights in the last hidden layer per time step to achieve high accuracy?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper numerically verify the second-order accuracy of the Crank-Nicolson scheme in its predictions?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the fine-tuning of all parameters performed between two adjacent timestamps?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "EyfOZKXpcN", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Improving Language Models via Plug-and-Play Retrieval Feedback", "authors": ["Wenhao Yu", "Zhihan Zhang", "Zhenwen Liang", "Meng Jiang", "Ashish Sabharwal"], "abstract": "Large language models (LLMs) exhibit remarkable performance across various NLP tasks. However, they often generate incorrect or hallucinated information, which hinders their practical applicability in real-world scenarios. Human feedback has been shown to effectively enhance the factuality and quality of generated content, addressing some of these limitations. However, this approach is resource-intensive, involving manual input and supervision, which can be time-consuming and expensive. Moreover, it cannot be provided during inference, further limiting its practical utility in dynamic and interactive applications. In this paper, we introduce ReFeed, a novel pipeline designed to enhance LLMs by providing automatic retrieval feedback in a plug-and-play framework without the need for expensive fine-tuning. ReFeed first generates initial outputs, then utilizes a retrieval model to acquire relevant information from large document collections, and finally incorporates the retrieved information into the in-context demonstration for output refinement, thereby addressing the limitations of LLMs in a more efficient and cost-effective manner. Experiments on four knowledge-intensive benchmark datasets demonstrate our proposed ReFeed could improve over +6.0% under zero-shot setting and +2.5% under few-shot setting, compared to baselines without using retrieval feedback.", "keywords": "Automatic feedback;information retrieval", "qa_pairs": [{"question": "How is the de-duplication process implemented to ensure diversity in the retrieved documents, and why was BM25 chosen as the retrieval method over other alternatives and What is the motivation for each component?", "answer": "The de-duplication process involves using each generated initial answer to retrieve a set of documents, calculating a similarity score for each document, merging all retrieved documents, and ranking them based on their similarity scores to keep only the top-k documents. This ensures diversity in the information gathered. BM25 was chosen as the retrieval method for two main reasons: (1) BM25 demonstrates excellent adaptability and generalizability across various contexts, and (2) dense retrieval methods like DPR, often trained with QA datasets or synthetic query-document pairs, might disproportionately reflect their ability to provide retrieval feedback, which could lead to inaccurate evaluations. BM25's performance within the retrieval feedback framework suggests potential for improvement with more advanced methods. The motivation for each component in the paper : (1) General Retrieval Feedback Pipeline: Language models often generate information that is either incorrect or hallucinated. The retrieved information allows the language model to reassess the initial outputs and, if necessary, refine them to produce new answers. (2) The diverse retrieval strategy enhances the diversity of the outputs and enables more comprehensive retrieval feedback based on these varied outputs. (3) The motivation for the ensemble method is that there are times when the retrieved documents may mislead the language model, leading to a correct output being revised into an incorrect one. So, the ensemble technique takes into account both the initial and refined outputs, ultimately enhancing the overall performance.", "category": "Motivation_Analysis"}, {"question": "Is the proposed method applicable to state-of-the-art models like GPT-4 and ChatGPT（gpt-3.5 turbo)?", "answer": "The paper conducted additional experiments using ChatGPT (gpt-3.5-turbo) and GPT-4, which reinforce the adaptability and effectiveness of the paper's approach across different model architectures. These results providing a comprehensive view of the paper's method's applicability.\n\n| Experiment on ChatGPT (gpt-3.5-turbo)                | NQ      | TriviaQA | HotpotQA |\n|--------------------------------------------------------|---------|----------|----------|\n| Closebook question answering (Question -> Answer)      | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer)  | 34.3 / 41.5 | 58.7 / 63.7 | 31.7 / 33.6 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 37.5 / 48.1 | 66.3 / 71.1 | 34.1 / 36.0 |\n\n| Experiment on GPT4 (gpt-4)                            | NQ      | TriviaQA | HotpotQA |\n|--------------------------------------------------------|---------|----------|----------|\n| Closebook question answering (Question -> Answer)      | 34.8 / 49.6 | 64.6 / 72.8 | 30.8 / 33.7 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer)  | 32.5 / 46.5 | 59.9 / 67.1 | 31.6 / 37.5 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 36.8 / 54.4 | 66.3 / 74.0 | 36.9 / 42.6|\n", "category": "Experimental_Analysis"}, {"question": "Why were Davinci-003 and Codex used as backbone models instead of ChatGPT and GPT-4?", "answer": "The paper’s initial decision to use Davinci-003 and Codex as backbone models was made to ensure a fair comparison with baseline methods (e.g., GenRead, RePLUG) that the paper has discussed. Besides, the ChatGPT and GPT-4 systems might involve large amounts of user data for instruction tuning, which could lead to contamination of many academic benchmark datasets, such as NQ and TriviaQA. Lastly, OpenAI’s announcement that both ChatGPT and GPT-4 will be subject to ongoing updates to their model parameters. These continual modifications would result in non-reproducible experiments, potentially compromising the reliability of the paper’s research outcomes.", "category": "Motivation_Analysis"}, {"question": "How does the hyper-parameter 'k' in module-1 affect the performance of the proposed method, and what experimental evidence supports this?", "answer": "Previous research, as mentioned in papers [1] and [2], indicates that increasing the value of ‘k’ beyond 10 results in only marginal improvements, while also increasing complexity. Additionally, as noted in [3], expanding the context length excessively for LLMs can result in the ‘lost in the middle’ issue, i.e., LLMs ignore the middle part of a long text input. To further investigate the effect of varying ‘k,’ the authors conducted experiments using ChatGPT (gpt-3.5-turbo) on datasets such as NQ, TriviaQA, and HotpotQA. The results are shown in the following tables (metric: left EM, right accuracy). The experimental results are consistent with the findings in [1] and [2]. As shown in the table, when ‘k’ is set to 10, it performs significantly better than when ‘k’ is set to 1 or 5. However, when ‘k’ is set to 20, the improvement is marginal compared to when ‘k’ is set to 10. When ‘k’ is set to a small number, such as 1, the language model could be misled because it tends to trust the single given document.\n\n\n| K=1 with ChatGPT (gpt-3.5-turbo) | NQ       | TriviaQA | HotpotQA |\n|----------------------------------|----------|----------|----------|\n| Closebook question answering (Question -> Answer) | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer) | 23.6 / 30.0 | 40.5 / 44.8 | 23.2 / 24.0 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 34.0 / 41.9 | 65.9 / 68.7 | 27.1 / 28.3 |\n\n| K=5 with ChatGPT (gpt-3.5-turbo) | NQ       | TriviaQA | HotpotQA |\n|----------------------------------|----------|----------|----------|\n| Closebook question answering (Question -> Answer) | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer) | 30.5 / 37.6 | 50.4 / 55.4 | 28.5 / 29.0 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 34.3 / 42.1 | 67.5 / 70.8 | 30.0 / 30.4 |\n\n| K=10 with ChatGPT (gpt-3.5-turbo) | NQ       | TriviaQA | HotpotQA |\n|-----------------------------------|----------|----------|----------|\n| Closebook question answering (Question -> Answer) | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer) | 34.3 / 41.5 | 58.7 / 63.7 | 31.7 / 33.6 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 37.5 / 48.1 | 66.3 / 71.1 | 34.1 / 36.0 |\n\n| K=20 with ChatGPT (gpt-3.5-turbo) | NQ       | TriviaQA | HotpotQA |\n|-----------------------------------|----------|----------|----------|\n| Closebook question answering (Question -> Answer) | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer) | 34.6 / 42.0 | 59.5 / 64.5 | 31.8 / 33.8 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 37.8 / 48.2 | 66.5 / 71.2 | 34.4 / 36.5 |\n\n[1] Prompting GPT-3 To Be Reliable. ICLR 2023 \n\n[2] REPLUG: Retrieval-Augmented Black-Box Language Models. EMNLP 2023.\n\n[3] Lost in the Middle: How Language Models Use Long Contexts. TACL 2023.", "category": "Experimental_Exposition"}, {"question": "How does the proposed approach generalize and perform across different backbone models, such as ChatGPT (gpt-3.5 turbo) and GPT-4, compared to the original Davinci-003 and Codex?", "answer": "The paper’s initial decision to use Davinci-003 and Codex as backbone models was made to ensure a fair comparison with baseline methods (e.g., GenRead, RePLUG) that the paper has discussed. Besides, the ChatGPT and GPT-4 systems might involve large amounts of user data for instruction tuning, which could contaminate many academic benchmark datasets, such as NQ and TriviaQA. Lastly, OpenAI’s announcement states that both ChatGPT and GPT-4 will be subject to ongoing updates to their model parameters. These continual modifications would lead to non-reproducible experiments, potentially compromising the reliability of the paper’s research outcomes. The paper conducted additional experiments using ChatGPT (gpt-3.5 turbo) and GPT-4, which reinforce the adaptability and effectiveness of the paper’s approach across different model architectures. These results provide a comprehensive view of the paper’s method’s applicability.\n\n| Experiment on ChatGPT (gpt-3.5-turbo)                  | NQ          | TriviaQA    | HotpotQA    |\n|--------------------------------------------------------|-------------|-------------|-------------|\n| Closebook question answering (Question -> Answer)      | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer)   | 34.3 / 41.5 | 58.7 / 63.7 | 31.7 / 33.6 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 37.5 / 48.1 | 66.3 / 71.1 | 34.1 / 36.0 |\n\n| Experiment on GPT4 (gpt-4)                             | NQ          | TriviaQA    | HotpotQA    |\n|--------------------------------------------------------|-------------|-------------|-------------|\n| Closebook question answering (Question -> Answer)      | 34.8 / 49.6 | 64.6 / 72.8 | 30.8 / 33.7 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer)   | 32.5 / 46.5 | 59.9 / 67.1 | 31.6 / 37.5 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 36.8 / 54.4 | 66.3 / 74.0 | 36.9 / 42.6 |\n\nFrom the above table, it can be observed:\n\n(1) The retrieval-feedback consistently improves the system, compared to both standard closed-book QA methods and the retrieve-then-read pipeline.\n\n(2) The traditional retrieve-then-read system is even sometimes hurt by noisy retrieval. On the contrary, the paper’s ReFeed can avoid misleading issues if the first-round generation is good.", "category": "Experimental_Analysis"}, {"question": "How can one verify that the language model (LLM) is well-calibrated to determine the plausibility of a refinement, and what steps are taken if the model is not well-calibrated?", "answer": "Verifying that the language model is well-calibrated is challenging, and the method assumes a relatively well-calibrated model. If the model is not well-calibrated, uncertainty can be measured using semantic-based metrics instead of relying solely on probabilities. The paper includes a limitation discussion to highlight this issue, noting that the ensemble strategy is only applicable to well-calibrated models. Additionally, the main retrieval feedback framework does not depend on model calibration.", "category": "Method_Mechanics"}, {"question": "What are the novel contributions of the paper compared to the related work by Jiang et al. on Active Retrieval Augmented Generation (2023)?", "answer": "Although there are similarities in the pipeline (generate-retrieve-regenerate) between Active RAG and the paper’s ReFeed approach, the paper offers distinct contributions compared to Active RAG. Firstly, ReFeed employs multi-pass decoding (diverse generation) to facilitate more comprehensive retrieval feedback based on varying outputs, addressing a limitation in Active RAG’s one-pass decoding, which may limit potential matches to diverse relevant information. Secondly, ReFeed introduces an ensemble strategy that re-evaluates the trustworthiness of answers to prevent retrieved documents from misleading the language model, enhancing the reliability of the output. These contributions present novel approaches to managing the complexities of retrieve-then-read pipelines.", "category": "Method_Comparison"}, {"question": "What are the specific steps and mechanisms involved in integrating the two enhanced modules together?", "answer": "The integration involves two distinct steps: (1) Diverse generation is implemented in the first step by generating multiple diverse initial answers via sampling, retrieving documents for each answer, and removing duplicates. (2) The ensemble method is applied in the second step to produce the final answer, which is based on the likelihood of all initial and refined answers.", "category": "Method_Mechanics"}, {"question": "What was the rationale behind selecting text-davinci-003 and Code-Davinci-002 (Codex) as the backbone language models for the study?", "answer": "The selection of text-davinci-003 and Code-Davinci-002 (Codex) as backbone models was based on ensuring a fair comparison with baseline methods (e.g., GenRead, RePLUG) discussed in the paper. Additionally, ChatGPT and GPT-4 were not chosen because they might involve user data for instruction tuning, potentially contaminating academic benchmark datasets like NQ and TriviaQA. Furthermore, ongoing updates to ChatGPT and GPT-4 could lead to non-reproducible experiments, compromising research reliability. Additional experiments with ChatGPT (gpt-3.5 turbo) and GPT-4 were conducted to demonstrate the adaptability and effectiveness of the approach across different model architectures.", "category": "Motivation_Analysis"}, {"question": "What are the key observations from the experiments conducted on ChatGPT (gpt-3.5-turbo) and GPT-4 regarding the retrieval-feedback mechanism and the traditional retrieve-then-read system?", "answer": "\n| Experiment on ChatGPT (gpt-3.5-turbo)                  | NQ          | TriviaQA    | HotpotQA    |\n|--------------------------------------------------------|-------------|-------------|-------------|\n| Closebook question answering (Question -> Answer)      | 32.3 / 39.9 | 65.6 / 69.5 | 23.5 / 24.0 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer)   | 34.3 / 41.5 | 58.7 / 63.7 | 31.7 / 33.6 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 37.5 / 48.1 | 66.3 / 71.1 | 34.1 / 36.0 |\n\n| Experiment on GPT4 (gpt-4)                             | NQ          | TriviaQA    | HotpotQA    |\n|--------------------------------------------------------|-------------|-------------|-------------|\n| Closebook question answering (Question -> Answer)      | 34.8 / 49.6 | 64.6 / 72.8 | 30.8 / 33.7 |\n| Retrieve-then-Read (Question -> Retrieval -> Answer)   | 32.5 / 46.5 | 59.9 / 67.1 | 31.6 / 37.5 |\n| ReFeed (Question -> Answer -> Retrieval -> Refinement) | 36.8 / 54.4 | 66.3 / 74.0 | 36.9 / 42.6 |\n\nFrom the experiments, it was observed that: (1) The retrieval-feedback mechanism consistently improves the system compared to both the standard closebook QA method and the retrieve-then-read pipeline. (2) The traditional retrieve-then-read system is sometimes hindered by noisy retrieval, while the ReFeed mechanism avoids being misled if the first-round generation is accurate.", "category": "Experimental_Exposition"}, {"question": "How does REFEED avoid generating incorrect or hallucinated information?", "answer": "REFEED avoids generating incorrect or hallucinated information by using retrieval feedback to validate the correctness of the generated output. Evaluations based on open-domain factual question-answering demonstrate that REFEED significantly reduces hallucination and improves factuality, as shown in Figure 5 of the paper.", "category": "Method_Mechanics"}, {"question": "What steps have been taken to address concerns about qualitative evaluation, and why does Table 1 lack implementation details?", "answer": "For qualitative evaluation, the methods were evaluated in a few-shot QA/generation setting across different domains, focusing on commonsense QA[1] and open-domain QA[2], aiming to improve large language model performance like GPT-3. Table 1 highlights the main differences in method design and application domain, excluding implementation details.\n\n[1] Rethinking with retrieval: Faithful large language model inference.\n\n[2] Check your facts and try again: Improving large language models with external knowledge and automated feedback.", "category": "Experimental_Setup"}, {"question": "Were experiments conducted on conventional smaller neural language models like BART and T5 as part of the initial study?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are comprehensive details for reproducibility included in the main paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were all methods evaluated in a few-shot QA/generation setting across different domains, focusing on commonsense QA [1] and open-domain QA [2]?\n\n[1] Rethinking with retrieval: Faithful large language model inference.\n\n[2] Check your facts and try again: Improving large language models with external knowledge and automated feedback.", "answer": "True", "category": "Claim_Verification"}, {"question": "Does Table 1 provide clear settings and implementation details for the methods?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "RadQVWAucN", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "LLM-Rec: Personalized Recommendation via Prompting Large Language Models", "authors": ["Hanjia Lyu", "Song Jiang", "Hanqing Zeng", "Yinglong Xia", "Qifan Wang", "Si Zhang", "Chris Leung", "Ren Chen", "Jiajie Tang", "Jiebo Luo"], "abstract": "Text-based recommendation holds a wide range of practical applications due to its versatility, as textual descriptions can represent nearly any type of item. However, directly employing the original item descriptions as input features may not yield optimal recommendation performance. This limitation arises because these descriptions often lack comprehensive information that can be effectively exploited to align with user preferences. Recent advances in large language models (LLMs) have showcased their remarkable ability to harness commonsense knowledge and reasoning. In this study, we investigate diverse prompting strategies aimed at $\\textit{augmenting the input text}$ to enhance personalized text-based recommendations. Our novel approach, coined  $\\textbf{LLM-Rec}$, encompasses four distinct prompting techniques: (1) basic prompting, (2) recommendation-driven prompting, (3) engagement-guided prompting, and (4) recommendation-driven + engagement-guided prompting. Our empirical experiments show that incorporating the augmented input text generated by the LLMs yields discernible improvements in recommendation performance. Notably, the recommendation-driven and engagement-guided prompting strategies exhibit the capability to tap into the language model's comprehension of both general and personalized item characteristics. This underscores the  significance of leveraging a spectrum of prompts and input augmentation techniques to enhance the recommendation prowess of LLMs.", "keywords": "recommendation;large language model;input augmentation;personalization", "qa_pairs": [{"question": "How does the engagement-guided prompting strategy utilize neighbor information, and what specific elements are included in the prompt to guide the LLM?", "answer": "The engagement-guided prompting strategy includes the descriptions of other important neighbors, not just item names. The prompt is 'Summarize the commonalities among the following descriptions: ‘description’; ‘descriptions of other important neighbors’', where 'description' refers to the target item's description and 'descriptions of other important neighbors' refers to the item descriptions of other important neighbors, as shown in Figure 1 and Section 2.", "category": "Method_Mechanics"}, {"question": "What are the potential limitations and challenges of using LLMs for personalized recommendation, specifically regarding computational complexity, model interpretability, and potential biases in the generated prompts?", "answer": "The potential limitations and challenges of using LLMs for personalized recommendation include:\n\n**Computational Complexity:** The LLM-Rec framework focuses on enhancing input for personalized recommendations through generating item descriptions. This approach addresses the problem of original item descriptions often being incomplete or insufficient for effective recommendations. The primary computational load in LLM-Rec arises during the augmentation phase including the output text length, number of items, and the concatenation of text embeddings. \n\nFor output text length, the paper's findings indicate that selecting important words for inclusion, rather than incorporating all response words, can lead to improved recommendation performance, as evidenced in Table 3. Future research could explore the balance between the number of words generated and the resulting performance enhancements. \n\nFor the number of items, to mitigate this, the paper proposes potential strategies to reduce the need for augmentation across numerous descriptions. One approach is developing a model that evaluates whether an original item description already possesses sufficient information for recommendations. This avenue warrants further exploration in subsequent studies to decrease computational demands.\n\nFor the concatenation of text embeddings, as demonstrated in Table 2 and Figure 6, combining the embeddings of augmented text substantially enhances performance. However, this also increases the input dimension for the recommendation module, potentially raising computational costs. A solution could be to opt for a simpler model structure, such as an MLP, which the paper has implemented in LLM-Rec, to manage this increase in computational complexity.\n\n**Model Interpretability:** LLM-Rec introduces an innovative approach to model interpretability by enabling direct comparisons between variations in generated prompts and corresponding shifts in recommendation performance. This relationship is illustrated in the paper's research, as showcased in Figures 4 and 5, as well as Tables 3. Nonetheless, delving into a more granular, word-level analysis could yield deeper insights. By examining both the generated responses and the subsequent recommendation module, a more nuanced understanding of how specific textual changes influence overall model behavior can be gained, enhancing comprehension of model interpretability.\n\n**Potential Biases in generated responses:** The presence of potential biases in the prompts generated by LLMs is a significant concern, primarily because these biases are deeply ingrained in the data used to train the LLMs. This challenge calls for a multifaceted approach, including algorithmic adjustments, ethical guidelines, and continuous monitoring for bias, which are important future directions.", "category": "Method_Disambiguation"}, {"question": "Why does the performance decrease when combining recommendation-driven and engagement-guided strategies in Table 1, while Figure 6 shows different results?", "answer": "The findings do not contradict each other. Table 1 reflects the performance of a strategy where LLMs are instructed to generate descriptions with both recommendation-driven and engagement-guided objectives, leaving the LLMs to determine the amount of information for each component. Figure 6, however, shows the results of directly concatenating the embeddings of separately generated prompts, leaving the recommendation module to determine the amount of information for each component.", "category": "Concept_Understanding"}, {"question": "How can the limitations of prompt template design sustainability be addressed, given that Table 1 shows that basic prompts outperform more complex prompts on Llama2?", "answer": "By adding experiments with prompts that have the same meaning but differ in wording to evaluate the consistency of LLM-Rec. The Llama-2-Chat variant (7B parameters) may not boost recommendation performance as much as GPT-3, but the findings remain consistent. Adding extra information augmented by LLMs outperforms models using only original item descriptions, and concatenating augmented text further improves performance, enabling simpler models like MLP to achieve comparable or better results than more complex feature-based models.", "category": "Method_Mechanics"}, {"question": "How does the proposed method, LLM-Rec, perform in scenarios where datasets lack textual information?", "answer": "LLM-Rec augments the input of items relevant to rich textual content, so the dataset itself does not need to contain rich textual information. For example, the Movielens dataset initially lacks movie descriptions. The limitation is whether the item is associated with rich features, which is a significant issue in recommendation systems.", "category": "Experimental_Exposition"}, {"question": "Is there an experiment specifically validating the effectiveness of engagement-guided prompting in utilizing neighbor information to guide the LLM?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the engagement-guided prompting strategy focus solely on item names without including descriptions of important neighbors?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "jolYuxpVn1", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "FacTool: Factuality Detection in Generative AI - A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios", "authors": ["Ethan Chern", "Steffi Chern", "Shiqi Chen", "Weizhe Yuan", "Kehua Feng", "Chunting Zhou", "Junxian He", "Graham Neubig", "Pengfei Liu"], "abstract": "The emergence of generative pre-trained models has facilitated the synthesis of high-quality text but has also posed challenges in identifying factual errors in the generated text. In particular: (1) A wider range of tasks now face an increasing risk of containing factual errors when handled by generative models. (2) The content generated by these models tends to be lengthy and lacks clearly defined granularity for individual facts. (3) There is a scarcity of explicit evidence available during the process of fact checking. With the above challenges in mind, in this paper, we propose FacTool, a tool augmented multi-task and multi-domain framework for detecting factual errors of texts generated by large language models (e.g., ChatGPT). Experiments on four different tasks (knowledge-based QA, code generation, mathematical reasoning, and scientific literature review) show the efficacy of the proposed method.", "keywords": "Factuality;LLM", "qa_pairs": [{"question": "Why are the open-source non-LLM methods[1]  and[2] not suitable for the main objective of the paper on factuality detection in Generative AI across multi-task and multi-domain scenarios?\n\n[1] https://github.com/titipata/detecting-scientific-claim\n\n[2] https://github.com/dwadden/multivers\n", "answer": "The open-source non-LLM methods are not suitable for the paper’s objective because: factuality detection in Generative AI across multi-task and multi-domain scenarios. Upon trying to run the open-source code, both of the repositories are not suitable for the paper’s purpose, as detailed below:\n\n**Regarding [1]** The claim extraction setting of [1] differs significantly from the paper's setting. According to their paper [3], the claim extraction task is defined as a binary classification task: predicting whether a 'sentence' in a paragraph is a 'claim' or not. However, (i) the paper defines 'claim' as atomic facts, code snippets, math calculations, or Tuples (e.g., paper title, year, authors) extracted from long-form text. This is clearly distinct from the definition of a 'sentence' in [3]. (ii) Additionally, the paper's claim extraction task involves extracting all verifiable claims from the generated text $x$ (denoted as $c_1, c_2, \\cdots, c_n$), which is a generative task, unlike the binary classification task defined in [3].\n\n**Regarding [2]** The scientific claim verification task defined in [4] requires a system to predict a label $y(c, a) \\in \\{\\text{SUPPORTS}, \\text{REFUTES}, \\text{NOT ENOUGH INFO}\\}$ given a claim $c$ and a collection of candidate abstracts potentially containing relevant evidence. This differs from the paper's task, where claims are not provided. The task of factuality detection in Generative AI requires a pipeline that can provide fine-grained claim-level factuality without given claims. Moreover, directly applying the MultiVerS model for agreement verification is impractical because it is trained on specific datasets (HealthVer, COVIDFact, SCIFACT), and its effectiveness in general domains and tasks like KBQA is unknown. Further fine-tuning is likely required. Given the superior capabilities of open-source LLMs, fine-tuning them would be a better choice than using non-LLM-based methods. Versatile claim extraction is highly non-trivial and posed significant challenges before the era of LLMs. The limitations of non-LLM methods in achieving this task underscore its complexity, thereby justifying the non-trivial nature of versatile claim extraction. In contrast, non-LLM-based methods are limited to specific scenarios, requiring significant effort to annotate datasets for training claim extractors with limited generalization ability.\n\n\n[3] Achakulvisut et al. Claim Extraction in Biomedical Publications Using Deep Discourse Model and Transfer Learning \n\n[4] Wadden et al. MULTIVERS: Improving scientific claim verification with weak supervision and full-document context", "category": "Experimental_Analysis"}, {"question": "What are the key innovations of FacTool's factuality detection framework beyond the claim extraction module, and how do they differentiate it from prior art?", "answer": "FacTool's factuality detection framework distinguishes itself from prior art in two key areas: (i) It simplifies the agreement verification process by utilizing the advanced reasoning abilities of LLMs to directly process collected evidence, avoiding the complexities of relevance matching that systems like RARR entail. (ii) It facilitates the integration of various tools into the factuality detection pipeline, making it a versatile approach capable of factuality detection across multiple scenarios and tasks, unlike prior art which typically focuses on a single task and scenario.", "category": "Method_Disambiguation"}, {"question": "What are the key differences between FacTool and RARR (Gao et al., 2022a)[1] in terms of claim extraction, agreement verification, and focus areas?\n\n[1]Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Y. Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. Rarr: Researching and revising what language models say, using language models, 2022a.\n", "answer": "The key differences between FacTool and RARR are: (i) FacTool provides versatile claim extraction, enabling not only the provision of fine-grained factuality information to help users more easily detect potential factual errors but also the integration of factuality detection across various domains and scenarios. In contrast, RARR neither supports fine-grained factuality nor multi-task factuality.  (ii) FacTool offers a more intuitive and straightforward agreement verification process. FacTool simplifies the agreement verification process by directly feeding collected evidence to LLMs, eliminating the need for relevance matching, whereas RARR requires parsing webpages and using a relevance matching algorithm. (iii)  FacTool focuses on 'factuality detection', i.e., detecting factual errors from the generated text from LLMs, while RARR focuses on 'text editing', i.e., editing text that may contain hallucinations.", "category": "Method_Comparison"}, {"question": "Why were notable fact-checking datasets like FEVER and SciFact excluded from the evaluation, despite the paper's emphasis on fact-checking?", "answer": "The paper focuses on the growing challenges of factual errors generated by models in generative AI like ChatGPT. To this end, the paper has constructed a dataset based on the natural distribution of errors generated by these models. Unlike fact-checking datasets like FEVER and SciFact, which do not reflect the natural distribution of errors generated by LLMs, this study did not focus on these datasets. Additionally, both FEVER and SciFact provide claims, which do not align with the paper’s task setting: offering an end-to-end system capable of delivering detailed, claim-level factuality information from long-form context.", "category": "Motivation_Analysis"}, {"question": "What were the results of the previous LLM systems as baseline comparisons for evidence-based fact-checking, and why were non-LLM models not included?", "answer": "**previous LLM-based systems**, The authors integrate FacTool with Llama2 and other open-source models, and provide detailed results for various tasks. Detailed results are:\n\n| Methods   | Tasks        | Foundation Model   | Claim-Level |      |      |      | Response-Level |      |      |      |\n|-----------|--------------|--------------------|-------------|------|------|------|----------------|------|------|------|\n|           |              |                    | Accuracy    | Recall | Precision | F1   | Accuracy    | Recall | Precision | F1   |\n| FACTOOL   | Knowledge-QA | llama2-70b-chat    | 73.99       | 93.49 | 77.07      | 84.49| 52.00       | 78.26 | 48.65      | 60.00|\n| FACTOOL   | Code         | llama2-70b-chat    | 55.75       | 50.00 | 76.00      | 60.32| 55.75       | 50.00 | 76.00      | 60.32|\n| FACTOOL   | Math         | llama2-70b-chat    | 86.62       | 100.00| 86.62      | 92.83| 47.00       | 100.00| 47.00      | 63.95|\n| FACTOOL   | Scientific   | llama2-70b-chat    | 96.77       | 81.82 | 100.00     | 90.00| 98.00       | 80.00 | 100.00     | 88.89|\n\nThe error detection ability of FacTool powered by Llama-2-70b-chat is slightly constrained due to Llama-2-70b's limited coding capabilities (Llama-2-70b scores only 29.9% on the HumanEval benchmark, in contrast to ChatGPT-3.5's 48.1%).\n\n**Regarding non-LLM-based systems**, it is challenging to find a non-LLM approach capable of performing the task defined here: the fine-grained assessment of factuality in long texts from various scenarios. Previous methods often focused on specific aspects; for example, FEVER judges the factuality of a given claim, which is a relatively simpler task than the one defined here. Developing a non-LLM method from scratch to conduct fine-grained assessment of factuality in long texts from various scenarios would require a substantial engineering effort, so significant that it might be considered a separate project.\n", "category": "Experimental_Exposition"}, {"question": "How does the fact-checking methodology of FacTool differ from prior evidence-based fact-checking methods that use search engines for evidence retrieval?", "answer": "The fact-checking methodology of FacTool differs from prior works that use search engines for evidence retrieval in evidence-based fact-checking, with its **versatile claim extraction** module. This module effectively provides **fine-grained factuality information for long-form text across various domains and scenarios** that enables a **unified framework for factuality detection across different scenarios**. Classical evidence-based fact-checking methods [1, 2, 3, 4] typically follow a three-step pipeline: document retrieval, sentence retrieval, and claim verification. This pipeline is designed to address scenarios where the **claims are already given**. In contrast, FacTool is capable of fact-checking **long-form texts in which claims are not explicitly given**. More recent prior works such as RARR [5] utilize search engines for evidence retrieval but also do not employ a claim extraction process to collect detailed factuality information. Other concurrent works like FactScore [6] employ a claim extraction module to extract atomic facts from long-form texts but are limited to the Knowledge-Based Question Answering (KBQA) task. The paper's work expands the definition of factuality beyond the KBQA task and does not solely focus on using search engines for evidence retrieval. FacTool is designed to be versatile and capable of handling multiple tasks, integrating various tools such as search engines, Python executors, and Google Scholar, as demonstrated in the paper. This comprehensive integration enables FacTool to tackle a wider spectrum of factuality issues in generative AI (GAI), overcoming the limitations of existing works that are focused on single tasks or specific scenarios. \n\nReferences: [1] Bekoulis et al.. A review on fact extraction and verification \n\n[2] Thorne et al. FEVER: a large-scale dataset for fact extraction and VERification \n\n[3] Zhong et al. Reasoning Over Semantic-Level Graph for Fact Checking \n\n[4] Zhou et al. GEAR: Graphbased Evidence Aggregating and Reasoning for Fact Verification \n\n[5] Gao et al. Rarr: Researching and revising what language models say, using language models \n\n[6] Min et al. FACTSCORE: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation", "category": "Method_Comparison"}, {"question": "How was the baseline method Self-Check set up, and were the prompts developed by the authors or based on existing code?", "answer": "The prompts for Self-Check were developed by the authors, inspired by the Self-Refine method. Self-Refine, a strong baseline for LLMs to refine their responses, starts by generating an output and then feeding it back into the same LLM for feedback. This aligns with the focus on factuality detection, so the first step of Self-Refine was adapted as the Self-Check baseline.", "category": "Experimental_Setup"}, {"question": "Why were existing methods for factual error detection like RARR and SELF-REFINE not included as baselines in the evaluations, particularly for KBQA?", "answer": "RARR was not included because it focuses on text editing and lacks a claim extraction process, which differs from the paper's focus on fine-grained claim-level factuality detection. Instead, the paper included 'Self-Check', the first step of 'Self-Refine', as the paper's baseline, aligning with the paper's focus on detection rather than refinement.", "category": "Motivation_Analysis"}, {"question": "How does the paper define and evaluate the factuality of code snippets in the context of code generation, given that there can be multiple valid ways to write a logical code snippet?", "answer": "The factuality of a code snippet is defined and measured using an execution-based approach, which involves executing the generated code against test case inputs and comparing its output to the expected (or golden) output to determine correctness.", "category": "Method_Mechanics"}, {"question": "What are the key differences between FacTool and RARR[1], and how does FacTool's approach to factuality detection and agreement verification compare to RARR's approach?Given that the intermediate question generation module to validate any claims made by the LLM against an external evidence has been proposed in other works ([1, 2])\n\n[1]https://arxiv.org/pdf/2210.08726.pdf\n\n[2]https://arxiv.org/pdf/2309.11495.pdf", "answer": "The key differences between FacTool and RARR are:(i) FacTool provides versatile claim extraction, enabling not only the provision of fine-grained factuality information to help users more easily detect potential factual errors but also the integration of factuality detection across various domains and scenarios. In contrast, RARR neither supports fine-grained factuality nor multi-task factuality. (ii) FacTool offers a more intuitive and straightforward agreement verification process. After evidence collection, the paper leverages the strong reasoning capabilities of LLMs to directly feed LLMs with the collected evidence, eliminating the need for extra steps in relevance matching. RARR requires an extra step to parse over webpages and uses a relevant matching algorithm to find the most relevant snippets, which is more complicated compared to FacTool. (iii) FacTool focuses on detecting factual errors in generated text, while RARR focuses on editing text containing hallucinations. Additionally, the work cited  [2]  was completed after FacTool and cites a pre-print of the FacTool paper.", "category": "Method_Comparison"}, {"question": "What are the novel technical contributions of the paper, particularly in terms of its unified framework for detecting factual errors and its applicability across tasks and domains?", "answer": "The novel technical contributions of the paper include: (i) A motivation to address the increasing risk of factual errors in generative AI tasks by providing a unified framework for detecting factual errors across various tasks and domains. (ii) FacTool, a unified tool-augmented framework for detecting factual errors across diverse tasks and domains, including KBQA, coding, math, and scientific. (iii) An automatic factuality evaluator for modern chatbots across different scenarios, which maintains a factuality leaderboard for chatbots and helps developers detect potential factual errors generated by their LLMs. (iv) A user-friendly API interface that supports ChatGPT plugin, making the tool accessible to both technical and general users.", "category": "Method_Disambiguation"}, {"question": "What steps are taken to address the reliance on ChatGPT for implementing the framework, and how does the integration of Llama2 and other open-source models affect the reproducibility and cost-effectiveness of the results?", "answer": "The authors worked on integrating FacTool with Llama2 and various other open-source models.Detailed results are:\n\n| Methods | Tasks        | Foundation Model | Claim-Level |      |      |      | Response-Level |      |      |      |\n|---------|--------------|------------------|-------------|------|------|------|----------------|------|------|------|\n|         |              |                  | Accuracy    | Recall | Precision | F1   | Accuracy    | Recall | Precision | F1   |\n| FACTOOL | Knowledge-QA | llama2-70b-chat  | 73.99       | 93.49 | 77.07      | 84.49| 52.00       | 78.26 | 48.65      | 60.00|\n| FACTOOL | Code         | llama2-70b-chat  | 55.75       | 50.00 | 76.00      | 60.32| 55.75       | 50.00 | 76.00      | 60.32|\n| FACTOOL | Math         | llama2-70b-chat  | 86.62       | 100.00| 86.62      | 92.83| 47.00       | 100.00| 47.00      | 63.95|\n| FACTOOL | Scientific   | llama2-70b-chat  | 96.77       | 81.82 | 100.00     | 90.00| 98.00       | 80.00 | 100.00     | 88.89|\n\nThe error detection ability of FacTool powered by Llama-2-70b-chat is slightly constrained due to Llama-2-70b's limited coding capabilities (Llama-2-70b scores only 29.9% on the HumanEval benchmark, in contrast to ChatGPT-3.5's 48.1%).\n", "category": "Experimental_Exposition"}]}
{"id": "w8eCnnq57m", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Composition", "authors": ["Chengsong Huang", "Qian Liu", "Bill Yuchen Lin", "Chao Du", "Tianyu Pang", "Min Lin"], "abstract": "Low-rank adaptations (LoRA) are often employed to fine-tune large language models (LLMs) for new tasks. This paper investigates LoRA composability for cross-task generalization and introduces LoraHub, a simple framework devised for the purposive assembly of LoRA modules trained on diverse given tasks, with the objective of achieving adaptable performance on unseen tasks. With just a few examples from a new task, LoraHub can fluidly combine multiple LoRA modules, eliminating the need for human expertise and assumptions.  Notably, the composition requires neither additional model parameters nor gradients.  Our empirical results, derived from the Big-Bench Hard benchmark, suggest that LoraHub can effectively mimic the performance of in-context learning in few-shot scenarios, excluding the necessity of in-context examples alongside each inference input.", "keywords": "Parameter Efficient Tuning;LoRA;Cross-Task Generalization", "qa_pairs": [{"question": "What is the practical impact of the reduced number of tokens required in the proposed LoraHub method, and how does it benefit practitioners despite the initial training burden of several LORA modules?", "answer": "While the in-context learning (ICL) for generalization is effective. The paper's proposed LoraHub is not intended to replace ICL but rather to offer a complementary strategy with performance-efficiency trade-offs. An essential benefit of LoraHub, highlighted in the paper, is its focus on improving inference efficiency by utilizing fewer prompt tokens per example. As illustrated in Table 1, LoraHub demonstrates a significant advantage with a reduced token requirement per example — `111.6` tokens compared to the `597.8` tokens needed for ICL. The efficiency of inference is paramount for real-world deployments, particularly when dealing with lots of inference examples. With fewer tokens per example during inference, the paper's method significantly reduces computational overhead and enables faster responses. The commitment to optimizing inference efficiency aligns seamlessly with a broader research trend, as evidenced by several recent studies actively exploring approaches to reduce the number of input tokens [1,2,3,4,5,6]. For example, as stated in [1], in-context learning can be inefficient because it makes the input prompt much longer, consuming valuable space in the context window and leading to larger computational costs. \n\n[1]. Wangchunshu Zhou, Yuchen Eleanor Jiang, Ryan Cotterell, Mrinmaya Sachan, Efficient Prompting via Dynamic In-Context Learning. ArXiv abs/2305.11170 (2023) \n\n[2]. Tao Ge, Jing Hu, Lei Wang, Xun Wang, Si-Qing Chen, Furu Wei, In-context Autoencoder for Context Compression in a Large Language Model. ArXiv abs/2307.06945 (2023) \n\n[3]. Alexis Chevalier, Alexander Wettig, Anirudh Ajith, Danqi Chen, Adapting Language Models to Compress Contexts. ArXiv abs/2305.14788 (2023) \n\n[4]. Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, Lili Qiu, LongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression. ArXiv abs/2310.06839 (2023) \n\n[5]. Yucheng Li, Bo Dong, Chenghua Lin, Frank Guerin, Compressing Context to Enhance Inference Efficiency of Large Language Models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing \n\n[6]. Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, Lili Qiu, LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing.", "category": "Experimental_Analysis"}, {"question": "How does the performance of LoraHub compare to in-context learning (ICL) in terms of both average and best performance across tasks, and what is the rationale behind the observed performance gap?", "answer": "Although the paper’s current approach lags behind ICL in terms of average performance, the inferior performance is primarily an artifact of the paper’s implementation rather than an inherent limitation of LoraHub. It is believed that the observed gap can be narrowed through the development of better module composition algorithms. This conviction finds support in the new experimental results (below) on the challenging Big-Bench Hard benchmark. The new experimental results report the best performance of ICL. The authors conducted a meticulous examination, collecting three distinct sets of few-shot examples, ensuring a comprehensive evaluation that covers both average (Table 1) and best (Table 3) performances across all tasks for various methods. The best performance of both ICL and LoraHub is as below.\n\n| Task                                      | ICL$_{best}$ | LoraHub$_{best}$ |\n|-------------------------------------------|-------------|-----------------|\n| Boolean Expressions                       | 62.7        | 60.7            |\n| Causal Judgement                          | 59.8        | 63.2            |\n| Date Understanding                        | 21.3        | 45.3            |\n| Disambiguation                             | 69.3        | 68.0            |\n| Dyck Languages                             | 2.0         | 2.7             |\n| Formal Fallacies                          | 59.3        | 59.3            |\n| Geometric Shapes                           | 20.0        | 18.7            |\n| Hyperbaton                                | 72.7        | 72.7            |\n| Logical Deduction (three objects)         | 52.7        | 52.7            |\n| Logical Deduction (five objects)          | 39.3        | 40.0            |\n| Logical Deduction (seven objects)         | 42.0        | 46.0            |\n| ...                                       |             |                 |\n| Tracking Shuffled Objects (three objects) | 31.3        | 31.3            |\n| Tracking Shuffled Objects (five objects)  | 12.0        | 16.7            |\n| Tracking Shuffled Objects (seven objects) | 6.7         | 15.3            |\n| Web of Lies                               | 54.0        | 57.3            |\n| Word Sorting                              | 0.7         | 1.3             |\n| **Avg Performance Per Task**              | **38.4**    | **41.2**        |\n| **Avg Tokens Per Example**                | **597.8**   | **111.6**       |\n\nThe findings indicate that LoraHub's best performance surpasses that of ICL on 18 tasks, highlighting its promising potential for future development. This underscores the untapped capabilities of LoraHub, emphasizing its capacity to excel in challenging tasks.", "category": "Method_Comparison"}, {"question": "What is the trade-off between the computational costs of in-context learning and LoraHub, particularly when considering one-time ad-hoc tasks versus recurring tasks?", "answer": "LoraHub addresses the challenge of reducing inference costs by eliminating the need for processing additional tokens, resulting in a noticeable reduction in overall inference expenses. However, it indeed introduces an inherent cost during the Adapt stage, necessitating extra inference steps. This introduces a trade-off between choosing the in-context learning approach and LoraHub, with the decision typically hinging on the nature of the situation. For one-time ad-hoc tasks, the in-context learning approach should be more pragmatic due to LoraHub's additional inference step costs. In such scenarios, where immediate, single-use solutions are preferred, the simplicity and efficiency of in-context learning might outweigh the benefits of potential savings offered by LoraHub. Conversely, for recurring or similar tasks, LoraHub emerges as a compelling option. Despite the initial fixed cost associated with the Adapt stage, LoraHub's ability to efficiently handle repetitive tasks, often occurring thousands of times, while simultaneously reducing overall expenses, establishes it as a practical and viable option in such scenarios. In summary, the paper's intention is not to replace in-context learning with LoraHub, but to present LoraHub as a complementary strategy with performance-efficiency trade-offs.", "category": "Method_Disambiguation"}, {"question": "Does Table 1 report results averaged on 5 random seeds for all methods or only for `LoraHub`? Additionally, do the different random seeds impact the choice of query few-shot samples, or only optimization aspects such as initialization and the library of Lora modules? What are the average (avg) and best performance of all methods?", "answer": "Table 1 now reports both the average performance across multiple runs for all methods. The paper conducted additional experiments by randomly sampling three sets of five examples each and applied various few-shot methods to assess their robustness. The results on both the average (avg) and best (best) performance across multiple runs for all methods in Table 1 and Table 4 (Appendix A) show that uncertainties exist in all methods when confronted by changes in few-shot examples. On average, LoraHub demonstrates a reasonable performance-efficiency trade-off, showcasing solid performance with a notable reduction in tokens.\n\n| Task                                      | ICL$_{best}$ | IA3$_{best}$ | LoRA$_{best}$ | FFT$_{best}$ | LoraHub$_{best}$ | ICL$_{avg}$ | IA3$_{avg}$ | LoRA$_{avg}$ | FFT$_{avg}$ | LoraHub$_{best}$ |\n|-------------------------------------------|-------------:|-------------:|-------------:|-------------:|-----------------:|-----------:|-----------:|-----------:|-----------:|-----------------:|\n| Boolean Expressions                       |         62.7 |         58.0 |         60.7 |         65.3 |            60.7 |       59.6 |       56.2 |       56.0 |       62.2 |            55.5 |\n| Causal Judgement                          |         59.8 |         62.1 |         57.5 |         60.9 |            63.2 |       59.4 |       60.2 |       55.6 |       57.5 |            54.3 |\n| Date Understanding                        |         21.3 |         20.7 |         40.7 |         67.3 |            45.3 |       20.4 |       20.0 |       35.8 |       59.3 |            32.9 |\n| Disambiguation                            |         69.3 |          0.0 |         68.7 |         70.7 |            68.0 |       69.1 |          0.0 |       68.0 |       68.2 |            45.2 |\n| Dyck Languages                            |          2.0 |          4.7 |         25.3 |         33.3 |             2.7 |        0.9 |         4.2 |       22.2 |       19.5 |             1.0 |\n| Formal Fallacies                          |         59.3 |         52.0 |         56.7 |         56.0 |            59.3 |       55.3 |       51.5 |       53.6 |       54.0 |            52.8 |\n| Geometric Shapes                          |         20.0 |         15.3 |         28.7 |         39.3 |            18.7 |       19.6 |       14.7 |       24.0 |       31.1 |             7.4 |\n| Hyperbaton                                |         72.7 |         49.3 |         57.3 |         82.0 |            72.7 |       71.8 |       49.3 |       55.3 |       77.3 |            62.8 |\n| ...                                       |             |             |             |             |                 |            |            |            |            |                 |\n| Tracking Shuffled Objects (seven objects) |         6.7 |         6.7 |         12.0 |         10.0 |            15.3 |        6.7 |        6.7 |        10.0 |        9.8 |             7.7 |\n| Tracking Shuffled Objects (three objects) |         31.3 |         30.7 |         32.0 |         36.0 |            31.3 |       31.1 |       30.7 |        30.9 |        32.0 |            29.2 |\n| Web of Lies                               |         54.0 |         54.7 |         55.3 |         54.0 |            57.3 |       53.8 |       54.2 |       52.7 |       48.2 |            50.1 |\n| Word Sorting                              |          0.7 |          1.3 |          5.3 |          6.0 |             1.3 |        0.5 |         1.3 |        4.9 |        4.9 |             1.1 |\n| Average Performance Per Task              |         38.4 |         32.1 |         40.9 |         46.2 |            41.2 |       37.3 |       31.6 |       37.7 |       42.1 |            34.7 |\n| Avg Tokens Per Example                    |       597.8 |       111.6 |        111.6 |        111.6 |          111.6 |     597.8 |      111.6 |      111.6 |      111.6 |          111.6 |\n| Gradient-based Training                   |         No  |        Yes  |        Yes  |        Yes  |            No   |        No  |       Yes  |       Yes  |       Yes  |             No  |\n\n\n", "category": "Experimental_Exposition"}, {"question": "Why was FLAN-T5-Large chosen as the base model instead of a model pre-trained on unsupervised text for fine-tuning or LoRA fine-tuning on upstream tasks from the FLAN collection?", "answer": "Firstly, the choice of FLAN-T5 as the paper’s base model is grounded in the pursuit of language models with exceptional few-shot capabilities, making it a robust baseline for the paper’s study. FLAN-T5 not only excels in both zero-shot and few-shot scenarios but also boasts strong problem-solving capabilities, positioning it as a formidable candidate for the paper’s consideration. Secondly, the choice of FLAN-T5 as the paper’s primary base model is crucial to ensuring a fair comparison with zero-shot learning. By utilizing LoRA modules trained on the FLAN collection, the paper’s intention is to insulate the method’s performance from the influence of additional datasets. Given that the FLAN collection is already seen during FLAN-T5’s training, the true efficacy lies in the composition of LoRA modules rather than the introduction of extra datasets. Lastly, the paper has also provided an analysis of the performance of the method with T5 as the base model in both Section 5 (‘Can LoraHub work well on non-instruction-tuning models’) and Appendix B. These sections explicitly showcase the efficacy of the method when applied to models trained exclusively on unsupervised text, underscoring the generality of the paper’s approach.", "category": "Motivation_Analysis"}, {"question": "What is the rationale for maintaining the same rank for the composed LoRA module, and how does the choice of rank impact the performance of LoraHub learning?", "answer": "The impact of different ranks was investigated in Section 5 and Appendix B. For FLAN-T5, the rank choice has minimal impact, but for T5, it still influences performance. Empirical findings show that a rank value of 16 consistently outperforms rank values of 4 or 64 across different runs, both in average and best performance.\n\n\n| Task                      |$Rank 4 _{avg}$ | $Rank 4_{best}$| $Rank 16_{avg}$ | $Rank 16_{best}$ | $Rank 64_{avg}$ | $Rank 64_{best}$ |\n| :------------------------ | :-----: | :-----: | :-----: | :-----: | :-----: | :-----: |\n| Average Performance Per Task |   16.1  |   24.2  |   20.8  |   30.7  |   14.8  |   21.4  |\n", "category": "Experimental_Exposition"}, {"question": "What is the rationale behind the selection of 20 LoRA module candidates for unseen tasks, and how was the selection process improved to address concerns about arbitrariness?", "answer": "The initial selection of 20 LoRA module candidates was conducted randomly, introducing a primary source of randomness. To address concerns about arbitrariness, the authors implemented an improved approach called LoraHub_filter, which selects 20 LoRA module candidates with the lowest loss on few-shot examples. This approach aims to minimize variance and improve consistency, resulting in a slight performance improvement, as detailed in Appendix D.\n\n| Task                          | $LoraHub_{avg}$| $LoraHub_{filter}$ |\n|-------------------------------|------------|---------------|\n| Boolean Expressions           | 55.5       | 60.0          |\n| Causal Judgement              | 54.3       | 52.9          |\n| Date Understanding            | 32.9       | 33.3          |\n| Disambiguation                | 45.2       | 62.7          |\n| Dyck Languages                | 1.0        | 0.0           |\n| ...                           | ...        | ...           |\n| Tracking Shuffled Objects (three objects) | 29.0 | 32.7          |\n| Web of Lies                   | 53.0       | 46.0          |\n| Word Sorting                  | 1.1        | 1.3           |\n| Avg Performance Per Task      | 34.7       | 35.4          |\n\nHowever, the method has also introduced additional reasoning overhead, as the losses of all LoRA module candidates need to be calculated.", "category": "Experimental_Exposition"}, {"question": "What steps were taken to address the absence of baselines, such as the average performance for BBH's top 5 upstream tasks, in the evaluation?", "answer": "New experimental results were included in the study, specifically evaluating the performance on BBH tasks of the top 5 generally useful LoRA modules presented in Table 2. The findings show that these top LoRA modules perform similarly to the original FLAN-T5-large in most cases, but their individual performance does not reach the level achieved by LoraHub. These results support the conclusion that module compositions can enhance overall performance.\n\n| Task                                   | WIQA: Last | RACE: Right | WIQA: First | AdversarialQA: BiDAF | WebQuestions: Answer |\n|----------------------------------------|------------|-------------|-------------|----------------------|----------------------|\n| Boolean Expressions                    | 52.7       | 58.0        | 52.7        | 54.7                 | 53.3                 |\n| Causal Judgement                       | 55.2       | 63.2        | 55.2        | 57.5                 | 57.5                 |\n| Date Understanding                     | 17.3       | 19.3        | 17.3        | 16.7                 | 15.3                 |\n| Disambiguation                         | 0.0        | 0.0         | 0.0         | 0.0                  | 0.0                  |\n| Dyck Languages                         | 0.7        | 0.7         | 0.7         | 1.3                  | 1.3                  |\n| Formal Fallacies                       | 51.3       | 51.3        | 51.3        | 51.3                 | 51.3                 |\n| Geometric Shapes                       | 8.0        | 13.3        | 8.0         | 6.7                  | 7.3                  |\n| Hyperbaton                             | 16.7       | 44.0        | 16.7        | 1.3                  | 6.0                  |\n| Logical Deduction (five objects)       | 23.3       | 28.0        | 23.3        | 19.3                 | 20.7                 |\n| Logical Deduction (seven objects)      | 22.0       | 26.0        | 22.0        | 10.7                 | 12.0                 |\n| Logical Deduction (three objects)      | 0.7        | 9.3         | 0.7         | 0.0                  | 0.0                  |\n| Movie Recommendation                    | 63.3       | 62.7        | 63.3        | 56.7                 | 63.3                 |\n| Multistep Arithmetic                   | 0.7        | 0.7         | 0.7         | 0.7                  | 0.7                  |\n| Navigate                               | 47.3       | 50.0        | 47.3        | 47.3                 | 47.3                 |\n| Object Counting                        | 34.7       | 34.0        | 34.7        | 35.3                 | 35.3                 |\n| Penguins in a Table                    | 45.7       | 41.3        | 45.7        | 39.1                 | 43.5                 |\n| Reasoning about Colored Objects        | 40.0       | 37.3        | 40.0        | 31.3                 | 30.7                 |\n| Ruin Names                             | 22.0       | 21.3        | 22.0        | 17.3                 | 22.7                 |\n| Salient Translation Error Detection    | 36.7       | 34.7        | 36.7        | 32.7                 | 37.3                 |\n| Snarks                                 | 52.6       | 55.1        | 52.6        | 47.4                 | 52.6                 |\n| Sports Understanding                    | 56.0       | 58.7        | 56.0        | 55.3                 | 55.3                 |\n| Temporal Sequences                     | 16.7       | 17.3        | 16.7        | 12.7                 | 17.3                 |\n| Tracking Shuffled Objects (five objects) | 12.0      | 12.0        | 12.0        | 10.7                 | 12.0                 |\n| Tracking Shuffled Objects (seven objects) | 6.7       | 6.7         | 6.7         | 6.7                  | 6.7                  |\n| Tracking Shuffled Objects (three objects) | 20.7      | 30.7        | 20.7        | 10.7                 | 25.3                 |\n| Web of Lies                            | 54.7       | 54.0        | 54.7        | 54.0                 | 54.0                 |\n| Word Sorting                           | 1.3        | 1.3         | 1.3         | 1.3                  | 1.3                  |\n| Avg Performance per Task               | 28.1       | 30.8        | 28.1        | 25.1                 | 27.0                 |\n| FLAN-T5-large                          | 1.1        | 3.8         | 1.1         | -1.9                 | 0.0                  |\n", "category": "Experimental_Exposition"}, {"question": "What baseline was used to evaluate the retrieval of trained LoRAs when given a handful of examples from unseen tasks, and how does it compare to the proposed method?", "answer": "The paper included a discussion of the LoRA retrieval method as a baseline in Section 5. It designed a LoRA retrieval mechanism based on the loss derived from few-shot examples, ranking all LoRA module candidates by their loss on the few-shot examples and evaluating the best candidate on the test set of the unseen task. The performance comparison shows that LoRA retrieval achieved an average performance of 31.7, while LoraHub achieved 34.7 (avg) and 41.2 (best), highlighting the effectiveness of LoRA module composition.\n\n| Task                                       | LoRARetrieval | $LoraHub_{avg}$ | $LoraHub_{best}$ |\n|--------------------------------------------|---------------|---------------|----------------|\n| Boolean Expressions                        | 53.3          | 55.5          | 60.7           |\n| Causal Judgement                           | 63.2          | 54.3          | 63.2           |\n| Date Understanding                         | 20.7          | 32.9          | 45.3           |\n| Disambiguation                             | 0             | 45.2          | 68             |\n| Dyck Languages                             | 0.7           | 1.0           | 2.7            |\n| Formal Fallacies                           | 51.3          | 52.8          | 59.3           |\n| Geometric Shapes                           | 9.3           | 7.4           | 18.7           |\n| Hyperbaton                                 | 57.3          | 62.8          | 72.7           |\n| Logical Deduction (five objects)           | 26            | 36.1          | 40             |\n| Logical Deduction (seven objects)          | 36.8          | 36.8          | 46             |\n| Logical Deduction (three objects)          | 10.0          | 45.7          | 52.7           |\n| Movie Recommendation                        | 63.3          | 55.3          | 62             |\n| Multistep Arithmetic Two                   | 1.3           | 0.4           | 1.3            |\n| Navigate                                    | 49.3          | 47.1          | 51.3           |\n| Object Counting                             | 37.3          | 33.7          | 36.7           |\n| Penguins in a Table                         | 41.3          | 35.9          | 47.8           |\n| Reasoning about Colored Objects             | 40            | 40            | 44.7           |\n| ruin_names                                  | 21.3          | 24.4          | 28.7           |\n| Salient Translation Error Detection         | 46            | 36            | 42.7           |\n| Snarks                                      | 50            | 56.9          | 61.5           |\n| Sports Understanding                        | 54.7          | 56.7          | 62.7           |\n| Temporal Sequences                          | 17.3          | 18.2          | 21.3           |\n| Tracking Shuffled Objects (five objects)    | 12.0          | 12.3          | 16.7           |\n| Tracking Shuffled Objects (seven objects)   | 6.7           | 7.7           | 15.3           |\n| Tracking Shuffled Objects (three objects)   | 30.7          | 29.2          | 31.3           |\n| Web of Lies                                 | 55.3          | 50.1          | 57.3           |\n| Word Sorting                                | 0.7           | 1.1           | 1.3            |\n| Average Performance                         | 31.7          | 34.7          | 41.2           |\n\n", "category": "Method_Comparison"}, {"question": "How are negative coefficients for the LoRA weights interpreted in the context of the equation $\\hat{m} = (w_1A_1+w_2A_2)(w_1B_1+w_2B_2)$?", "answer": "When the weight parameter ($w$) in the equation $\\hat{m} = (w_1A_1+w_2A_2)(w_1B_1+w_2B_2)$ assumes a negative value, the operation involves subtracting the corresponding parameters of LoRA at the specified positions, with no additional adjustments made for negative coefficients.", "category": "Concept_Understanding"}, {"question": "What is the rationale for maintaining a rank of 16 for the composed LoRA module, and what were the findings when higher rank matrices were explored?", "answer": "The choice of rank has minimal impact for FLAN-T5 but still influences T5. Empirical findings show that a rank value of 16 consistently outperforms ranks of 4 or 64 in terms of average and best performance across different runs.", "category": "Experimental_Analysis"}, {"question": "What is the procedure for selecting 20 LoRA module candidates for unseen tasks, and how does this selection impact the method’s performance on unseen tasks?", "answer": "The selection of 20 LoRA module candidates is based on identifying those with the lowest loss on few-shot examples, using a straightforward approach called LoraHub_filter. This selection aims to minimize introduced variance and foster consistent performance. Experimental results show that this approach slightly improves the average performance on unseen tasks.", "category": "Experimental_Setup"}, {"question": "What are the effects of not including a weight threshold on the method's performance, and are there specific tasks where this adjustment negatively impacts the results?", "answer": "An ablation study was conducted by removing the weight threshold. While the adjustment had minimal impact on most tasks, three tasks—'Date Understanding,' 'Disambiguation,' and 'Hyperbaton'—showed notable performance declines, with an average decrease of 1.2% compared to using the threshold. This demonstrates the importance of a reasonable threshold to mitigate extreme scenarios.\n\n| Task                                       | $LoraHub_{avg}\\ w/\\ Threshold$ | $LoraHub_{avg}\\ w/o\\ Threshold$ |\n|--------------------------------------------|-----------------------------|------------------------------|\n| Boolean Expressions                        | 55.5                        | 54.0                         |\n| Causal Judgement                           | 54.3                        | 54.8                         |\n| Date Understanding                         | 32.9                        | 17.7                         |\n| Disambiguation                             | 45.2                        | 40.6                         |\n| ...                                        | ...                         | ...                          |\n| Temporal Sequences                         | 18.2                        | 16.7                         |\n| Tracking Shuffled Objects (five objects)   | 12.3                        | 12.3                         |\n| Tracking Shuffled Objects (seven objects)  | 7.7                         | 8.5                          |\n| Tracking Shuffled Objects (three objects)  | 29.2                        | 29.8                         |\n| Web of Lies                                | 50.1                        | 50.3                         |\n| Word Sorting                               | 1.1                         | 1.3                          |\n| Average Performance                        | 34.7                        | 33.5                         |\n\n", "category": "Experimental_Exposition"}, {"question": "How do the optimization methods for $w$ scale with an increasing number of upstream LoRA modules in terms of cost and performance, and are alternative techniques considered to mitigate optimization noise?\n", "answer": "The primary source of variance in the optimization process is associated with the LoRA candidate modules, not the gradient-free optimization algorithms. Once the candidates are determined, random seeds have minimal impact on the final performance. The observed instability arises from balancing the quantity and quality of LoRA module candidates. This hypothesis receives partial support from the paper's analysis in Section 5, where attempts to increase the number of LoRA modules led to a noticeable rise in model performance volatility. \nIn response to this observation, the authors explored a straightforward approach in the selection of LoRA module candidates. Specifically, $20$ LoRA module candidates with the lowest loss on the few-shot examples were first identified, with the goal of minimizing introduced variance and fostering a more consistent performance. This strategic selection process, named as $ LoraHub_{filter}$, has notably contributed to a slight enhancement in average performance compared to $LoraHub_{avg}$. A brief summary of the results is presented below, and the detailed results can be found in Appendix D. However, the method has also introduced additional reasoning overhead, as the losses of all LoRA module candidates need to be calculated. The variance is further discussed in Appendix G\n\n| Task                                  | $LoraHub_{avg}$ | $LoraHub_{filter}$ |\n|---------------------------------------|-----------------|--------------------|\n| Boolean Expressions                   | 55.5            | 60.0               |\n| Causal Judgement                      | 54.3            | 52.9               |\n| Date Understanding                    | 32.9            | 33.3               |\n| Disambiguation                        | 45.2            | 62.7               |\n| Dyck Languages                        | 1.0             | 0.0                |\n| ...                                   | ...             | ...                |\n| Tracking Shuffled Objects (three objects) | 29.0            | 32.7               |\n| Web of Lies                           | 53.0            | 46.0               |\n| Word Sorting                          | 1.1             | 1.3                |\n| Avg Performance Per Task              | 34.7            | 35.4               |\n", "category": "Experimental_Analysis"}, {"question": "Do the top 5 generally useful LoRA modules in the study perform at a level that surpasses the original FLAN-T5-large individually?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "OqlmgmS4Wr", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "AgentTuning: Enabling Generalized Agent Abilities for LLMs", "authors": ["Aohan Zeng", "Mingdao Liu", "Rui Lu", "Bowen Wang", "Xiao Liu", "Yuxiao Dong", "Jie Tang"], "abstract": "Open large language models (LLMs) with great performance in various tasks have significantly advanced the development of LLMs. However, they are far inferior to commercial models such as ChatGPT and GPT-4 when acting as agents to tackle complex tasks in the real world. These agent tasks employ LLMs as the central controller responsible for planning, memorization, and tool utilization, necessitating both fine-grained prompting methods and robust LLMs to achieve satisfactory performance. Though many prompting methods have been proposed to complete particular agent tasks, there is lack of research focusing on improving the agent capabilities of LLMs themselves without compromising their general abilities. In this work, we present AgentTuning, a simple and general method to enhance the agent abilities of LLMs while maintaining their general LLM capabilities. We construct AgentDataset, a lightweight instruction-tuning dataset containing high-quality interaction trajectories. We employ a hybrid instruction-tuning strategy by combining AgentDataset with open-source instructions from general domains. AgentTuning is used to instruction-tune the Llama 2 series, resulting in AgentLlama. Our evaluations show that AgentTuning enables LLMs' agent capabilities without compromising general abilities. The AgentLlama-70B is comparable to GPT-3.5-turbo on unseen agent tasks, demonstrating generalized agent capabilities. We open source the AgentDataset and AgentLlama-7B, 13B, and 70B models at https://anonymous.4open.science/r/AgentTuning, serving open and powerful alternatives to commercial LLMs for agent tasks.", "keywords": "Large language models;Autonomous agents;Instruction tuning;Reasoning", "qa_pairs": [{"question": "What are the unique aspects and significant contributions of the paper's methodology, given that agent trajectories are very valuable but costly to collect, they are mainly extracted from public tasks/benchmarks using ReAct with GPT models. The overall process, which includes instruction generation, trajectory collection, and filtering, can be useful for other data collection efforts, but it is relatively straightforward. For example, ReAct/COT prompting is used with no significant change, and reward filtering is also a standard practice in imitation learning.", "answer": "Some unique aspects and significant contributions of the paper's work:\n\n- **Limitations of prior works:** Existing studies on LLMs as agents have thus far largely focused on designing prompts or a framework for completing one particular agent task [1][2], rather than fundamentally enhancing the agent capabilities of the LLMs themselves. In addition, many efforts are dedicated to improving LLMs in specific aspects, involving fine-tuning the LLMs using datasets tailored to specific tasks [1][3]. This overemphasis on specialized capabilities comes at the expense of the LLMs’ general abilities and also compromises their generalizability.\n- **First General Agent Model to Match GPT-3.5-turbo**:  As illustrated in Figure 1(b), there is a substantial gap between open-source and API-based models like GPT-3.5 and GPT-4. The paper's work introduces the first open-source model, AgentLlama-70B, which **matches the capabilities of GPT-3.5-turbo in general (unseen) agent tasks while preserving its general abilities**. This makes a significant contribution **both academically and in the development of downstream applications**.\n- **First attempt to instruct-tune LLMs with agent trajectories across multiple tasks**: **The paper is the first to use agent trajectories from multiple tasks for instruction tuning**, achieving performance comparable to that of GPT-3.5-turbo. Furthermore, the paper's error analysis (Cf Sec 3.3) shows that the lack of alignment is the main issue with the agent capabilities of models, which can be greatly improved with minimal data training. It is believed that this will provide insights into how to enhance the inherent agent capabilities of models.\n- **Automated and Scalable Trajectory Collection Methodology:** Given that agent trajectories are precious and challenging to collect, the paper has established an **automated and scalable method for collecting agent trajectories and has verified its effectiveness**. For the Instruction Generation process, the paper used two methods, Self-Instruct and Task Derivation, to supplement tasks lacking training instructions. This **ensures that the paper has sufficiently diverse agent trajectories**. In the Trajectory Collection process, the paper interacts using GPT-4 and aggressively apply a filtering of $r=1$ to almost all trajectories, **ensuring it has high-quality agent trajectories**. Both of these steps **require no manual intervention and can be easily extended to other agent tasks**. By fine-tuning the model on 1,866 trajectories from 6 held-in tasks, there was a significant improvement in performance on 6 held-out (unseen) Agent tasks, demonstrating the effectiveness of the paper's method.\n\nRegarding the concerns about the innovativeness of using the ReACT/CoT prompt method. While previous works [4][5] have indeed innovated in prompting methods to enhance the performance of API-based models, **they have not focused on improving the agent capabilities of the models themselves**, which is the main focus of the paper. Considering the fact that open-source models are far behind API models in agent capabilities, the paper has for the first time **elevated the general-purpose agent capability of an open-source model to the level of GPT-3.5-turbo**, even just by using the common ReACT prompt method. This still to be a significant contribution.\n\n[1] Deng, Xiang, et al. \"Mind2Web: Towards a Generalist Agent for the Web.\" arXiv preprint arXiv:2306.06070 (2023).\n\n[2] Yao, Shunyu, et al. \"Webshop: Towards scalable real-world web interaction with grounded language agents.\" Advances in Neural Information Processing Systems 35 (2022): 20744-20757.\n\n[3] Qin, Yujia, et al. \"Toolllm: Facilitating large language models to master 16000+ real-world apis.\" arXiv preprint arXiv:2307.16789 (2023).\n\n[4] Xu, Binfeng, et al. \"ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models.\" arXiv preprint arXiv:2305.18323 (2023).\n\n[5] Shinn, Noah, et al. \"Reflexion: Language Agents with Verbal Reinforcement Learning, June 2023.\" arXiv preprint arXiv:2303.11366.", "category": "Method_Disambiguation"}, {"question": "How is the reward $r$ calculated for filtering low-quality trajectories, and how is the correctness of the trajectories generated by GPT-3.5/4 ensured?", "answer": "**Reward Calculation** As the paper adopts 6 held-in agent tasks from AgentBench [1], the detailed reward calculation formula for each task can be found in the appendix of AgentBench. Reward represents the completion status of a task, with values ranging from 0 to 1, where 1 indicates full completion. The 6 held-in tasks are summarized as a table below:\n\n|   Task   |        Description        |                     Example                      |       Reward       |                      Reward Calculation                      |\n| :------: | :-----------------------: | :----------------------------------------------: | :----------------: | :----------------------------------------------------------: |\n| ALFWorld | Daily Household  Routines |                    Heat food                     |    Success Rate    |           If task is finished, r=1, otherwise r=0            |\n| WebShop  |      Online Shopping      |                   Buy a shirt                    |       Reward       |     Score for selecting the correct item during shopping     |\n| Mind2Web |    Website Navigation     |                  Book a ticket                   | Step Success  Rate | Evaluate the predicted action correctness compared to reference actions. |\n|    KG    | Retrieve Entity from  KG  | Which team won the 2014 AFC  Championship Game?  |         F1         | Compare the model’s predicted answers to the gold standard answers |\n|    DB    |   Database  Operations    | How many games did the badgers  play in october? |    Step Success    |           If MySQL query is correct, r=1, otherwise r=0            |\n|    OS    |    Interacting with OS    |               Count specific files               |    Step Success    |           If result from operating system is correct, r=1, otherwise r=0            |\n\n**Human Evaluation and Dataset Verification**: Since the paper uses trajectory filtering with $r=1$ for most tasks (except for MindWeb due to its difficulty), the correctness of the trajectories in these tasks can be ensured. Combining the Self-Instruct and Task Derivation methods for Instruction Generation, the paper's data construction process **does not require manual participation, which makes the paper's method scalable and easily extendable to other agent tasks**. Table 2 in the paper demonstrates that using trajectory filtering to eliminate incorrect trajectories significantly enhances the model's capabilities in Agent tasks. As for using GPT-4 to generate the Chain of Thought (CoT) reasoning process, its effectiveness has been verified by many works [2][3], and comparative experiments show that training with the CoT process yields better performance for the model.\n\n|  | AgentLlama-7B (w/ CoT) | AgentLlama-7B (w/o CoT) |\n| :---: | :---: | :---: |\n| Held-in | **1.96** | 1.38 |\n| Held-out | **0.67** | 0.56 |\n| General | **0.63** | 0.57 |\n\n[1] Liu, Xiao, et al. \"Agentbench: Evaluating llms as agents.\" arXiv preprint arXiv:2308.03688 (2023).\n\n[2] Yue, Xiang, et al. \"Mammoth: Building math generalist models through hybrid instruction tuning.\" arXiv preprint arXiv:2309.05653 (2023).\n\n[3] Mukherjee, Subhabrata, et al. \"Orca: Progressive learning from complex explanation traces of gpt-4.\" arXiv preprint arXiv:2306.02707 (2023).", "category": "Method_Mechanics"}, {"question": "Why were typical continual learning strategies like weight regularization[1] and parameter allocation[2] not directly applied in the fine-tuning of large language models within the AgentTuning methodology, and what alternative approaches are proposed?\n\n[1] Kirkpatrick et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences 2017.\n\n[2] Serra et al. Overcoming catastrophic forgetting with hard attention to the task. ICML 2018.", "answer": "Discussing and comparing the paper's approach with established continual learning methods like weight regularization and parameter allocation would enrich the paper. However, **there are specific challenges in directly applying many continual learning techniques to the fine-tuning of large language models (LLMs)**.\n\nFor example, Elastic Weight Consolidation (EWC) necessitates access to the original supervised fine-tuning (SFT) data from the first task to estimate Fisher information for each parameter. Similarly, Hard Attention to the Task (HAT) requires the original SFT data to learn a task embedding with the original SFT data and modifications to the model architecture, such as adding gates between layers. However, the fine-tuning data for many open-source LLMs, like the Llama series, is not typically available. (Still, it's not totally impossible to combine these methods. For EWC, Fisher information for each parameter could potentially be estimated using additional general task data, such as ShareGPT. Similarly, for HAT, the task embedding may also be estimated using extra data.)\n\nAnother challenge of implementing EWC/HAT entails the techniques used to train language models with huge parameters, like model/pipeline parallelism, distributed optimizers, etc. It requires huge modifications to Megatron training infrastructures (for example, reducing over the square of gradients, and accessing all parameters for EWC loss terms calculation) and would likely necessitate considerable effort in tuning method-specific hyper-parameters.\n\nInstead of integrating EWC and HAT methods, **the authors explored LoRA and P-Tuning, both of which are prevalent and parameter-efficient fine-tuning methods in the realm of LLMs and are frequently applied in the continual learning of LLMs**. LoRA learns a low rank delta for the parameter matrices, while P-Tuning introduces an additional prefix in each layer.", "category": "Motivation_Analysis"}, {"question": "What is the rationale for setting the sampling ratio between GPT-3.5 and GPT-4 as 1:4, and why is the reward threshold $r$ set to 2/3 in the Mind2Web task?", "answer": "- **About sampling ratio between GPT-4 and GPT-3.5.**\nThe sampling ratio between GPT-4 and GPT-3.5 was determined through a grid search on a 7B scale, varying from 0.0 to 0.5 with increments of 0.1. The ratio of 0.2 (GPT4 : GPT3.5 = 1 : 4) was chosen because it yielded the maximum average score across general tasks, balancing quality and diversity.\n\n| GPT-4 Sampling Ratio | 0.5  | 0.4  | 0.3  | 0.2  | 0.1  | 0.0  |\n|----------------------|------|------|------|------|------|------|\n| MMLU                 | 48.4 | 48.4 | 49.3 | 48.7 | 46.0 | 49.3 |\n| HumanEval            | 14.4 | 14.9 | 11.3 | 15.4 | 6.20 | 5.50 |\n| MT-Bench             | 6.00 | 6.08 | 6.15 | 6.34 | 6.40 | 6.18 |\n| GSM8K                | 21.8 | 24.6 | 23.7 | 24.6 | 23.7 | 26.5 |\n| **AVG**              | 0.60 | 0.62 | 0.59 | **0.63** | 0.55 | 0.57 |\n\n- **About the reward threshold 2/3** \nAs mentioned in Section 2.1.3 of the paper, the reward filtering threshold for all held-in tasks is set as 1 (which means the result of the operation is totally correct), except for Mind2Web, which is set as 2/3. The Mind2Web task is particularly hard, and even GPT-4 can only fully complete a small portion of tasks in the training set. This figure[1] shows the remaining number of trajectories under different filtering thresholds in the Mind2Web task. It can be observed that using a higher reward value for filtering results in a very small proportion of trajectories being retained. The filtering threshold for Mind2Web was chosen to strike a balance between data quality and diversity, whereas 2/3 is a relatively appropriate value.\n\n\n[1]https://i.imgur.com/5G3MR2h.png", "category": "Motivation_Analysis"}, {"question": "Is the comparison in Figure 1 (b) fair, given that the proposed model was trained on AgentBench while the other LLMs were not, potentially making AgentBench in-distribution for the proposed model?", "answer": "AgentBench was not the primary benchmark used to validate the agent capabilities of the paper's model. Figure 1 (b) was intended to illustrate the performance gap between open-source models and API-based models. Additionally, the paper conducted systematic evaluations on 6 held-out agent tasks that were not included in the training set, demonstrating robust generalization performance. These results show that the paper's model exhibits strong agent capabilities on these held-out tasks.", "category": "Method_Disambiguation"}, {"question": "What strategies can be employed to extend the capabilities of AgentTuning beyond reliance on GPT-4 generated data?", "answer": "The use of GPT-4 generated data in the paper's experiments was a strategic choice, primarily for its convenience of data generation and the quality of the trajectories. AgentTuning can serve as a foundational method that can be readily combined with other data-centric approaches to foster improvements that are not dependent on the external models like GPT-4. For instance, a key step in AgentTuning is trajectory filtering, which rejects trajectories with poor performance with environment reward. Self-instruct may be combined with trajectory filtering to perform AgentTuning, where a model can be iteratively self-improving, like [1], while preserving general abilities through hybrid instruction tuning.\n\n[1] Micheli, Vincent, and François Fleuret. \"Language models are few-shot butlers.\" arXiv preprint arXiv:2104.07972 (2021).", "category": "Method_Mechanics"}, {"question": "How does the proposed model demonstrate generalization ability to agent tasks not captured in AgentBench, such as driving or operating Minecraft?", "answer": "Generalization ability is one of the paper's main goals. To demonstrate this, the paper carried out extensive experiments showing that the paper's models exhibit strong generalization ability. As described in the paper's evaluation setup (Cf Sec. 3.1), it conducted experiments on 6 held-out tasks that are not covered in the paper's training set, with 5 of them not from AgentBench. The tasks vary greatly in their respective domains. Specifically, the 6 tasks cover tasks of digital card games, daily computer tasks, science experiments, web interaction, wiki retrieval, and question answering. As shown in Section 3.2 Main Results, the paper's model achieves robust performance on these diverse tasks, further validating its strong generalization capabilities.\nRegarding the Mind2Web task, the filtering threshold was set to 2/3. A lower value for filtering results in a very small proportion of trajectories being retained. The filtering threshold for Mind2Web was chosen to strike a balance between data quality and diversity, whereas 2/3 is a relatively appropriate value.\nIn comparison to prior work on Minecraft [1], even GPT-3.5 performs poorly according to [1]. They found that GPT-3.5 and open-source LLMs struggle to perform well in playing Minecraft. In their experiments, substituting GPT-4 with GPT-3.5 in code generation leads to a sharp performance degradation, with over 80% drop in the number of unique items obtained. Besides, [1] utilizes a set of complicated prompts to enable GPT-4 to play the game effectively. However, the paper's work focuses on improving the inherent agent ability of models, which will naturally benefit from advancements in prompting methods. Therefore, in the current stage, the paper did not test the models on these highly difficult tasks.\n\n[1]Wang, Guanzhi, et al. \"Voyager: An open-ended embodied agent with large language models.\" arXiv preprint arXiv:2305.16291 (2023).", "category": "Experimental_Analysis"}, {"question": "How does the overall score of the proposed model compare to GPT-4, from which the training data was collected, as shown in Table 4?", "answer": "The main results in the paper show GPT-4's overall score in both held-in and held-out tasks (refer to Table 4). The proposed model uses Llama 2 as its base model, which is comparable to GPT-3.5-turbo in general domains but has a significant gap in agent tasks. While surpassing GPT-4 in specific agent tasks is possible with open-sourced LLMs, the focus of this work is on general agent capabilities, making GPT-3.5-turbo a reasonable baseline for comparison.", "category": "Method_Comparison"}, {"question": "What are the reasons that the AlfWorld results in Table 4 cannot be directly compared to the results reported in [1], given that [1] reported a higher score than the best open-source model?\n\n[1] Micheli, Vincent, and François Fleuret. \"Language models are few-shot butlers.\" arXiv preprint arXiv:2104.07972 (2021).", "answer": "The ALFWorld evaluation setting in [1] differs from the evaluation setting used in AgentTuning (which follows the setting in [2]), making the scores not directly comparable. Additionally, the model in [1] was task-specific fine-tuned from a pre-trained model and is limited to solving ALFWorld tasks, lacking the capability to follow general instructions or perform generalized agent tasks, which makes it unsuitable for direct comparison.", "category": "Experimental_Analysis"}, {"question": "What is the justification for filtering trajectories based on a threshold, especially considering that it drastically reduces the dataset size? Additionally, what results demonstrate the effectiveness of using as much data as possible without injecting COT annotations?", "answer": "The justification for filtering trajectories is that more diverse, higher-quality data is more effective than a large amount of low-quality data, as supported by the conclusions in [1]. This approach enables the AgentTuning method to generalize to unseen agent tasks while maintaining general capabilities. The ablation results in Table 2 (Cf Sec. 2.1.3) demonstrate that filtered trajectories yield significantly better performance than unfiltered trajectories, confirming that a small amount of high-quality data is superior to training with all available data. Additionally, all trajectories interacted by GPT-4 include CoT rationale, except those constructed using Task Derivation, where GPT-4 is used to supplement the reasoning process. Comparative experiments show that training with the CoT process improves model performance.\n\n|  | AgentLlama-7B (w/o Filter) | AgentLlama-7B (w/o CoT) | AgentLlama-7B |\n| :---: | :---: | :---: | :---: |\n| Held-in | 1.34 | 1.38 | **1.96** |\n| Held-out | 0.47 | 0.56 | **0.65** |\n\n[1]Zhou, Chunting, et al. \"Lima: Less is more for alignment.\" arXiv preprint arXiv:2305.11206 (2023).", "category": "Experimental_Analysis"}, {"question": "How was the hyper-parameter η chosen, and what is its impact on the tradeoff between broader generalization and agent-task performance? Are the models sensitive to this parameter?", "answer": "The hyper-parameter η, which balances general language ability and agent task ability, was chosen through grid searches on 7B and 13B scales. The evaluation metrics included scores on held-in agent tasks, general tasks, and their average. Thus, no information about held-out tasks is leaking through this hyper-parameter. The scores of held-in tasks and general tasks are not directly comparable. To compose the average score, all scores are normalized with the maximum score in each category (i.e. the score of GPT-4). \n\n- Grid Search in 7B scale\n\n    |$η$|0.0|0.1|0.2|0.3|0.4|0.5|0.6|0.7|0.8|0.9|1.0|\n    |---|---|---|---|---|---|---|---|---|---|---|---|\n    |7B Held-in| 0.13 | 0.72 | 0.68 | 0.72 | 0.73 | 0.74 | 0.72 | 0.75 | 0.74 | 0.69 | 0.50 |\n    |7B General| 0.42 | 0.42 | 0.41 | 0.41 | 0.40 | 0.40 | 0.40 | 0.41 | 0.37 | 0.36 | 0.14 |\n    |7B Average| 0.27 | **0.57** | 0.55 | **0.57** | 0.56 | **0.57** | 0.56 | **0.58** | 0.55 | 0.53 | 0.32 |\n\n- Grid Search in 13B scale\n\n    |$η$|0.0|0.1|0.2|0.3|0.4|0.5|0.6|0.7|0.8|0.9|1.0|\n    |---|---|---|---|---|---|---|---|---|---|---|---|\n    |13B Held-in| 0.15 | 0.72 | 0.77 | 0.74 | 0.61 | 0.66 | 0.70 | 0.60 | 0.67 | 0.72 | 0.57 |\n    |13B General| 0.45 | 0.45 | 0.47 | 0.45 | 0.47 | 0.47 | 0.46 | 0.47 | 0.42 | 0.35 | 0.12 |\n    |13B Average| 0.30 | **0.59** | **0.62** | **0.59** | 0.54 | 0.56 | 0.58 | 0.54 | 0.54 | 0.54 | 0.35 |\n\nThe models are not very sensitive to η, except for extreme values (η=0 and η=1). The average score is evenly distributed in 7B models and peaks at η=0.2 in 13B models. The final η value of 0.2 was selected for all models based on these results. ", "category": "Method_Mechanics"}]}
{"id": "Fq8tKtjACC", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Textbooks Are All You Need", "authors": ["Suriya Gunasekar", "Yi Zhang", "Jyoti Aneja", "Caio Cesar Teodoro Mendes", "Allie Del Giorno", "Sivakanth Gopi", "Mojan Javaheripi", "Piero Conti Kauffmann", "Gustavo Henrique de Rosa", "Olli Saarikivi", "Adil Salim", "Shital Shah", "Harkirat Behl", "Xin Wang", "Sebastien Bubeck", "Ronen Eldan", "Adam Tauman Kalai", "Yin Tat Lee", "Yuanzhi Li"], "abstract": "We introduce phi-1, a new large language model for code, with significantly smaller size than competing models: phi-1 is a Transformer-based model with 1.3B parameters, trained for 4 days on 8 A100s, using a selection of ``textbook quality\" data from the web (6B tokens) and synthetically generated textbooks and exercises with GPT-3.5 (1B tokens). Despite this small scale, phi-1 attains pass@1 accuracy 50.6% on HumanEval and 55.5% on MBPP. It also displays surprising emergent properties compared to phi-1-base, our model before our finetuning stage on a dataset of coding exercises, and phi-1-small, a smaller model with 350M parameters trained with the same pipeline as phi-1 that still achieves 45% on HumanEval.", "keywords": "large language models;coding models", "qa_pairs": [{"question": "Is the structure and compositionality of programming code essential for high performance in small models and datasets, and can this approach be applied to other structured tasks like recipes or unstructured tasks?", "answer": "Yes, the composability and structure of code generation was certainly helpful for the paper in effectively selecting high quality data through filtering and generation. For example, it is intuitively easier to define specifications for filtering existing web-datasets for good quality algorithmic code compared to more imprecise and subjective skills such as common-sense reasoning. Textbooks can be built with sufficient coverage based on a beginner level syllabus for programming, ensuring coverage. This methodology is not specific to Python and can extend to other languages and structured tasks. For unstructured tasks like common-sense reasoning, while filtering and generating 'textbook' quality data is more challenging, it is not impossible and can benefit from creative approaches to enhance reasoning skills in the training corpus.", "category": "Method_Mechanics"}, {"question": "How can the emergent abilities of Phi-1, particularly in API usage, be evaluated more systematically and qualitatively beyond the examples provided in Section 3?", "answer": "The emergent ability of Phi-1 in API usage is demonstrated by its improved performance on popular APIs despite their absence in the fine-tuning dataset. Due to the lack of standard benchmarks for evaluating API usage, the paper relies on providing more qualitative examples of correct API usage.", "category": "Experimental_Analysis"}, {"question": "What are the unique contributions of this paper compared to prior work on tailored training data and prompted training in LLMs?", "answer": "While the philosophy of good quality training data being good for learning is not new, the predominant trend in LLMs has been on scaling-up data and model sizes, as evidenced by numerous papers on scaling laws. The paper is among the first to provide strong and compelling evidence on the effectiveness of high-quality data to achieve state-of-the-art performance in LLMs. The paper contributes uniquely in several ways: (1) Targeted synthetic data: The paper demonstrates the effective usage of targeted synthetic data to address specific gaps in existing data sources, such as synthetic textbooks that enhance signals from natural language text explaining simple algorithms in detail. (2) Filtering methodology: It employs a combination of GPT-4, a small LM classifier, and human-in-the-loop for textbook-quality data filtering, contributing to early-stage experimentation with frontier LLMs. Identifying best practices in this area is significant for future models. (3) Model performance: The model achieves top-tier performance on algorithmic code generation(next only to latest GPT models) despite being 10x-100x smaller than other top models. This provides strong evidence that selecting high quality data in creative ways is key to unlocking efficient learning.", "category": "Method_Disambiguation"}, {"question": "Why was the evaluation of the model limited to short Python functions, and how does it perform on more complex, real-world coding tasks?", "answer": "The evaluation was intentionally limited to Python programming on algorithmic tasks to focus on high-quality data for learning reasoning and planning-based skills, which are arguably among the hardest to learn from data scaling alone. Despite this narrow focus, qualitative examples in the paper demonstrate the model's ability to perform well on broader tasks, such as using popular APIs and understanding general natural language formats. The limitations of the model are discussed in the appendix, and qualitative evaluations are used due to the lack of standard benchmarks for broad use-cases in code generation.", "category": "Motivation_Analysis"}, {"question": "What methods were employed to evaluate the quality and diversity of the synthetic textbook data, and how were these measures validated?", "answer": "The quality and diversity of the synthetic textbook data were primarily assessed through manual evaluation. Quantitative validation was achieved by evaluating the performance of a small model trained on representative examples of the generated data. The evaluation of quality and diversity is subjective and depends on the specific task at hand; as such, there are no good metrics to gauge utility for LLMs. The paper considered these difficulties and the limitations of manual evaluation when crafting prompts for the generation process. The prompts were tailored to the generating-model's (GPT3.5) known strengths and weaknesses. Taking the example of synthetic textbooks, these were designed to cover the entirety of beginner programming topics, but at the same time, the topics selected were fine-grained enough that GPT-3.5 could reliably generate high accuracy text. The paper ensured diversity through variations in format and target audience.", "category": "Experimental_Setup"}, {"question": "What are the limitations of the proposed model, and in what types of tasks or data is it most likely to fail?", "answer": "The model is most likely to fail in tasks that require memorization, such as those involving less-known libraries in Python, as it is not trained on all available Python code. Additionally, it struggles with tasks requiring subjective data filtering or generation, such as common sense reasoning. These limitations are discussed in the appendix and conclusions, and potential remedies include continual training to incrementally add knowledge.", "category": "Method_Disambiguation"}, {"question": "Why was a thorough decontamination analysis not applied to the synthetic textbook dataset, unlike the CodeExercises dataset?", "answer": "The thorough decontamination analysis for HumanEval on the CodeExercises dataset was possible because both datasets consisted of the same type (short python functions) and contained self-contained code snippets. This allowed for the development of robust similarity detection methods beyond simple string matching. Many decontamination procedures in the literature (including those for The Stack dataset) rely on string-matching or min-hash matching against a test set, which the analysis shows is not very robust (for example, the analysis identified similarities that N-gram analysis did not). On the other hand, the synthetic textbook is a fundamentally different NLP-heavy dataset. The code snippets within the text are often very short and primarily used for illustrative purposes, such as exercises, usage-specific functions or structures like print and if-then statements, or class definitions. These are not directly comparable to HumanEval questions, which require specific function-level algorithmic solutions. Therefore, this dataset was not deemed suitable for in-depth AST-style analysis.", "category": "Experimental_Analysis"}, {"question": "How to justify the use of both high-quality textbook data and GPT-3.5-generated data, which may contain noise, given the apparent contradiction between the two?", "answer": "The authors justify the use of GPT-3.5-generated data by comparing it to random code/text snippets from the web, which, even after filtering, contain higher levels of non-educational content. They argue that, with appropriate prompting, GPT-3.5-generated data has more educational value for simple coding problems compared to typical random text from GitHub or StackOverflow. The harder issue to tackle with using Synthetic data is that the generated text might not have sufficient coverage of rarer and more complex use-cases. This is why the paper uses a mix of web-filtered data and synthetic data, the latter targeted to address specific shortcomings observed in the webdata. The paper's prompts were tailored to the generating model (GPT3.5)’s known strengths and weaknesses. Taking the example of synthetic textbooks, these were designed to cover the entirety of beginner programming topics, but at the same time, the topics selected were fine-grained enough that GPT-3.5 can reliably generate high accuracy text. The paper ensured diversity through variations in format and target audience. Of course, if the paper uses a more powerful GPT-4 model for the paper's generations the quality of models will very likely improve further, but generating data at the paper's scale using GPT-4 is significantly more expensive and slower (due to access limits).", "category": "Experimental_Analysis"}, {"question": "Why were other high-quality data sources, such as legal documents and government documents, not included in the training set alongside textbook data?", "answer": "Other high-quality data sources like legal and government documents were not included due to two main issues: (a) much of this material, including real textbooks, is copyrighted, and (b) the size of the datasets obtained from such sources was very small (<<100MB), making negligible impact on the training set.", "category": "Experimental_Analysis"}, {"question": "What are the training data and methods used by CodeGen-Mono, Replit, and StarCoder, and how do they differ from Phi-1 in terms of data quality and training approach?", "answer": "CodeGen-Mono uses a combination of Python source code from GitHub and NLP data from The Pile, both of which are large, unfiltered web datasets. StarCoder and Replit are trained on a deduplicated subset of The Stack, which includes GitHub source code cleaned of PII and permissive licenses, and they use multiple code languages in pretraining. These models were trained on a total of 1T and 500B tokens, respectively. In contrast, Phi-1 (a) uses only the Python subset of The Stack and StackOverflow, (b) filters the data to reduce it to 20-25% of the original size (6B unique tokens), and (c) strategically adds targeted synthetic data during pretraining and fine-tuning. Phi-1 is trained on 7B unique tokens for slightly less than 8 epochs. The Python-only choice was made to restrict the scope for controlled experiments, while the filtering and synthetic data additions are new contributions in this work.", "category": "Method_Comparison"}, {"question": "What is the performance of GPT-4 on the 50 new unconventional coding problems, and how does it compare to other models?", "answer": "The performance of GPT-4 on the 50 new unconventional coding problems is 73% on the GPT-Score and 67% on HumanEval, as reported in the table. This performance is higher than other models\n\n| Model | GPT-Score | HumanEval |   \n|---|---|---|   \n| CodeGen-Mono-16.1B Nijkamp et al. (2023b) | 38% | 29% |   \n| Replit Replit (2023) | 37% | 22% |   \n| StarCoder Li et al. (2023) | 51% | 34% |   \n| phi-1-base | 37% | 29% |   \n| phi-1 | 52% | 51% | \n| GPT-4 | 73% | 67% (reported) |", "category": "Method_Comparison"}, {"question": "Why was a decontamination analysis not conducted for the 'synthetic textbook' dataset, similar to the analysis performed for the 'synthetic exercise' dataset?", "answer": "A thorough decontamination analysis for HumanEval on the CodeExercises dataset was possible because both datasets were of the same type (short python functions) and are self-contained code snippets. This meant that robust versions of similarities beyond string matching could be devised. Many decontamination procedures in the literature (including The Stack dataset) resort to string-matching or min-hash matching against a test set which the paper’s analysis shows is not very robust (e.g., the paper’s analysis picked up similarities that N-gram analysis did not). In contrast, the ‘synthetic textbook’ dataset is NLP-heavy, with code snippets that are often very short and used for illustrative purposes (e.g., exercises or usage-specific functions or structures like print and if-then, class definition, etc.). These snippets are not directly comparable to HumanEval questions, which involve specific function-level algorithmic tasks. Therefore, the ‘synthetic textbook’ dataset was not suitable for in-depth AST-style analysis.", "category": "Method_Disambiguation"}, {"question": "Why are finetuning results for other models like StarCoder not included, and how does this impact the assessment of whether 'textbook' quality data alone is sufficient for model performance?", "answer": "The benefit of the ‘textbook’ approach not only lies in the final performance but also in overall training compute. Finetuning results for other models like StarCoder are not included because these models are pretrained with significantly larger compute resources. StarCoder, for instance, is 10x larger in terms of parameters and computationally expensive to finetune. While CodeExercise finetuning might boost performance, it does not diminish the importance of high-quality ‘textbook’ data for pretraining. The paper's phi-1-base model, trained on the CodeTextbook dataset without finetuning, achieves 29% HumanEval performance with only 1.3B parameters, outperforming the previous smallest model, Replit-Finetuned, which achieves close to 30% performance on HumanEval at 2.7B parameters and was trained with 100 times more training tokens than Replit (2023).", "category": "Experimental_Analysis"}, {"question": "Why was an in-depth contamination study not conducted for the synthetic textbook dataset, and how does it differ from the HumanEval dataset in terms of suitability for such analysis?", "answer": "A thorough decontamination analysis for HumanEval on the CodeExercises dataset was possible because both datasets were of the same type (short python functions) and are self-contained code snippets. This meant that robust versions of similarities beyond string matching could be devised. Many decontamination procedures in the literature (including The Stack dataset) resort to string-matching or min-hash matching against a test set which the paper’s analysis shows is not very robust (e.g., the paper’s analysis picked up similarities that N-gram analysis did not). In contrast, the ‘synthetic textbook’ dataset is NLP-heavy, with code snippets that are often very short and used for illustrative purposes (e.g., exercises or usage-specific functions or structures like print and if-then, class definition, etc.). These snippets are not directly comparable to HumanEval questions, which involve specific function-level algorithmic tasks. Therefore, the ‘synthetic textbook’ dataset was not suitable for in-depth AST-style analysis.", "category": "Experimental_Analysis"}, {"question": "What is the extent of overlap between the HumanEval dataset and the 'synthetic textbooks' dataset, and why was a thorough decontamination analysis not performed on the latter?", "answer": "A thorough decontamination analysis was conducted for HumanEval on the CodeExercises dataset due to their similar nature as self-contained short Python functions, allowing for robust similarity assessments beyond string matching. However, the synthetic textbook dataset is fundamentally different, being NLP-heavy with very short code snippets used for illustrative purposes, such as exercises or specific functions like print and if-then statements. These snippets are not directly comparable to the specific function-level algorithmic style of HumanEval questions, making in-depth ast-style analysis unsuitable.", "category": "Experimental_Analysis"}, {"question": "Was the diversity of synthetic textbook data ensured by varying format and target audience?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the reliance on GPT-4 as an annotator in Section 2.1 justified by comparing GPT-4's high-quality and low-quality data annotations?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was a thorough decontamination analysis performed on the 'synthetic textbook' dataset similar to the one conducted for the 'synthetic exercise' dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was a decontamination analysis performed to rule out overlap between HumanEval and the 'synthetic textbooks' dataset?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the quality of the synthetic textbook data evaluated solely through automated metrics?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the evaluation of synthetic textbook data include manual assessment of quality and diversity?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper perform manual quality checks and prompt refinements to validate GPT-4's annotation performance?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the prompts for generating synthetic textbooks tailored to address GPT-3.5's strengths and weaknesses?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "DY6uhcv4Xm", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "FedSecurity: A Benchmark for Attacks and Defenses in Federated Learning and Federated LLMs", "authors": ["Shanshan Han", "Baturalp Buyukates", "Zijian Hu", "Han Jin", "Weizhao Jin", "Lichao Sun", "Xiaoyang Wang", "Wenxuan Wu", "Chulin Xie", "Yuhang Yao", "Kai Zhang", "Qifan Zhang", "Yuhui Zhang", "Salman Avestimehr", "Chaoyang He"], "abstract": "This paper introduces FedSecurity, an end-to-end benchmark designed to simulate adversarial attacks and corresponding defense mechanisms in Federated Learning (FL). FedSecurity comprises two major components: FedAttacker, which simulates attacks injected during FL training, and FedDefender, which simulates defensive mechanisms to mitigate the impacts of the attacks. FedSecurity is open-\nsource and can be customized to cover a wide range of machine learning models (e.g., Logistic Regression, ResNet, and GAN) and federated optimizers (e.g., FedAVG, FedOPT, and FedNOVA). We also demonstrate the use of FedSecurity during federated training of Large Language Models (LLMs), showcasing its adaptability and applicability in more complex scenarios.", "keywords": "Security;attack;defense;byzantine;backdoor;federated learning;LLM", "qa_pairs": [{"question": " Why does the library exclude certain privacy attacks, such as adversarial examples, property inference attacks, and membership inference attacks, that can occur during the training and deployment of federated learning models?", "answer": "The paper has cited the following privacy attacks [1,2,3,4,5]The library excludes certain privacy attacks because they often require strong assumptions that are difficult to satisfy in real-world applications. For example, in DLG attack [6], to revert an image training data that can be recognized by the naked eye, the model should be trained using only one sample, and the model inversion attack [7] requires knowing the distribution of the training data in advance. So, the paper focuses more on the attacks that are more practical in real-world scenarios.\n\n[1] Fowl, Liam, et al. \"Robbing the fed: Directly obtaining private data in federated learning with modified models.\" arXiv preprint arXiv:2110.13057 (2021).\n\n[2] Wang, Zhibo, et al. \"Poisoning-assisted property inference attack against federated learning.\" IEEE Transactions on Dependable and Secure Computing (2022).\n\n[3] Luo, Xinjian, et al. \"Feature inference attack on model predictions in vertical federated learning.\" 2021 IEEE 37th International Conference on Data Engineering (ICDE). IEEE, 2021.\n\n[4] Zhang, Jingwen, et al. \"Gan enhanced membership inference: A passive local attack in federated learning.\" ICC 2020-2020 IEEE International Conference on Communications (ICC). IEEE, 2020.\n\n[5] Melis, Luca et al. “Exploiting Unintended Feature Leakage in Collaborative Learning.” 2019 IEEE Symposium on Security and Privacy (SP) (2018): 691-706.\n\n[6] Zhu, Ligeng, Zhijian Liu, and Song Han. \"Deep leakage from gradients.\" Advances in neural information processing systems 32 (2019).\n\n[7] Geiping, Jonas, et al. \"Inverting gradients-how easy is it to break privacy in federated learning?.\" Advances in Neural Information Processing Systems 33 (2020): 16937-16947.", "category": "Experimental_Analysis"}, {"question": "What analytical and visual evidence does the paper provide to demonstrate the impact of attacks and defenses on model convergence in Federated Learning?", "answer": "The paper includes 10 different experiments that analyze how common attacks and defenses in Federated Learning affect model accuracy and loss. These experiments cover various computer vision and NLP tasks, including experiments with large language models, and a real-world experiment using 70 edge devices from Theta Network. Performance under different configurations (number of clients, malicious clients, data distributions, and models) is visualized. Key takeaways include: i) the Byzantine attack of random mode effectively decreases global model test accuracy, while m-Krum provides robustness against various attacks; ii) defense mechanisms mitigate attacks but may negatively impact aggregation results and model performance. However, attacks are infrequent in real FL systems. Therefore, it’s crucial to weigh the benefits against potential drawbacks before integrating a defense mechanism into real systems.", "category": "Experimental_Analysis"}, {"question": "Why did the paper choose to scale the number of clients to 100 instead of the commonly suggested 1000 clients for benchmarking in federated learning (FL) experiments?", "answer": "The paper chose to scale the number of clients to 100 based on the feasibility and relevance to real-world scenarios, as real-world FL applications typically use fewer than 20 clients and none utilize more than 100 clients. Although the theoretical appeal of scaling up to 1000 clients, the practical utility and relevance of such a scale are not yet established in real-world FL applications. Additionally, the paper conducted an experiment using 70 real edge devices from the Theta network to validate the scalability of the paper's benchmark to real-world applications.", "category": "Experimental_Analysis"}, {"question": "What are the detailed information about the actual training times measured on different devices?", "answer": "The paper included an experiment (Exp10 in Appendix) that evaluates the paper’s benchmark using 70 real-world devices from Theta Network. Details for training on each device are included in Figure 16, where local models are trained in an asynchronous way. From Figure 16, the training time for each device in each iteration can be clearly seen. Detailed information for each device is also included in Figure 17. The description of the experiment is as follows:\n\nExp10: Evaluations in real-world applications.\nEdge devices from the Theta network are utilized to validate the scalability of FedSecurity to real-world applications. The FL client package is integrated into Theta’s edge nodes, which periodically fetch data from the Theta back-end. Subsequently, the FL training platform capitalizes on these Theta edge nodes and their associated data to train, fine-tune, and deploy machine learning models. m-Krum is selected as the defense, and the Byzantine attack of random mode is chosen as the attack.\n\nConsidering the challenges posed by real-world environments, such as devices equipped solely with CPUs (lacking GPUs), potential device connectivity issues, network latency, and limited storage on edge devices (for instance, some mobile devices might have less than 500MB of available storage), a simple task is chosen by employing the MNIST dataset for a logistic regression task.\n\nIn the paper’s experimental setup, 70 client edge devices are deployed, with 7 of these designated as malicious for each FL training round. For m-Krum, m is set to 35, meaning that 35 out of the 70 local models are involved in aggregation during each FL training round. As illustrated in Figure 18, m-Krum mitigates the adversarial effect of the random-mode Byzantine attack. A screenshot of the platform is also included, as shown in Figure 16 for the FL training process and Figure 17 for the training status of each device.", "category": "Experimental_Exposition"}, {"question": "What ablation studies have been conducted to evaluate the impact of various attacks and defenses on model accuracy and loss in federated learning?", "answer": "The paper includes 10 different experiments that evaluate the impact of common attacks and defenses on model accuracy and loss in federated learning. These experiments cover various computer vision and NLP tasks, including those involving large language models, and a real-world experiment using 70 edge devices from Theta Network. The experiments assess performance under different configurations, including varying numbers of clients, malicious clients, data distributions, and models. Key findings include: i) the Byzantine attack of random mode effectively decreases the global model's test accuracy, while m-Krum provides robustness against various attacks; ii) defense mechanisms, while mitigating attacks, may negatively affect aggregation results and model performance. However, in actual FL systems, attacks are infrequent. Therefore, it’s crucial to weigh the benefits against potential drawbacks before integrating a defense mechanism into real systems.", "category": "Experimental_Analysis"}, {"question": "What is the current status of implementing attacks and defenses in the paper's benchmark, and why did the paper choose not to include the backdoor success rate metric?", "answer": "Currently, the benchmark includes 14 defenses and 8 attacks, as detailed in the provided link[1]. For the scope of this paper, the primary goal is to maintain a fair comparison of different attacks, including various modes of Byzantine attacks aimed at preventing global model convergence, as well as backdoor attacks attempting model replacement via a malicious local model. In this context, test accuracy is deemed the most suitable metric to evaluate the effectiveness of different defenses against both Byzantine and backdoor attacks. While the backdoor success rate is indeed an important metric for evaluating backdoor attacks, considering the scope and page limitations of this paper, it was not included in the comparative analysis.\n\n[1]https://gitfront.io/r/Submission1595/4rUJj4M2NPzy/FedSecurity/blob/readme.md", "category": "Motivation_Analysis"}, {"question": "What are the unique challenges and differences between using Large Language Models (LLMs) and traditional machine learning models, particularly in the context of federated learning and security mechanisms?", "answer": "Existing research has not yet investigated applying the attack and defense mechanisms to federated LLMs. In contrast to traditional small models, LLMs are distinguished by the large number of parameters and complex training datasets obtained from unregulated sources, which could introduce challenges when applying attacks and defenses on top of them. Thus, investigating the performance of the most common and widely utilized federated learning attack and defense mechanisms in the context of federated LLMs is a valuable endeavor that may stimulate a deeper understanding of the security concept in federated LLMs.", "category": "Method_Comparison"}, {"question": "How are data reconstruction attacks addressed within the open-sourced FL benchmark?", "answer": "Basically, the paper’s attack and defense library serves as a complementary part of an open-sourced FL benchmark. The paper enables the attacks or defenses by activating the security components using a configuration file. Similarly, the FL benchmark also includes a privacy library—a differential privacy (DP) library—which can protect the models against data construction attacks. To enable privacy-related attacks and defenses, the paper activates attacks in the FedSecurity and a DP mechanism in the DP library.", "category": "Experimental_Setup"}, {"question": "Were certain privacy attacks excluded from the benchmark due to their strong assumptions and limited practicality in real-world scenarios?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the paper focused on evaluating the most commonly considered and state-of-the-art defense mechanisms?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the backdoor success rate used as a metric to evaluate backdoor attacks in the current version of the paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the library cover all types of privacy attacks, including adversarial examples, property inference attacks, and membership inference attacks?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are attacks/defenses for NLP tasks like Neurotoxin currently included in the benchmark for LLM and smaller models?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "3mXJ9o2DNx", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Connecting Domains and Contrasting Samples: A Ladder for Domain Generalization", "authors": ["Tianxin Wei", "Yifan Chen", "Xinrui He", "Jingrui He"], "abstract": "Distribution shifts between training and testing datasets, contrary to classical machine learning assumptions, frequently occur in practice and impede model generalization performance. Studies on domain generalization (DG) thereby arise, aiming to predict the label on unseen target domain data by only using data from source domains. In the meanwhile, the contrastive learning (CL) technique, which prevails in self-supervised pre-training, can align different augmentation of samples to obtain invariant representation. It is intuitive to consider the class-separated representations learned in CL are able to improve domain generalization, while the reality is quite the opposite: people observe directly applying CL deteriorates the performance. We analyze the phenomenon with the CL theory and discover the lack of domain connectivity in the DG setting causes the deficiency. Thus we propose domain-connecting contrastive learning (\\model) to enhance the conceptual connectivity across domains and obtain generalizable representations for DG. Specifically, more aggressive data augmentation and cross-domain positive samples are introduced into self-contrastive learning to improve domain connectivity. Furthermore, to better embed the unseen test domains, we propose model anchoring to exploit the domain connectivity in pre-trained representations and complement it with generative transformation loss. Extensive experiments on five standard DG benchmarks are provided. The results verify that \\model~outperforms state-of-the-art baselines even without domain supervision.", "keywords": "Distribution shift;contrastive learning;self-supervised learning", "qa_pairs": [{"question": "How does the proposed method for augmenting positive samples differ from existing approaches in [1,2] and contribute to domain generalization?\n\n[1] Towards Principled Disentanglement for Domain Generalization, CVPR 2022\n\n[2] Model-based domain generalization, NeurIPS 2021", "answer": "As referenced in [3,4], prior studies primarily tackled the positive data alignment issue in domain generalization by developing proxy objectives. Meanwhile, studies in [1,2] advanced the alignment of training domains, utilizing specially designed image generators to create domain-transformed positive samples. However, these approaches primarily involve data transforms. In contrast, the paper's work diverges by exhaustively exploring the concept of connectivity from both data and model perspectives. On the data side, the paper improves upon traditional methods that align augmentation views of the same input, as explained in Appendix C. The paper proposes two direct improvements for enhancing domain connectivity through data: 1. implementing more aggressive data augmentation, and 2. broadening the range of positive samples, extending beyond self-augmented outputs to include intra-class samples across different domains. Nevertheless, this data alignment strategy merely enhances the clustering of representations within known domains, and may not effectively bridge the gap between unseen test domains and trained domains within the same class. Therefore, the paper's focus  is extended to the model perspective. The paper observes that using a pre-trained ResNet-50 model, intra-class samples from training and testing domains are scattered yet interconnected. By integrating this observation, the paper presents a unified approach, combining data and model perspectives with self-supervised objectives to bolster connectivity, thereby advancing domain generalization.\n\n\n[3] SelfReg: Self-supervised Contrastive Regularization for Domain Generalization. ICCV 2021\n\n[4] PCL: Proxy-based Contrastive Learning for Domain Generalization. CVPR 2022\n\n\n", "category": "Method_Comparison"}, {"question": "What is the rationale behind using an additional decoder for generative transformation to improve domain generalization?", "answer": "The additional decoder for generative transformation serves as a pivotal proxy objective that facilitates model anchoring. This can be understood from multiple angles: 1. Echoing the findings in [1] that directly aligning positive pairs across vastly different domains often results in poor performance, the paper's research similarly identifies a substantial gap in the representations of pre-trained and fine-tuned models. Direct alignment using contrastive learning, as evidenced by the paper's findings in Table 2, tends to be sub-optimal. In response, the paper introduces the concept of variational generative loss to comprehend the transformation process and bridge these representational gaps. 2. Additionally, the generative transformation module is designed to reconstruct the features of the pre-trained model at an intra-sample level. This complements the inter-sample level supervision provided by contrastive loss. The module, along with its associated loss function, is intended to provide a more enriched supervised signal, encapsulating crucial within-sample information. This, in turn, supports and enhances the anchoring of the pre-trained model.\n\n[1]PCL: Proxy-based Contrastive Learning for Domain Generalization. CVPR 2022", "category": "Motivation_Analysis"}, {"question": "How does the experimental improvement of the proposed methods (DCCL, SelfReg, PCL, and MIRO) compare to the current state-of-the-art (SOTA) benchmarks on the VLCS and OfficeHome datasets?", "answer": "The proposed methods– DCCL, SelfReg, PCL, and MIRO – show improvements over the SOTA benchmarks on the VLCS and OfficeHome datasets by margins of 0.4/1.1, 0.3/-0.8, -1.0/1.0, and 0.5/0.8 for DCCL, SelfReg, PCL, and MIRO, respectively. Additionally, achieving significant absolute improvements in the domain generalization community is challenging due to factors like the same backbone, fixed data splitting, and long-term development, often requiring the combined efforts of multiple research papers.", "category": "Method_Comparison"}, {"question": "Was an ablation study conducted on benchmarks that achieved a performance gain of 1.0 or less, and what were the results?", "answer": "Yes, an ablation study was conducted on the VLCS dataset, which achieved a performance gain of 0.9. The results confirm that the three components identified contribute consistently to the effectiveness, as detailed in the paper. \n\n|   Algorithm  |   C    |    L   |   S   |    V|Avg          |\n|:------------:|:---------------:|:------:|:---------------:|:---------------:|:----------------------:|\n|     SWAD     |       98.8      |  63.3  |       75.3      |       79.2      |          79.1          |\n| DCCL w/o CDC |       98.9      |  63.8  |       75.6      |       79.5      |          79.4          |\n| DCCL w/o PMA |       98.6      |  63.7  |       75.7      |       79.3      |          79.3          |\n|  DCCL w/o GT |       98.7      |  **64.3**  |       75.2      |       80.2      |          79.6          |\n| DCCL| **99.1** | 64.0 | **76.1** | **80.7** | **80.0** |", "category": "Experimental_Exposition"}, {"question": "What is the central theme of the paper, and how does the generative transformation module contribute to the model's connectivity enforcement?", "answer": "The paper's central theme is not just introducing the concept of connectivity and theoretical analysis, but rather emphasizing connectivity enforcement from both data and model perspectives. The generative transformation is designed as a pivotal proxy objective that facilitates model anchoring. This can be understood from multiple angles: (i) Echoing the findings in [1], which point out that directly aligning positive pairs across vastly different domains often results in poor performance, the paper similarly identifies a substantial gap in the representations of pre-trained and fine-tuned models. Direct alignment using contrastive learning, as evidenced by the paper’s findings in Table 2, tends to be sub-optimal. In response, the paper introduces the concept of variational generative loss to comprehend the transformation process and bridge these representational gaps. (ii) Additionally, the generative transformation module is designed to reconstruct the features of the pre-trained model at an intra-sample level. This complements the inter-sample level supervision provided by contrastive loss. The module, along with its associated loss function, is intended to provide a more enriched supervised signal, encapsulating crucial within-sample information. This, in turn, supports and enhances the anchoring of the pre-trained model.\n\n[1]Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap. ICLR 2022.", "category": "Method_Mechanics"}, {"question": "How does the toy experiment in Appendix C illustrate the drawbacks of directly applying SCL to DG?", "answer": "In Figure 6, the experiment visualizes domains $d_1$ and $d_2$ using slashes and spots, and classes 1 and 2 using orange and blue rectangles. The mapping function $\\varphi (\\theta(x)) := (\\cos (\\theta), \\sin(\\theta))$ with $\\theta(x) = (x - \\mathrm{sgn}(y)) \\pi$ learned on domain $d_1$ can perfectly classify the samples, and the mapping attains perfect alignment and uniformity (the objective of SCL). However, when applied to domain $d_2$, the classifier fails completely(0% acc), demonstrating the drawback.", "category": "Experimental_Exposition"}, {"question": "Why does the study include both samples from the same domain and different domains as 'cross-domain' positive pairs, and how does this approach impact the results?", "answer": "The study includes both samples from the same and different domains as 'cross-domain' positive pairs because domain labels are unavailable. The rationale for using cross-domain contrast (CDC) is detailed in Appendix A.5, showing an accuracy of 71.8% (Table 2). A baseline experiment using only within-domain positive samples resulted in a drop in accuracy (from 71.8% to 70.4%), while an oracle experiment using exclusively cross-domain positive pairs yielded similar results (from 71.8% to 71.9%). These results validate the approach and suggest that effective utilization of domain information, if available, requires thoughtful design for notable improvements.", "category": "Experimental_Analysis"}, {"question": "Why is the generative transformation loss necessary in Section 3.4, given that a pre-trained model anchoring is already in place?", "answer": "The generative transformation is designed as a pivotal proxy objective that facilitates model anchoring. This can be understood from multiple angles: (i) Echoing the findings in [1], which point out that directly aligning positive pairs across vastly different domains often results in poor performance, the paper similarly identifies a substantial gap in the representations of pre-trained and fine-tuned models. Direct alignment using contrastive learning, as evidenced by the paper’s findings in Table 2, tends to be sub-optimal. In response, the paper introduces the concept of variational generative loss to comprehend the transformation process and bridge these representational gaps. (ii) Additionally, the generative transformation module is designed to reconstruct the features of the pre-trained model at an intra-sample level. This complements the inter-sample level supervision provided by contrastive loss. The module, along with its associated loss function, is intended to provide a more enriched supervised signal, encapsulating crucial within-sample information. This, in turn, supports and enhances the anchoring of the pre-trained model.\n\n[1]Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap. ICLR 2022.", "category": "Motivation_Analysis"}, {"question": "What are the specific challenges encountered when applying Self-Contrastive Learning (SCL) in the domain generalization (DG) setting, and how do these challenges affect performance?", "answer": "Self-Contrastive Learning (SCL) aligns augmentation views of the same input and performs well in unsupervised pre-training tasks. However, it does not naturally fit the domain generalization setting because it assumes access to sample $x$ from the whole data distribution, whereas in DG, only partial domains are accessible during training. This mismatch can lead to sub-optimal performance if the classical contrastive learning loss is used. Even achieving optimal SCL loss does not guarantee good performance in DG. For example, in Figure 6, the paper visualizes the example. Slashes and spots are used to represent domains $d_1$ and $d_2$; orange and blue rectangles respectively denote classes 1 and 2. The mapping function $\\varphi (\\theta(x)) := (\\cos (\\theta), \\sin(\\theta))$ with $\\theta(x) = (x - \\mathrm{sgn}(y)) \\pi$ learned on domain $d_1$ can perfectly classify the samples, and the mapping attains perfect alignment and uniformity (the objective of SCL). However, when applied to a new domain $d_2$, the classifier completely fails (0% acc). This example highlights the potential challenges in applying self-contrastive learning effectively in the domain generalization context. ", "category": "Method_Mechanics"}, {"question": "What is the motivation behind the need for more intensive data augmentation and the inclusion of cross-domain positive samples?", "answer": "The need for more intensive data augmentation and the inclusion of cross-domain positive samples arose because SCL (Supervised Contrastive Learning) in the previous example failed to achieve intra-class connectivity due to insufficient data augmentation and domain-separated representations, leading to poor generalization performance. To address this, two approaches were proposed: 1. applying more aggressive data augmentation and 2. expanding the scope of positive samples, from solely self-augmented outputs $a(x)$ to the augmentation of intra-class samples across domains.", "category": "Motivation_Analysis"}, {"question": "What is the rationale behind using model anchoring with domain connectivity in pre-trained representations and how does it synergize with generative transformation loss?", "answer": "The rationale for model anchoring lies in addressing the limitation of previous data transform-based connectivity improvements, which only enhance clustering within known domains and fail to bridge gaps between unseen test and trained domains. By focusing on the model perspective, it is observed that intra-class samples from training and testing domains are scattered yet interconnected in a pre-trained ResNet-50 model. The paper's unified approach combines data and model perspectives with self-supervised objectives to improve connectivity and domain generalization. Generative transformation acts as a crucial proxy objective to facilitate model anchoring.", "category": "Motivation_Analysis"}, {"question": "How is the connectivity defined in the paper, and how does it differ from the connectivity proposed in [1]?\n\n[1] Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation, ICML 2022", "answer": "The connectivity in the paper is defined based on cross-domain sample connectivity within the class. In contrast, [1] assumes that data adheres to the stochastic block model (SBM) and uses SBM's edge probability to define connectivity. The paper has referenced [1] and expanded the discussion to better articulate the approach. The paper's central theme is not just introducing the concept of connectivity, but rather emphasizing its enforcement from both data and model perspectives. For an intuitive understanding of the connectivity in the paper, it introduced another quantitative metric to help evaluate whether the representation space is 'well-connected' in Appendix A.4. For images within the same class, images are taken as nodes and a graph is constructed, only connecting two nodes when their distance on the pre-trained space is smaller than a threshold. The smallest possible threshold which makes the graph connected is denoted as τ, and the mean and the std of the pairwise distances are respectively denoted as μ and σ. The paper can thus use (τ - μ)/σ as a metric to describe the connectivity of the representations.", "category": "Method_Mechanics"}, {"question": "What is the high-level motivation for using the generative transformation loss, and why is it necessary to use augmentation of the same sample to acquire class knowledge more effectively?", "answer": "The generative transformation is designed as a pivotal proxy objective that facilitates model anchoring. This can be understood from multiple angles: (i) Echoing the findings in [1], which point out that directly aligning positive pairs across vastly different domains often results in poor performance, the paper similarly identifies a substantial gap in the representations of pre-trained and fine-tuned models. Direct alignment using contrastive learning, as evidenced by the paper’s findings in Table 2, tends to be sub-optimal. In response, the paper introduces the concept of variational generative loss to comprehend the transformation process and bridge these representational gaps. (ii) Additionally, the generative transformation module is designed to reconstruct the features of the pre-trained model at an intra-sample level. This complements the inter-sample level supervision provided by contrastive loss. The module, along with its associated loss function, is intended to provide a more enriched supervised signal, encapsulating crucial within-sample information. This, in turn, supports and enhances the anchoring of the pre-trained model.\n\n[1]Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap. ICLR 2022.", "category": "Motivation_Analysis"}, {"question": "Does the paper overstate its claims regarding the theoretical analysis of why Contrastive Learning harms the performance of Domain Generalization models?", "answer": "The paper's claim regarding the theoretical analysis is conservative. The paper's central theme is not just introducing the concept of connectivity and theoretical analysis, but rather emphasizing connectivity enforcement from both data and model perspectives. Meanwhile, to clearly convey the main contribution of the paper to the audience, the paper doesn't present any decorative theorems/propositions. It instead aims to propose useful DG methods for practitioners and attributes most of the theoretical analysis to the motivating theory paper [1]. The paper leverages the theoretical motivation behind [1], as clearly indicated in Section 3.1, for the new setting of DG; the analysis in [1] can be applied to DG settings with proper adaptations and regularity conditions. The paper does not take that as its contribution.\n\n[1]Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap. ICLR 2022.", "category": "Method_Disambiguation"}, {"question": "What is the primary focus of the paper in terms of improving domain generalization, and how does it differ from traditional methods of aligning representations across different domains?", "answer": "The key idea of the paper is not aligning the representations across different domains, but **a unified approach**, combining data and model perspectives with self-supervised objectives to bolster **connectivity**, thereby advancing domain generalization. From the data perspective, the paper builds upon the shortcomings of traditional methods that align augmentation views of the same input, as explained in Appendix C. Two direct improvements for enhancing domain connectivity through data: 1. implementing more aggressive data augmentation, and 2. broadening the range of positive samples, extending beyond self-augmented outputs to include intra-class samples across different domains. Nevertheless, this data alignment strategy merely enhances the clustering of representations within known domains, and may not effectively bridge the gap between unseen test domains and trained domains within the same class. Therefore, the paper's focus is extended to the model perspective. The paper observes that using a pre-trained ResNet-50 model, intra-class samples from training and testing domains are scattered yet interconnected. By integrating this observation, the paper presents a holistic approach.", "category": "Method_Comparison"}, {"question": "What is the role and rationale of the proposed Generative Transformation Loss in addressing the representational gaps between pre-trained and fine-tuned models?", "answer": "The generative transformation is designed as a pivotal proxy objective that facilitates model anchoring. This can be understood from multiple angles: (i) Echoing the findings in [1], which point out that directly aligning positive pairs across vastly different domains often results in poor performance, the paper similarly identifies a substantial gap in the representations of pre-trained and fine-tuned models. Direct alignment using contrastive learning, as evidenced by the paper’s findings in Table 2, tends to be sub-optimal. In response, the paper introduces the concept of variational generative loss to comprehend the transformation process and bridge these representational gaps. (ii) Additionally, the generative transformation module is designed to reconstruct the features of the pre-trained model at an intra-sample level. This complements the inter-sample level supervision provided by contrastive loss. The module, along with its associated loss function, is intended to provide a more enriched supervised signal, encapsulating crucial within-sample information. This, in turn, supports and enhances the anchoring of the pre-trained model.\n\n[1]Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap. ICLR 2022.", "category": "Method_Disambiguation"}, {"question": "What are the key differences between the connectivity concepts and methods proposed in this paper and those in DN$^2$A [1], and how does this paper's approach differ in terms of novelty and contribution?", "answer": "As the mathematical notations are non-standard in [1] (such as the expectation of distribution in Def. 1 and minimum product $\\min x_i^+ x_j^+$ of two augmented samples in Eqn. (14)), there might be misunderstanding.\n\n- The concepts of 'connectivity' differ in two papers.\n\n[2] motivates both this project and [1] (as claimed in Section 3.2 of [1]). In [2], the original definition of intra-class connectivity appears in Assumption 4.5 just means the augmentation graph (c.f. their Def. 4.4) is connected. The paper follows the same definition of 'connectivity' in the paper.\n\nHowever, [1] is based on the concepts of intra-domain, intra-class, and semantic connectivity as \n\n> Definition 1. (Semantic Connectivity) For any input $x \\in \\mathcal{X}$, let $\\mathcal{A}(\\cdot \\mid x)$ be the distribution of its augmentations $A$. Let $C$ be the joint distribution on $\\mathcal{X} \\times \\mathcal{X}$ of augmented views of images $x_i, x_j$ as $C\\left(x_i^{+}, x_j^{+}\\right)=\\mathcal{A}\\left(x_i^{+} \\mid x_i\\right) \\mathcal{A}\\left(x_j^{+} \\mid x_j\\right)$. Then the paper defines intra-domain $C_\\alpha$ and intra-class $C_\\beta$ connectivity.\n> \n> $$C_\\alpha:=\\mathbb{E}_{d \\sim P_D^S} \\mathbb{E}_{x_i, x_j \\sim P_d^{S_{\\mathrm{UL}}}} C\\left(x_i^{+}, x_j^{+}\\right), \\quad C_\\beta:=\\mathbb{E}_{y \\sim P_Y^S} \\mathbb{E}_{x_i, x_j \\sim P_y^{S_{\\mathrm{UL}}}} C\\left(x_i^{+}, x_j^{+}\\right)$$\n\n> Then, the semantic connectivity is defined as $C_s:=C_\\beta / C_\\alpha$.\n\nFor the reader's convenience, the paper notes its definition of the joint distribution $C()$ is uncommon: it is a constant multiple of the so-called minimum product \n\n$$e_A\\left(x_i^{+}, x_j^{+}\\right)=\\min _{x_i^{+} \\in A\\left(x_i\\right), x_j^{+} \\in A\\left(x_j\\right)} x_i^{+} x_j^{+}.$$\n\nIt is obvious that the connectivity concept based on the joint distribution in their paper is quite different from the paper and [2]; therefore the theoretical motivation behind two papers are indeed different.\n\n- The methods proposed are different.\n\nExcept for the common strategy of strong augmentation recommended by the contrastive learning theory paper [2], the paper's proposed methods are different from the ones in [1]. They propose two nearest-neighbor-based methods for constructing positive pairs, while the paper's main contribution lies in the exploitation of both the pre-trained models and the intra-class data connectivity.\n\n[1] Promoting Semantic Connectivity: Dual Nearest Neighbors Contrastive Learning for Unsupervised Domain Generalization. CVPR 2023.\n\n[2] Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap. ICLR 2022.\n\n", "category": "Method_Comparison"}, {"question": "What is the rationale behind designing the generative transformation module and how does it facilitate model anchoring?", "answer": "The generative transformation is designed as a pivotal proxy objective that facilitates model anchoring. This can be understood from multiple angles: (i) Echoing the findings in [1], which point out that directly aligning positive pairs across vastly different domains often results in poor performance, the paper similarly identifies a substantial gap in the representations of pre-trained and fine-tuned models. Direct alignment using contrastive learning, as evidenced by the paper’s findings in Table 2, tends to be sub-optimal. In response, the paper introduces the concept of variational generative loss to comprehend the transformation process and bridge these representational gaps. (ii) Additionally, the generative transformation module is designed to reconstruct the features of the pre-trained model at an intra-sample level. This complements the inter-sample level supervision provided by contrastive loss. The module, along with its associated loss function, is intended to provide a more enriched supervised signal, encapsulating crucial within-sample information. This, in turn, supports and enhances the anchoring of the pre-trained model.\n\n[1]PCL: Proxy-based Contrastive Learning for Domain Generalization. CVPR 2022", "category": "Motivation_Analysis"}, {"question": "What are the results of the additional ablation study conducted on the VLCS dataset to verify the effectiveness of the proposed modules?", "answer": "The additional ablation study on the VLCS dataset shows the following results: \n\n|   Algorithm  |        C        |    L   |        S        |        V        |           Avg          |\n|:------------:|:---------------:|:------:|:---------------:|:---------------:|:----------------------:|\n|     SWAD     |       98.8      |  63.3  |       75.3      |       79.2      |          79.1          |\n| DCCL w/o CDC |       98.9      |  63.8  |       75.6      |       79.5      |          79.4          |\n| DCCL w/o PMA |       98.6      |  63.7  |       75.7      |       79.3      |          79.3          |\n|  DCCL w/o GT |       98.7      |  **64.3**  |       75.2      |       80.2      |          79.6          |\n|     DCCL     | **99.1** | 64.0 | **76.1** | **80.7** | **80.0** |", "category": "Experimental_Exposition"}, {"question": "Does the paper primarily focus on presenting decorative theorems or propositions related to the theoretical analysis of Contrastive Learning in Domain Generalization?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper claim the theoretical analysis of why Contrastive Learning harms Domain Generalization models as its original contribution?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "p09XyFxZkc", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models", "authors": ["Yaohui Wang", "Xinyuan Chen", "Xin Ma", "Shangchen Zhou", "Ziqi Huang", "Yi Wang", "Ceyuan Yang", "Yinan He", "Jiashuo Yu", "Peiqing Yang", "Yuwei Guo", "Tianxing Wu", "Chenyang Si", "Yuming Jiang", "Cunjian Chen", "Chen Change Loy", "Bo Dai", "Dahua Lin", "Yu Qiao", "Ziwei Liu"], "abstract": "This work aims to learn a high-quality text-to-video (T2V) generative model by leveraging a pre-trained text-to-image (T2I) model as a basis. It is a highly desirable yet challenging task to simultaneously a) accomplish the synthesis of visually realistic and temporally coherent videos while b) preserving the strong creative generation nature of the pre-trained T2I model. To this end, we propose LaVie, an integrated video generation framework that operates on cascaded video latent diffusion models, comprising a base T2V model, a temporal interpolation model, and a video super-resolution model. Our key insights are two-fold: 1) We reveal that the incorporation of simple temporal self-attentions, coupled with relative positional encoding, adequately captures the temporal correlations inherent in video data. 2) Additionally, we validate that the process of joint image-video fine-tuning plays a pivotal role in producing high-quality and creative outcomes. To enhance the performance of LaVie, we contribute a comprehensive and diverse video dataset named Vimeo25M, consisting of 25 million text-video pairs that prioritize quality, diversity, and aesthetic appeal. Extensive experiments demonstrate that LaVie achieves state-of-the-art performance both quantitatively and qualitatively. Furthermore, we showcase the versatility of pre-trained LaVie models in various long video generation and personalized video synthesis applications.", "keywords": "text-to-video generation;diffusion models", "qa_pairs": [{"question": "What are the specific system-level contributions of the proposed method, given that it combines existing techniques such as text-to-image generation, frame interpolation, and super-resolution?", "answer": "Since the 'base-interpolation-SR' framework has become a standard design for high-quality video generation, the paper's contributions are system-level rather than model-level. For the entire system, this paper presents the following contributions:\n\n- **Model design**: The paper proposes a simple yet effective method to involve RoPE-equipped temporal self-attention in the paper's base, interpolation, and video super resolution models. Experiments show that such design is able to produce videos with temporal coherency.\n- **Training**: The paper studied different training schemes in the paper's base, interpolation, and super-resolution models, and proved that when given a pre-trained text-to-image generation model, joint image-video fine-tuning is an effective way to overcome catastrophic forgetting, as well as to obtain high quality videos.\n- **Evaluation**: T2V is a novel video generation task, where novel evaluation is required. Different from other approaches that simply report human preference and FVD, the paper conducts a systematic human evaluation on various dimensions to compare with state-of-the-art.\n- **Dataset**: The paper proposes a novel text-video dataset Vimeo25M to improve the performance of text-to-video generation tasks.", "category": "Method_Disambiguation"}, {"question": "How does the proposed framework in this paper differ from Dandi et al.'s[1] work in terms of methodology and training approach?\n\n[1]Yatin Dandi, Aniket Das, Soumye Singhal, Vinay Namboodiri, and Piyush Rai. Jointly trained image and video generation using residual vectors. In WACV, 2020.", "answer": "This paper[1] proposed a GAN-based framework to generate videos. Dandi et al.'s work combined an LSTM and an image GAN to learn disentangled temporal and spatial information. Joint image-video training is conducted to learn the entire framework. The proposed framework is a large-scale diffusion-based system, unlike Dandi et al.'s GAN-based framework. In the proposed framework, the temporal module is added after each spatial block due to the absence of a compressed latent space in diffusion models. Joint image-video fine-tuning uses different large-scale image and video datasets, with images and videos concatenated in the temporal axis within each batch. In contrast, Dandi et al. used the same video to optimize both spatial and temporal modules. Additionally, the proposed framework trains spatial and temporal modules jointly, whereas Dandi et al. employed a two-step training scheme to train the framework separately.", "category": "Method_Comparison"}, {"question": "What is the impact of the temporal module, joint image-video training, and the Vimeo25M dataset on the video generation performance, and how do these components compare in terms of their contributions to the final model?", "answer": "For the temporal module, small-scale comparative experiments were conducted on UCF101. Three different settings were explored: 1) The paper's proposed temporal module, 2) replacing RoPE in the paper's module with absolute PE, and 3) spatial-temporal self-attention used in [1]. Three models were trained for the same iterations and the FVD results are reported in Table:\n\n| Method | the paper | absolute PE | spatio-temporal self-attention |\n|--------|-------------|-------------|-------------------------------|\n| FVD    | 611         | 625         | 635                           |\n\nQuantitative results demonstrate the effectiveness of the paper's proposed temporal module.\n\nFor joint image-video fine-tuning, several qualitative examples show that without joint image-video fine-tuning, the network catastrophically forgets previously learned concepts and overfits to training data very quickly. On the contrary, using images with videos together to fine-tune the entire framework helps combine different concepts in both image and video data, thereby improving the diversity of the results.\n\nRegarding the usage of Vimeo25M data, without using Vimeo, higher FVD is obtained on zero-shot generation on UCF101. In addition, as shown in Tab. 4, since Vimeo25M has better aesthetics and large amounts of human data, the generated results have smoother motion and better face/body/hand quality compared with other methods.\n\n[1] Wu et al, Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation, CVPR 2023\n", "category": "Experimental_Exposition"}, {"question": "What is the benefit of the technique described on page 4, where each frame generated through interpolation replaces the corresponding input frame, compared to conventional video frame interpolation methods?", "answer": "The benefit of the technique is that it reconstructs input frames during the interpolation process, unlike conventional methods that directly combine input frames with generated frames. Directly reusing the input frames in generated videos leads to flickering in the final results. Generating the entire video, including input frames, makes the results have more temporally coherency.", "category": "Method_Comparison"}, {"question": "How do the experimental results on UCF101 support the claim that simple temporal self-attention with RoPE(Su et al., 2021[1]) adequately captures temporal correlations in video data compared to more complex architectural designs?\n\n[1]Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021.", "answer": "For the temporal module, small-scale comparative experiments were conducted on UCF101. Three different settings were explored: 1) The paper’s proposed temporal module, 2) replacing RoPE in the paper’s module with absolute PE, 3) spatial-temporal self-attention used in [1]. Three models were trained for the same number of iterations, and their FVD results are reported in Table:\n\n| Method | Ours | absolute PE | spatio-temporal self-attention |\n|--------|------|-------------|--------------------------------|\n| FVD    | 611  | 625         | 635                            |\n\nWhile 3) is much more complex than the paper’s proposed temporal module, no improvement in visual quality was found, and training is slower. Small-scale quantitative evaluation demonstrates the effectiveness of the paper’s proposed temporal module.", "category": "Experimental_Analysis"}, {"question": "What are the specific advantages of using Rotary Positional Encoding (RoPE) compared to other positional encoding methods like learnable embeddings, particularly in the context of video generation?", "answer": "Experimental results show that RoPE achieves lower FVD (Fréchet Video Distance) compared to absolute positional encoding on the UCF101 dataset. Additionally, RoPE enables extrapolation to longer video lengths beyond the training data, such as generating 32, 48, or 64 frames from a 16-frame training setup, while maintaining good spatial quality. However, flickering increases as the generated videos become longer, indicating that designing better positional encoding for video generation remains an open challenge.", "category": "Method_Comparison"}, {"question": "How does text conditioning, which primarily interacts with spatial attention, influence or control motion within the model?", "answer": "The paper has explored adding an extra temporal-level cross-attention for text input. However, no huge differences were observed in the generated results. The reason could be that Text conditioning, although primarily interacting with spatial attention, serves as a condition for the entire video. When the spatio-temporal networks are optimized, these conditions also affect the temporal module, thereby influencing motion control.", "category": "Method_Mechanics"}, {"question": "What is the technical novelty of the proposed work, given that the paper's two findings have both been proposed in previous works? 1) temporal attention over a pretrained text-to-image model could address the temporal consistency model well. This idea is first stated in 'AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning' and then pointed out in 'MagicEdit: High-Fidelity and Temporally Coherent Video Editing'. 2) Joint image-video training is a well-known technique to improve the effectiveness of training text-to-video models. It was first proposed by Jonathan Ho et al. in 'Video Diffusion Models'.", "answer": "LaVie is a concurrent work compared to AnimateDiff and MagicEdit since both works have not been peer-reviewed and are only on arXiv. 2) Joint image-video training is a well-known technique for both video understanding [1] and video generation for a long time. In this work, the paper studied how to leverage this technique to fine-tune a pretrained Stable Diffusion initialized T2V system. The paper demonstrates that joint image-video fine-tuning is the key to obtaining good results for base, interpolation, as well as video super resolution models. The paper has also shown that such technique helps to prevent catastrophic forgetting. However, there are still open questions about using this technique, such as 1) is there any balance between image and video, and 2) how to also leverage images in temporal modules to make the best usage of the entire training data. Such questions are also related to design effective learning approaches in video diffusion models. Therefore, the paper's contributions are not in this technique itself, but the analysis of this technique, as well as opening new doors for future usage and improvement.\n\n[1] Bain et al., Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval, ICCV 2021\n\n", "category": "Method_Disambiguation"}, {"question": "What are the specific contributions of the proposed cascaded diffusion model compared to the standard 'base-interpolation-SR' framework used in other works like 'Make-A-Video', 'Imagen Video', and 'Align your Latents'?", "answer": "Since the 'base-interpolation-SR' framework has become a standard design for high-quality video generation, the paper's contributions are system-level rather than model-level. For the entire system, this paper presents the following contributions:\n\n- **Model design**: The paper proposes a simple yet effective method to involve RoPE-equipped temporal self-attention in the paper's base, interpolation, and video super resolution models. Experiments show that such design is able to produce videos with temporal coherency.\n- **Training**: The paper studied different training schemes in the paper's base, interpolation, and super-resolution models, and proved that when given a pre-trained text-to-image generation model, joint image-video fine-tuning is an effective way to overcome catastrophic forgetting, as well as to obtain high quality videos.\n- **Evaluation**: T2V is a novel video generation task, where novel evaluation is required. Different from other approaches that simply report human preference and FVD, the paper conducts a systematic human evaluation on various dimensions to compare with state-of-the-art.\n- **Dataset**: The paper proposes a novel text-video dataset Vimeo25M to improve the performance of text-to-video generation tasks.", "category": "Method_Comparison"}, {"question": "What are the distinguishing aspects of the technical innovations in this work compared to previous works like VDM and Align your latent, particularly in terms of system-level contributions?", "answer": "Since the 'base-interpolation-SR' framework has become a standard design for high-quality video generation, the paper's contributions are system-level rather than model-level. For the entire system, this paper presents the following contributions: \n\n- **Model design**: The paper proposes a simple yet effective method to involve RoPE-equipped temporal self-attention in the paper's base, interpolation, and video super resolution models. Experiments show that such design is able to produce videos with temporal coherency.\n- **Training**: The paper studied different training schemes in the paper's base, interpolation, and super-resolution models, and proved that when given a pre-trained text-to-image generation model, joint image-video fine-tuning is an effective way to overcome catastrophic forgetting, as well as to obtain high quality videos.\n- **Evaluation**: T2V is a novel video generation task, where novel evaluation is required. Different from other approaches that simply report human preference and FVD, the paper conducts a systematic human evaluation on various dimensions to compare with state-of-the-art.\n- **Dataset**: The paper proposes a novel text-video dataset Vimeo25M to improve the performance of text-to-video generation tasks.", "category": "Method_Comparison"}, {"question": "What experimental evidence supports the effectiveness of RoPE as a positional encoding and the claim that joint image-video fine-tuning plays a pivotal role?", "answer": "Three different settings were explored: 1) The paper’s proposed temporal module, 2) replacing RoPE in the paper’s module with absolute PE, 3) spatial-temporal self-attention used in [1]. Three models were trained for the same number of iterations, and their FVD results are reported in Table:\n\n| Method                     | The paper | absolute PE | spatio-temporal self-attention |\n|----------------------------|------|-------------|-------------------------------|\n| FVD                        | 611  | 625         | 635                          \n\nSmall-scale experimental results demonstrate the effectiveness of the paper's proposed temporal module. The results of with and without image-video joint fine-tuning show that, without image-video joint tuning, catastrophic forgetting appears in the generated results. For example, the teddy bear and Iron Man could not be well generated, and the entire system overfits to training video data.\n\n[1] Wu et al, Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation, CVPR 2023\n", "category": "Experimental_Exposition"}, {"question": "What does the letter 'E' represent in Figure 2, and what specific criteria were used to construct the Vimeo25M dataset?", "answer": "'E' represents the Encoder in VAE. The Vimeo25M dataset was constructed using PySceneDetect for scene detection and segmentation, camera motion detection, and static video analysis to filter out videos with camera shake and static footage. Segments with stationary or gently moving cameras were retained, while those with entirely static footage were eliminated. Captions with fewer than three words and video segments with fewer than 16 frames were also excluded.", "category": "Experimental_Setup"}]}
{"id": "gDlsMWost9", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Multimodal Chain-of-Thought Reasoning in Language Models", "authors": ["Zhuosheng Zhang", "Aston Zhang", "Mu Li", "hai zhao", "George Karypis", "Alex Smola"], "abstract": "Large language models (LLMs) have shown impressive performance on complex reasoning by leveraging chain-of-thought (CoT) prompting to generate intermediate reasoning chains as the rationale to infer the answer. However, existing CoT studies have primarily focused on the language modality. We propose Multimodal-CoT that incorporates language (text) and vision (images) modalities into a two-stage framework that separates rationale generation and answer inference. In this way, answer inference can leverage better generated rationales that are based on multimodal information. Experimental results on ScienceQA and A-OKVQA benchmark datasets show the effectiveness of our proposed approach. With Multimodal-CoT, our model under 1 billion parameters achieves new state-of-the-art performance on the ScienceQA benchmark. Our analysis indicates that Multimodal-CoT offers the advantages of mitigating hallucination. Code is publicly available at Anonymous.", "keywords": "Chain of Thought Prompting;Language Models;Multimodal Reasoning;Fine-tuning;Natural Language Processing.", "qa_pairs": [{"question": "What is the motivation behind separating the reasoning process into two-stage works in the proposed model?", "answer": "The motivation is to allow the answer inference to leverage better generated rationales based on multimodal information. Separating the training stages enables the obtainment of more effective rationales through a specific rationale generation model, which are then used to predict the final answer. In contrast, as results reported in Table 2, modeling the reasoning process by combining the rationale and answer (RA or AR) would lead to inferior results.", "category": "Motivation_Analysis"}, {"question": "Is multimodal-CoT a prompting technique, and are the references in Table 4 misleading if it is considered as such?", "answer": "No, multimodal-CoT is not a prompting technique; it is a fine-tuned model. While prompting and fine-tuning are both potential solutions for multimodal-CoT, fine-tuning approaches generally achieve better performance and are widely adopted by related studies. Therefore, the paper chose the fine-tuning paradigm to design Multimodal-CoT and focus on the comparison with fine-tuning approaches.", "category": "Method_Disambiguation"}, {"question": "Why were the ScienceQA and A-OKVQA datasets chosen for evaluating the approach?", "answer": "The paper adopts the ScienceQA and A-OKVQA datasets for evaluation because they have annotated rationales as golden labels to assess the quality of the generated rationales and to have an in-depth study on the problem of multimodal CoT, though the annotations are dispensable in real use. The paper’s results compellingly show the feasibility of adaptation to scenarios even without human-annotated rationales, thereby establishing the effectiveness of the approach across diverse tasks. ScienceQA is a large-scale multimodal QA benchmark that covers rich domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills. Indeed, these datasets are relevant and challenging. It is worth noting that existing publications only used one of the datasets for evaluation. In contrast, the paper has taken both of them to verify the effectiveness of Multimodal-CoT [1-5].\n\n[1] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Hauptmann, Yonatan Bisk, Jianfeng Gao. KAT: A Knowledge Augmented Transformer for Vision-and-Language. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (pp. 956-968).\n\n[2] Yuanze Lin, Yujia Xie, Dongdong Chen, Yichong Xu, Chenguang Zhu, Lu Yuan. Revive: Regional visual representation matters in knowledge-based visual question answering. Advances in Neural Information Processing Systems, 35, 10560-10571.\n\n[3] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. The Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023).\n\n[4] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. The Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS 2023).\n\n[5] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention. arXiv preprint arXiv:2303.16199.", "category": "Motivation_Analysis"}, {"question": "Why does the paper focus on encoder-decoder models instead of left-to-right language models, and how does the proposed method perform when compared to popular left-to-right language models?", "answer": "The focus on encoder-decoder models is due to their widespread adoption in multimodal frameworks and existing studies[1] showing that encoder-decoder architectures (e.g., T5) outperform left-to-right decoder-only architectures (e.g., LLaMA/Vinuca) on most tasks. Additionally, the proposed method has been compared to various left-to-right language models (LLaMA-Adapter, LLaVA, InstructBLIP, and Chameleon) in Table 4, demonstrating that it performs better than most of these models despite having fewer than 1B parameters.\n\n[1]Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023.", "category": "Method_Comparison"}, {"question": "What are the key aspects that differentiate this work from prior multimodal LLM research, such as BLIP-2[1] and MINIGPT-4[2], in terms of novelty and contribution?\n\n[1]https://arxiv.org/pdf/2301.12597.pdf)\n\n[2]https://arxiv.org/pdf/2304.10592.pdf)", "answer": "This work differentiates itself from prior multimodal LLM research in three key aspects: (1) It is the first to study Chain-of-Thought (CoT) reasoning across different modalities. (2) It introduces a novel approach to multimodal CoT, distinct from text-only reasoning or VQA methods, by enabling explicit intermediate reasoning with multimodal input using a frozen image encoder and a two-stage framework. (3) The two-stage approach analyzes why naive CoT methods fail in this context and demonstrates how incorporating vision features addresses the problem. Additionally, the approach is shown to be effective across tasks and backbone models, and the work includes in-depth studies to explore the benefits and limitations of Multimodal-CoT.", "category": "Method_Comparison"}, {"question": "What intricate alignment strategies have been explored to enhance the two-stage framework, and how do they compare in terms of performance?", "answer": "The authors explored intricate alignment strategies inspired by BLIP, specifically the Image-grounded text encoder. This encoder injects visual information by adding a cross-attention layer between the self-attention layer and the feed-forward network in each transformer block of the text encoder. The paper's current strategy is similar to the Unimodal encoder used in BLIP, which serves as a baseline for comparison. These results show that using other alignment strategies also outperforms direct answering.\n\n| Method | Accuracy |  \n| --- | --- |  \n| Direct Answering | 82.62 |  \n| Unimodal encoder | 85.31 |  \n| Image-grounded text encoder | 84.60 |  \n\nAdditionally, the authors explore the potential of integrating Multimodal-CoT with more complex models. They find that the paper's method is orthogonal to existing multimodal models (e.g., InstructBLIP) and can potentially be used in conjunction with them to further improve generality. A detailed discussion is provided in Appendix C.3.\n", "category": "Method_Comparison"}, {"question": "What are the potential benefits or shortcomings of adopting a 'QCM->RA' strategy, especially when paired with few-shot demonstrations, and how does it compare to the proposed two-stage pipeline framework?", "answer": "The 'QCM->RA' strategy, which is similar to the standard CoT approach, was initially considered but found to be inferior to direct reasoning (QCM->A), as shown in Table 2. The 'X->RA' strategy is beneficial when the model can generate effective rationales, such as in complex reasoning tasks like arithmetic, commonsense, and logical reasoning. However, in the paper's tasks, the rationales often contain commonsense or logical mistakes. When using the rationales on the output side (RA), the decoder generates the answer conditioned on the previously generated rationales. Consequently, the answer generation is more likely to suffer from the imperfect rationales. In contrast, when the rationales are provided on the input side, they do not directly affect the answer generation process of the decoder, and the model may also easily ignore the imperfect rationales. This finding also applies to the in-context learning baselines with few-shot demonstrations. According to the results in [1], GPT-3 also yields inferior performance with X->RA.\n\n| Method  | Accuracy |\n| :------ | :-------- |\n| X -> A  | 74.0     |\n| X -> AR | 73.6     |\n| X -> RA | 55.4     |\n\nTherefore, the paper finally adopts the two-stage framework and succeeds in improved performance.\n\n[1] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35, 2507-2521.", "category": "Experimental_Analysis"}, {"question": "Is multimodal-CoT considered a prompting technique rather than a fine-tuned model in the study?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "3Zm6wR5Mvc", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "LangNav: Language as a Perceptual Representation for Navigation", "authors": ["Bowen Pan", "Rameswar Panda", "SouYoung Jin", "Rogerio Feris", "Aude Oliva", "Phillip Isola", "Yoon Kim"], "abstract": "We explore the use of language as a perceptual representation for vision-and-language navigation. Our approach uses off-the-shelf vision systems (for image captioning and object detection) to convert an  agent's egocentric panoramic view at each time step  into natural language descriptions. We then finetune a pretrained language model to select an action, based on the current view and the  trajectory history, that would best fulfill the navigation instructions. In contrast to the standard setup which adapts a pretrained language model to work directly with continuous visual features from pretrained vision models, our approach instead uses (discrete) language  as the perceptual representation. We explore two use cases of our language-based navigation ours approach on the R2R vision-and-language navigation benchmark: generating synthetic trajectories from a prompted large language model (GPT-4) with which to finetune a smaller  language model; and sim-to-real transfer where we transfer a policy learned on a simulated environment (ALFRED) to a real-world environment (R2R). Our approach is found to improve upon strong baselines that rely on visual features in settings where only a few gold trajectories (10-100) are available, demonstrating the potential of using language as a perceptual representation for learning navigation agents.", "keywords": "Language Models;Vision-and-Language Navigation;Learn from Synthetic Data;Sim-to-Real", "qa_pairs": [{"question": "Why does the LangNav agent trained on synthetic data converge to notably lower success rates (SR) compared to an agent trained on the full R2R dataset, and what are the broader limitations of synthetic data generation using LLMs’ ‘common-sense’ as opposed to hand-collected datasets like R2R and RxR?", "answer": "The lower success rates (SR) of the LangNav agent trained on synthetic data are primarily due to differences in scene diversity. To compare the quality of synthetic and real-world data, an experiment was conducted using 93 training trajectories for both types of data, where LLaMA2-7B was fine-tuned. The results show that, under the same scene diversity constraints, synthetic trajectories outperform real-world trajectories, likely due to more accurate and diverse language descriptions. Additionally, scaling experiments were conducted using 1000 navigation instructions sampled from various R2R scenes to prompt GPT-4-turbo to generate 2000 synthetic trajectories. These experiments show that increasing synthetic data volume improves performance, even if individual trajectory quality is slightly lower.", "category": "Experimental_Analysis"}, {"question": "How does the paper address the inherent merits of a language-based representation in a VLN system, beyond its focus on low-data regimes?", "answer": "The paper addresses the inherent merits of a language-based representation through three case studies: (1) efficient synthetic data generation, (2) ALFRED-to-R2R transfer, and (3) a new case study focusing on interpretability and editability. In the third case study, the authors randomly selected 10 R2R trajectories where LangNav failed to choose the correct direction, similar to the second example in Figure 4 of the paper. For the steps where LangNav was incorrect, the captions were inspected, and incorrect captions were manually corrected. The paper's LangNav was able to revise its decision correctly in 70% of these trajectories. This case study highlights another benefit of working in language space—improved interpretability and editability.\n", "category": "Experimental_Analysis"}, {"question": "Why is the assumption of low data availability for VLN in home environments considered artificial, and how does the proposed system address this by testing in more varied environments?", "answer": "The assumption of low data availability for VLN in home environments is considered artificial due to the rich and widely available datasets for VLN [1-4] and VLN-CE [5] tasks in household environments. The potential application in some exotic environments would be a strong motivation for LangNav. To validate that LLMs can synthesize useful data in more exotic environments, the authors conducted an experiment where a trajectory was handcrafted in a real office environment and then GPT-4 was prompted to generate synthetic trajectories within the scope of the office environment. The synthetic trajectories contain abundant common object-scene correlations in office environments, exhibit great spatial consistency, and incorporate a significant amount of captions and objects that do not directly relate to the given instruction.However, due to the lack of benchmarks for exotic environments, only qualitative results are available.\n\n[1] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In CVPR, pages 9982–9991, 2020.\n\n[2] Fengda Zhu, Xiwen Liang, Yi Zhu, Qizhi Yu, Xiaojun Chang, and Xiaodan Liang. Soon: Scenario oriented object navigation with graph-based exploration. In CVPR, pages 12689–12699, 2021.\n\n[3] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen ¨ Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In CVPR, pages 3674–3683, 2018.\n\n[4] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4392–4412, 2020.\n\n[5] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In European Conference on Computer Vision, 2020.", "category": "Experimental_Analysis"}, {"question": "What are the fundamental differences between LangNav's sim2real experiments and those of [1,2] in terms of novelty and transfer complexity?\n\n[1] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Cote, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2021. ALFWorld: Aligning text and embodied environments for interactive learning. In International Conference on Learning Representations.\n\n[2] Narasimhan, K., Barzilay, R., and Jaakkola, T. (2018). Grounding language for transfer in deep reinforcement learning. JAIR, 63(1):849–874.", "answer": "LangNav's sim2real experiments differ fundamentally from [1] and [2]. In [1], the agent learns the text policy from TextWorld (which is extended to analog the ALFRED scenes) and then evaluates it in the ALFRED environment. The learning and testing are in the two views (text & vision) of the same underlying world, while LangNav transfers knowledge between two very different underlying worlds (R2R and ALFRED). In [2], the agent transfers knowledge between two game environments (Boulderchase and Bomberman), which have lower transfer complexity compared to LangNav's testbeds (ALFRED and R2R). Additionally, LangNav restricts the policy learned in ALFRED to only a few scenes in R2R, making the transfer more challenging.", "category": "Method_Comparison"}, {"question": "How does the paper address the need for deeper exploration of the coupling between visual perception and language-based decision-making in LangNav, particularly in light of VLMs' failure to identify necessary features for LM policy reasoning?", "answer": "To explore the coupling between visual perception and language-based decision-making, the authors extended RecBERT by concatenating language features to the visual features to represent the candidate image. The results showed that language-based features improved performance in both few-shot and full training set cases, indicating that language as a perceptual representation can provide additional benefits on top of continuous visual features.", "category": "Experimental_Analysis"}, {"question": "How does the LangNav system generalization ability demonstrate that it does not overfit to visual captions/objects and can handle incomplete visual/spatial information in real-world scenarios?", "answer": "The LangNav system demonstrates generalization ability by achieving nontrivial performance when transferring a policy trained on ALFRED to R2R with zero R2R examples. This shows that the agents do not overfit to the captions/objects from the domain and can generalize navigation knowledge across domains even with incomplete visual/spatial information. Additionally, language captions improve performance even in the presence of continuous visual features.", "category": "Experimental_Analysis"}, {"question": "How does the LangNav approach address the concern of limited applicability to real-world robotic systems due to the complexity and variability of real-world environments compared to the training data?", "answer": "Real-world scenarios are much more complex than the R2R environment. Beyond the more complex and varied scene configurations, objects in the real world are also dynamic, rather than static. Because of this, language is an important perceptual representation to bridge the gap between real-world complexity and simulation environments. Case Study 2 demonstrates that combining language’s high-level abstraction with LLMs’ common-sense reasoning offers a more promising way to transfer knowledge between environments. On the other hand, language-based features can also enhance the traditional visual perception module of the navigation agent by providing additional context. To further explore language’s assistance with continuous visual features, the authors extended RecBert by concatenating language features to visual features for candidate image representation. The results, listed in the tables, show that language features improve performance in both few-shot and full training set cases, indicating their effectiveness in aiding the visual perception module of the VLN agent.\n\n| eval env   | train trajectories | perceptual features | SR↑  | SPL↑ |\n|------------|--------------------|---------------------|------|------|\n| val_seen   | 100                | vision only         | 19.9 | 18.6 |\n| val_unseen | 100                | vision only         | 19.0 | 17.4 |\n| val_seen   | 100                | vision + language   | 22.8 | 21.5 |\n| val_unseen | 100                | vision + language   | 19.3 | 18.0 |\n| val_seen   | full train         | vision only         | 58.5 | 55.4 |\n| val_unseen | full train         | vision only         | 47.1 | 43.4 |\n| val_seen   | full train         | vision + language   | 61.4 | 57.1 |\n| val_unseen | full train         | vision + language   | 48.8 | 44.1 |\n", "category": "Method_Mechanics"}, {"question": "How does the performance of the LangNav model vary with the length of the navigation episode, and what empirical evidence supports this observation?", "answer": "The performance of the LangNav model decreases as the episode length increases. This is evidenced by the evaluation of the model fine-tuned on 2,000 GPT-4 synthetic trajectories and 100 real-world trajectories, which measured the success rate (SR) and navigation error (NE). The results, detailed in the table, show that longer episodes make the task more complex, leading to reduced navigation performance.\n\n| Trajectory Steps | 4    | 5    | 6    | 7    | Overall |\n|------------------|------|------|------|------|---------|\n| NE↓              | 3.56 | 5.98 | 6.88 | 8.25 | 6.99    |\n| SR↑              | 54.2 | 37.1 | 35.7 | 28.3 | 33.8    |\n", "category": "Experimental_Exposition"}, {"question": "How does the expressivity of observations and actions influence the size and structure of the prompt, and what is its impact on overall task performance?", "answer": "The method updates the history by expanding the prompt with the current output action, without any additional complexities. The key aspect is to retain all the history information from the beginning within the prompt. Experimental results comparing the single-step action accuracy between two different models: (1) the LLaMA-7B model fine-tuned on prompts with full history, and (2) the LLaMA-7B model fine-tuned on prompts with a maximum of 3 history steps. The evaluation was conducted on the R2R val unseen split. As can be seen from the table, having full history information is essential for the model’s accuracy.\n\nEval Env | History Information | Step Accuracy↑\n---------|---------------------|-----------------\nval_unseen | full | 70.5%\nval_unseen | 3 steps | 66.1%\n\n", "category": "Method_Mechanics"}, {"question": "What are the potential biases in the synthetic data generation process for the LangNav system, given that 'LangNav underperforms baselines in data-rich regimes', and how can these biases be addressed, particularly in the context of scaling up the generation of synthetic data?", "answer": "The primary bias in the current pipeline is scene bias, which arises from using only 10 real trajectories from a single scene to prompt the LLMs, resulting in limited diversity. \nFor example, here are four generated instructions from the same output of the LLM:\n\n1. Start from the main entrance door, pass the living room, and enter the kitchen on your right. Locate the refrigerator, then turn left and stop just before the dining table.\n2. Navigate from the couch in the living room, move towards the mantel, and then stop next to the fireplace. Avoid any furniture and obstacles on your path.\n3. Begin at the foot of the bed in the master bedroom. Walk forward and enter the attached bathroom. Once you're inside, stop next to the bathtub.\n4. Start in the family room, walk towards the TV, then turn right and pass the bookshelf. Stop when you reach the large bay window overlooking the garden.\n\nThe synthetic instructions provided above show that (1) the patterns of these instructions are similar, typically following a format like 'Start from place A, go past place B, stop at place C'; (2) the scenes are limited to living areas and a single floor. However, R2R tasks often require the agent to navigate across multiple floors and in various non-living areas.\n\nTo address this, one viable solution is to increase the diversity of the seed navigation instructions, which is more cost-effective than obtaining a full trajectory. An experiment was conducted using 1,000 navigation instructions from various R2R scenes to prompt GPT-4-turbo, generating 2,000 synthetic trajectories without scene constraints. Although these trajectories are not of the same quality as those generated by GPT-4, scaling with them outperforms the results from the 10,000-trajectory set.\n\n| Number of synthetic trajectories | Seed trajectories | LLM             | OSR↑ | SR↑  | SPL↑ |\n|----------------------------------|--------------------|------------------|------|------|------|\n| 2,000                            | 10                 | GPT-4            | 42.2 | 31.1 | 26.6 |\n| 2,000                            | 1,000              | GPT-4-turbo      | 42.9 | 24.9 | 19.6 |\n| 2,000 + 2,000                    | 10 + 1,000         | GPT-4 + GPT-4-turbo | 43.2 | 32.6 | 28.3 |\n| 10,000                           | 10                 | GPT-4            | 41.9 | 31.6 | 27.5 |\n\n", "category": "Method_Mechanics"}, {"question": "How does the quality of synthetic training trajectories compare to real-world training trajectories when both are used to finetune the LLaMA2-7B model under controlled scene constraints?", "answer": "An experiment was conducted in which 93 training trajectories were prepared for both real-world data and synthetic data, and the LLaMA2-7B model was fine-tuned. To control the scene constraint, all 93 real-world trajectories originate from the same scan used to sample the 10 trajectories for prompting LLMs. The evaluation results are shown in the table. The evaluation shows that, under the same constraints of training scene diversity, synthetic trajectories have comparable or even better quality than the real-world trajectories. It is inferred that the gain from the synthetic data stems from the more accurate language description of the scene and the higher abundance of object concepts, attributed to GPT-4’s hallucination.\n\n|  Eval Env  | Training data |   NE↓  |   SR↑  |  SPL↑  |\n|:----------:|:-------------:|:-----:|:-----:|:-----:|\n|  val_seen  |   real-world  | 10.40 | 14.89 | 10.60 |\n|  val_seen  |    synthetic  | **8.81**  | **16.85** | **13.54** |\n| val_unseen |   real-world  | 10.16 | 14.09 |  9.54 |\n| val_unseen |    synthetic  | **8.29**  | **17.41** | **14.59** |", "category": "Experimental_Exposition"}, {"question": "What are the core contributions of the paper, particularly in the context of using language as a perceptual representation for navigation in low-data regimes?", "answer": "The paper’s core contribution is not a model or method, as it utilizes existing methods/models; rather, its core contribution is to suggest that language itself can serve as a perceptual representation for navigation, especially in low-data regimes, which has been underexplored compared to uses of language in other embodied tasks. The paper conducts two experiments that exploit the use of language via (1) synthetic data generation and (2) sim-to-real transfer to improve VLN.\n\nMoreover, the paper conducted additional experiments where a traditional vision-based navigation model is augmented with language features. This represents a ‘new model’ insofar as existing VLN systems only rely on visual perception features. More specifically, the original RecBert uses ResNet-152 to extract visual features to represent images, while the paper’s extension additionally concatenates the language feature of the caption to form the image representation. The captions are the same as those used in LangNav, and the features are extracted by BERT-base. The RecBert and its extension are pre-trained and fine-tuned with 100-shot training trajectories and the full R2R train set. The results are listed in the tables below.\n\n| eval env   | train trajectories | perceptual features | SR↑   | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n|  val_seen  |          100         |           vision   only         | 19.9 | 18.6 |\n| val_unseen |          100         |           vision  only         | 19.0 | 17.4 |\n|  val_seen  |          100         |           vision + language       |**22.8**|**21.5**|\n| val_unseen |          100         |           vision + language        |**19.3**|**18.0**|\n\n| eval env   | train trajectories  | perceptual features | SR↑   | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n|  val_seen  |      full train      |            vision  only           | 58.5 | 55.4 |\n| val_unseen |      full train      |            vision  only           | 47.1 | 43.4 |\n|  val_seen  |      full train      |           vision + language        |**61.4**|**57.1**|\n| val_unseen |      full train      |           vision + language        |**48.8**|**44.1**|\n\nIt can be seen that the language features improve the performance in both few-shot and full learning cases, which indicates that language-based features can help the visual perception module for the VLN agent.", "category": "Method_Disambiguation"}, {"question": "How would incorporating an end-to-end learning setup for the vision-to-text module improve the performance of the VLN task?", "answer": "End-to-end learning for the vision-to-text module would be an interesting direction to explore. With these settings, the vision-to-text module can benefit from LM policy and learn to focus more on navigation-relevant details. To investigate how a better vision-to-text module can help improve the VLN performance, and interpret the failure reason, the authors randomly pick 10 R2R trajectories where the LangNav fails to choose the correct direction, as illustrated by the second example in Figure 4 of the paper. For the steps where the LangNav acts wrong, incorrect captions are manually corrected using visual analysis and the given instructions. Three examples are provided which demonstrate that, by merely correcting the incorrect captions, LangNav is able to change the decision from incorrect to correct. The authors applied this procedure to 10 selected R2R trajectories and discovered that the paper's LangNav was able to revise its decision correctly in **70%** of these trajectories by only rectifying incorrect captions. The results indicate that a navigation-oriented vision-to-text module is essential for VLN tasks. End-to-end learning is a promising way to learn such a model.", "category": "Method_Mechanics"}, {"question": "How does the proposed strategy in Section 3.3 compare to the DAgger algorithm in terms of making the model robust to deviations from the logged imitation policy?", "answer": "Alongside the strategy outlined in Section 3.3, the authors conducted an external experiment using the DAgger algorithm for imitation learning. This process involves utilizing a trained language model (LM) policy to infer trajectories within the training environment. These inferred trajectories are then combined with the existing training dataset to form a new training set. The foundational concepts of the DAgger algorithm and the strategy, which is similar to RLHF—using a trained model policy pi_{t} to infer trajectories from the training set, assigning rewards (simple goal-based as an example) to the inferred trajectories, keeping those with high reward (above some threshold), and training pi_{t+1} on the union of the existing training dataset and those with high reward—bear a strong resemblance. In the comparison table, it illustrates the performance differences among the baseline LM policy, the LM policy enhanced through DAgger, and the paper’s method designed to address deviations. The findings demonstrate that although DAgger substantially improves VLN performance, the approach developed in the paper, as described in Section 3.3, ultimately achieves superior results compared to the DAgger algorithm.\n\n| eval env   |     policy     |  OSR↑ |  SR↑  |  SPL↑ |\n|------------|:--------------:|:----:|:----:|:----:|\n| val_unseen | base LM policy | 43.3 | 38.6 | 36.9 |\n| val_unseen |     DAgger     | 47.2 | 40.7 | 37.8 |\n| val_unseen |      the paper      | **50.3** | **43.2** | **37.9** |", "category": "Method_Comparison"}, {"question": "What is the impact of using image captions and object labels on trajectory generation and VLN results, and what components does the history $H$ include in different case studies?", "answer": "The comparison of LangNav trained on the full R2R train set with and without object labels shows a marginal boost in performance when object labels are added. The results don't involve the random approach which is described in section 3.3. Using image captions alone is more efficient due to prompt length considerations. In case study 1 and Table 4, only image captions are used, and history $H$ includes only image captions. In case study 2, history $H$ includes both image captions and object labels.\n\n| eval env   | train trajectories  | observations | SR↑   | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n| val_unseen |      full train     |captions| 35.3 | 33.5 |\n| val_unseen |      full train      |captions+object labels| 36.4 | 34.6 |", "category": "Experimental_Exposition"}, {"question": "What are the quantitative results that support the claims regarding the strong priors, spatial consistency, and descriptive quality of the generated trajectories in Figure 3?", "answer": "The authors conduct an experiment where 93 training trajectories are prepared for both real-world data and synthetic data, and the LLaMA2-7B model is fine-tuned. In order to control the scene constraint, the 93 real-world trajectories all originate from the same scan used to sample the 10 trajectories for prompting LLMs. The evaluation results are shown in the table. The evaluation shows that, under the same constraints of the training scene diversity, synthetic trajectories have comparable or even better quality with the real-world trajectories. It is inferred that the gain of the synthetic data is from the more accurate language description of the scene and the higher abundance of the object concepts due to GPT-4’s hallucination.\n\n|  Eval Env  | Training data |   NE↓  |   SR↑  |  SPL↑  |\n|:----------:|:-------------:|:-----:|:-----:|:-----:|\n|  val_seen  |   real-world  | 10.40 | 14.89 | 10.60 |\n|  val_seen  |    synthetic   | **8.81**  | **16.85** | **13.54** |\n| val_unseen |   real-world  | 10.16 | 14.09 |  9.54 |\n| val_unseen |    synthetic   | **8.29**  | **17.41** | **14.59** |", "category": "Experimental_Exposition"}, {"question": "Why are the gains marginal when increasing the number of trajectories from 2,000 to 10,000, and is this due to a lack of diversity in the generated trajectories? How would using more diverse scenes and gold instructions affect the quality of the generated trajectories and the fine-tuned LangNav?", "answer": "It is inferred that the fast convergence of synthetic data is due to the use of only 10 real trajectories from a single scene to prompt LLMs. For example, the following are four generated instructions from the same LLM output:\n\n> *1. Start from the main entrance door, pass the living room, and enter the kitchen on your right. Locate the refrigerator, then turn left and stop just before the dining table.*\n> *2. Navigate from the couch in the living room, move towards the mantel, and then stop next to the fireplace. Avoid any furniture and obstacles on your path.*\n> *3. Begin at the foot of the bed in the master bedroom. Walk forward and enter the attached bathroom. Once you're inside, stop next to the bathtub.*\n> *4. Start in the family room, walk towards the TV, then turn right and pass the bookshelf. Stop when you reach the large bay window overlooking the garden.*\n\nThe above synthetic instructions show that: (a) patterns of the synthetic instructions are similar, which are like ‘Start from place A, go past place B, stop at place C,’ (b) scenes are limited to the living area and a single floor; however, the R2R tasks always require the agent to navigate across floors and in some non-living areas.\n\nThis indicates that the underperformance of the generated trajectories compared to real trajectories stems from scene diversity. To further investigate the influence of the scene diversity, the authors conduct an experiment where 1000 navigation instructions sampled from various R2R scenes are used to prompt GPT-4-turbo to generate 2000 synthetic trajectories. As shown in the table, the scaling results indicate that although the 2000 trajectories generated by GPT-4-turbo are not of the same quality as those generated by GPT-4, scaling up using these trajectories outperforms the results from the 10000-trajectory set.\n\n| Number of synthetic trajectories | Seed trajectories |         LLM         |  OSR↑ | SR↑   | SPL↑  |\n|--------------------------------|-----------------|-------------------|----|------|------|\n|               2,000              |         10        | GPT-4               | 42.2 | 31.1 | 26.6 |\n|               2,000              |         1,000        | GPT-4-turbo               | 42.9 | 24.9 | 19.6 |\n|           2,000 + 2,000          |     10 + 1,000    | GPT-4 + GPT-4-turbo | **43.2** | **32.6** | **28.3** |\n|              10,000              |         10        | GPT-4               | 41.9 | 31.6 | 27.5 |", "category": "Experimental_Analysis"}, {"question": "What experiments were conducted to address the domain gap between vision and language representations in vision-and-language navigation, and what were the results?", "answer": "Additional experiments were conducted by augmenting a traditional vision-based navigation model (RecBert) with language features. The original RecBert uses ResNet-152 for visual feature extraction, while the extension concatenates the language feature of the caption to be the image representation. The captions are the same as what are used in LangNav and the features are extracted by BERT-base. The model was pre-trained and fine-tuned with 100-shot training trajectories and the full R2R train set. Results showed that adding language features improved performance in both few-shot and full learning cases, indicating that language-based features can enhance visual perception for the VLN agent.\n\n| eval env   | train trajectories | perceptual features | SR↑   | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n|  val_seen  |          100         |           vision   only         | 19.9 | 18.6 |\n| val_unseen |          100         |           vision  only         | 19.0 | 17.4 |\n|  val_seen  |          100         |           vision + language       |**22.8**|**21.5**|\n| val_unseen |          100         |           vision + language        |**19.3**|**18.0**|\n\n| eval env   | train trajectories  | perceptual features | SR↑   | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n|  val_seen  |      full train      |            vision  only           | 58.5 | 55.4 |\n| val_unseen |      full train      |            vision  only           | 47.1 | 43.4 |\n|  val_seen  |      full train      |           vision + language        |**61.4**|**57.1**|\n| val_unseen |      full train      |           vision + language        |**48.8**|**44.1**|", "category": "Experimental_Exposition"}, {"question": "Does the integration of language as a perceptual representation improve the performance of LangNav, given the concerns that (1) continuous visual features are likely to be almost always available, and (2) language-based representations could be incomplete and ambiguous?", "answer": "To address these concerns, the authors conducted an additional study where both vision and language are used as a perceptual representation by augmenting RecBERT to take text representations (given by a pretrained BERT-base model). Under this setup, using language as a perceptual representation improves performance in both low-data and full-data regimes, as shown below.\n\n| eval env   | train trajectories | perceptual features | SR↑  | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n|  val_seen  |          100         |           vision   only         | 19.9 | 18.6 |\n| val_unseen |          100         |           vision  only         | 19.0 | 17.4 |\n|  val_seen  |          100         |           vision + language       |**22.8**|**21.5**|\n| val_unseen |          100         |           vision + language        |**19.3**|**18.0**|\n\n| eval env   | train trajectories  | perceptual features | SR↑   | SPL↑  |\n|------------|:--------------------:|:----------------------:|------|------|\n|  val_seen  |      full train      |            vision  only           | 58.5 | 55.4 |\n| val_unseen |      full train      |            vision  only           | 47.1 | 43.4 |\n|  val_seen  |      full train      |           vision + language        |**61.4**|**57.1**|\n| val_unseen |      full train      |           vision + language        |**48.8**|**44.1**|\n\nThese results indicate that language as a perceptual representation can provide additional benefits on top of continuous visual features, even in non-low-data settings.", "category": "Experimental_Analysis"}, {"question": "Was numerical analysis provided to support the three claims of the quality of all generated trajectories (Strong Prior, Spatial Consistency, and Descriptive)?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the experiment comparing real-world and synthetic training trajectories show that synthetic trajectories have comparable or better quality under the same scene constraints?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the additional study show that using both vision and language as perceptual features improves performance in both low-data and full-data regimes?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "BrwIZVSc7b", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D Understanding, Generation, and Instruction Following", "authors": ["Ziyu Guo", "Renrui Zhang", "Xiangyang Zhu", "Yiwen Tang", "Xianzheng Ma", "Jiaming Han", "Aojun Zhou", "Kexin Chen", "Peng Gao", "Xianzhi Li", "Hongsheng Li", "Pheng-Ann Heng"], "abstract": "With the growing diversity of large-scale data, learning from multi-modality has attained notable progress in language and 2D vision. However, in 3D domains, how to develop an all-purpose multi-modal framework is still under-explored. To this end, we introduce Point-Bind, a 3D multi-modality model aligning point clouds with 2D image, language, and audio. Guided by ImageBind, we construct a joint embedding space between 3D and multi-modalities, enabling many promising applications, e.g., any-to-3D generation, 3D embedding arithmetic, and 3D open-world understanding. On top of this joint embedding space, we further present Point-LLM, a 3D large language model (LLM) following 3D and multi-modal instructions. Without any 3D instruction data, our Point-LLM injects the semantics of Point-Bind into pre-trained LLMs, e.g., LLaMA, and exhibits superior 3D and multi-modal question-answering capacity. We have conducted extensive experiments to demonstrate the effectiveness and generalizability of our approach for aligning 3D and multi-modality.", "keywords": "3D point cloud learning;multi-modality learning;large language model", "qa_pairs": [{"question": "What steps were taken to conduct quantitative experiments for evaluating the multi-modal arithmetic ability of Point-Bind, and what were the results?", "answer": "The authors constructed a new multi-modal arithmetic retrieval benchmark, containing 200 multi-modal data pairs selected from the ModelNet40, ESC-50, and ImageNet datasets. These data pairs fall into three categories: 3D-Text-Image, 3D-Audio-Image, and 3D-Image-Image. The authors use the combination of the first two modalities to retrieve the third. The retrieval patterns must be realistic, for example: Bottle (3D) + ‘Car’ (Text) -> Bottles in the car (Image), Tent (3D) + Water (Audio) -> Tents beside the river (Image), Person (3D) + Motor (Image) -> Human on the motorbike (Image). Point-Bind was evaluated on this benchmark based on feature similarity rankings. As reported in the table below, pre-training with more modalities achieves better performance, leading to superior multi-modal alignment and cross-modal understanding.\n\n|Text                 | 3D   | Image | Audio |    Acc. |\n| ---- | ----- | ----- | -------- | -------------- | \n| ✓  | ✓ | -    | -     | 44.5 | \n| ✓  | ✓ | ✓ | -    | 53.1% | \n| ✓  | ✓ | ✓ | ✓ | 56.4% |", "category": "Experimental_Exposition"}, {"question": "Are there examples of scene understanding, particularly for indoor scenes, provided in the study?", "answer": "Yes, qualitative examples of indoor scene understanding are provided in Figure 13 of the Appendix, focusing on the ScanNet[1] dataset. A scene-level variant of the model, Point-LLM$_{Scene}$, was developed by fine-tuning the object-level Point-LLM using a 3D question-answering dataset[2] constructed from ScanRefer[3]. Three MLP layers with residual connections were added between Point-Bind’s 3D encoder and the LLM to enable scene-level 3D geometry learning. Only the new MLP layers were made trainable, while keeping other components frozen to preserve the pre-trained cross-modal knowledge.\n\n[1]ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes. CVPR 2017.\n\n[2]Chat-3D: Data-efficiently Tuning Large Language Model for Universal Dialogue of 3D Scenes. arxiv 2023.\n\n[3]Scanrefer: 3d object localization in rgb-d scans using natural language. ECCV 2020.", "category": "Experimental_Exposition"}, {"question": "What are the key methodological differences between Point-LLM and existing BLIP-2[1]/MiniGPT-4[2], particularly regarding the handling of modalities and alignment processes?\n\n1] BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. arxiv 2023.\n[2] Minigpt-4: Enhancing vision-language understanding with advanced large language models. arxiv 2023", "answer": "The methodology of Point-LLM is very distinct from BLIP-2 [1] or MiniGPT-4 [2]. Point-LLM doesn't simply inject other modalities into LLMs, but proposes a more efficient paradigm by the joint embedding space of Point-Bind. Besides the different 2D and 3D modalities, the paper's novelty compared to existing BLIP-2 or MiniGPT-4 is three-fold as follows.\n\n1. Solve 3D-language alignment by converting it into easier 2D-language alignment. Considering the paucity of accessible 3D-caption data, it's hard to directly align the 3D modality with LLMs like existing 2D methods using extensive image-caption data. Fortunately, Point-Bind presents a joint embedding space between 3D and multi-modality. This allows us to still align LLMs with 2D using ImageBind's image encoder by sufficient image-caption data, and then replace the 2D encoder with Point-Bind's 3D encoder. A visual cache model is specifically adopted to alleviate the 2D-3D domain gap. By regarding 2D as an intermediary, the paper's cross-modal alternative solution can efficiently bridge the gap between 3D and LLMs, while overcoming the paucity issue of 3D-caption data, which fully demonstrates the emergent advantage of Point-Bind.\n\n2. Do not need 3D instruction data collection and tuning. Existing 2D visual LLMs normally conduct a specific instruction-tuning process after the 2D-LLM alignment, which requires collecting high-quality 2D QA pairs. For example, LLaVA [3] collects images with object-level annotations from MSCOCO [4] and generates an instruction dataset using GPT-4. Such data collection and tuning processes are expensive and time-consuming, especially in the complex 3D domain. Fortunately, also by the joint embedding space of Point-Bind, the paper still conducts the instruction tuning of Point-LLM in 2D space using the already collected 2D QA data. This contributes to superior data and computation efficiency.\n\n3. Process 3D-centric multi-modality instructions within LLMs. enabled by Point-Bind, Point-LLM can accept instructions of both 3D and other multi-modality input, as shown in Figure 3 of the paper. BLIP-2 and MiniGPT-4 can only tune LLMs to understand 2D images, but the paper's joint embedding space enables Point-LLM to directly understand multi-modality instructions and reason their relationship without specific multi-modality QA training, e.g., responding to a question based on a point cloud and a picture or audio.\n\nTherefore, instead of directly transferring the training pipeline of 2D visual LLMs into 3D, the paper's Point-LLM proposes an alternative solution emergent from Point-Bind’s joint embedding space, which is both efficient and effective for 3D domains.\n\n[3] Visual Instruction Tuning. NeurIPS 2023.\n\n[4] Microsoft COCO: Common objects in context. ECCV 2014.", "category": "Method_Comparison"}, {"question": "What are the main contributions of Point-Bind, distinguishing it from previous works like ULIP that also use the contrastive paradigm?", "answer": "The paper is inspired by previous works like ULIP to adopt the contrastive paradigm for pre-training, but the main contributions of Point-Bind are in two aspects as follows.\n\n1. A 3D multi-modal embedding space. For the first time, the paper aligns 3D with more than two modalities into a joint embedding space. Previous works only verify that 3D can be connected with 2D images and language from CLIP, while Point-Bind indicates the potential for more modalities including audio, video, depth, and infrared data. As shown in Figure 8 of the Appendix, the paper adds the visualization of cross-modal retrieval between 3D and three new modalities (video, depth, and infrared data). This unified space is expected to expand the scope of 3D models to wider cross-modal scenarios.\n\n2. Emergent new applications without specific training. Benefiting from the joint space, Point-Bind naturally motivates several new cross-modal applications, which do not need domain-specific training. For example, previously, to achieve question answering (QA) with 3D objects, one had to collect large-scale 3D data with paired QA and spend many computational resources for training. Instead, with Point-Bind, the 3D QA capability is totally emergent from the joint embedding space given an off-the-shelf 2D QA model, avoiding the expensive data collection and training. Likewise, specific 3D-audio training is not needed to achieve audio-referred 3D zero-shot classification, or specific 3D-video training for video-to-3D retrieval. Such emergent abilities of Point-Bind significantly lower the bar for efficiently achieving many new cross-modal applications.\n\nTherefore, Point-Bind, as the first work, develops a joint embedding space for 3D multi-modality and enables many emergent applications without domain-specific training. Besides, the paper also constructs a 3D-image-audio-text paired dataset and introduces a triplet contrastive loss for pre-training, which can be viewed as a minor contribution.", "category": "Method_Comparison"}, {"question": "What is the role of Point-Bind in the 'Any-to-3D Generation' process, and how does it complement the capabilities of ImageBind?", "answer": "Within the joint 3D embedding space, the basic text/image/audio-to-mesh generation is mainly powered by encoders of ImageBind. However, by the 3D encoder of Point-Bind, point-to-mesh generation can be achieved, and it further enables 3D editing with multi-modal instructions, as visualized in Figure 10 of Appendix.\n\nFor example, given a 3D airplane, a language instruction, ‘Color the 3D shape in red,’ or a pure yellow picture as the visual instruction can be provided. Then, they are respectively fed into Point-Bind’s 3D encoder and ImageBind’s text or image encoder. Due to the joint embedding space, the generative decoder can incorporate their semantics and output the airplane in red/yellow. Likewise, given an ordinary 3D bench, instructions like ‘Modify the material to wooden.’ can be provided. The model can correspondingly generate a wooden chair.\n\nTherefore, Point-Bind plays an important role in ‘Any-to-3D Generation’ by encoding input point clouds into the joint multi-modal embedding space.", "category": "Method_Mechanics"}, {"question": "How does Point-Bind contribute to 3D open-world understanding and what specific tasks does it address in this domain?", "answer": "Point-Bind contributes to 3D open-world understanding by eliminating the need for additional task-specific training. It achieves the best performance in traditional text-referred classification and explores a new task, audio-referred open-world understanding, setting a baseline for future 3D open-world learning.", "category": "Method_Mechanics"}, {"question": "How did the incorporation of JM3D and CG3D techniques into the Point-Bind framework enhance its performance on the benchmarks of 3D zero-shot classification and cross-modal retrieval on ModelNet40?", "answer": "1. **JM3D** two innovative approaches to enhance the multi-modal pre-training of 3D models: the Structured Multimodal Organizer (SMO) and the Joint Multi-modal Alignment (JMA). SMO utilizes multi-view rendered images and hierarchical text to generate more comprehensive representations, while JMA aims to achieve better multi-modal synergy by generating joint vision-language features. The authors integrate these two techniques into Point-Bind, which handles the image and text modalities within ImageBind, and evaluate their performance on two benchmarks: 3D zero-shot classification and cross-modal retrieval on ModelNet40. As shown in the table below, integrating SMO and JMA significantly enhances the capabilities of Point-Bind, highlighting the importance of more comprehensive vision-language guidance.\n\n| Method         | Zero-shot Cls.| 3D → 3D  | 2D → 3D  | 3D → 2D  | Text → 3D |\n| -------------- | ---- | -------- | -------- | -------- | --------- |\n| Point-Bind | 78.0 | 63.2 | 34.6 | 42.8 | 64.5 |\n| Point-Bind w JM3D | **78.4** | **64.1** | **35.5** | **43.9** | **64.7**|\n\n2. **CG3D** shares a similar contrastive learning paradigm with ULIP, and introduces learnable visual prompts for CLIP's image encoder for better adaptation of 2D rendered images. For Point-Bind, the authors add learnable visual prompts to the image encoder of ImageBind, and report the results in the table below. On both benchmarks, the prompting approach from CG3D can improve the performance of Point-Bind, which demonstrates the effectiveness of fine-tuning the pre-trained image embeddings.\n\n| Method         | Zero-shot Cls.| 3D → 3D  | 2D → 3D  | 3D → 2D  | Text → 3D |\n| -------------- | ---- | -------- | -------- | -------- | --------- |\n| Point-Bind | 78.0 | 63.2 | **34.6** | 42.8 | 64.5 |\n| Point-Bind w CG3D | **78.2** | **63.5** | 34.3 | **43.2** | **64.8**|", "category": "Experimental_Analysis"}, {"question": "How many categories from ShapeNet are used in the final training data, and how are the audio and non-audio categories handled?", "answer": "The full 55 categories from ShapeNet are used for pre-training. Out of these, 9 categories contain paired audio data from ESC-50, while for the remaining 45 categories, the contrastive loss is calculated using only the image and text encoders of ImageBind, without the audio modality.", "category": "Experimental_Setup"}, {"question": "How does Point-Bind demonstrate its efficacy and generalization ability for downstream 3D tasks, particularly in 3D shape classification on ModelNet40 and ScanObjectNN datasets?", "answer": "Point-Bind, regarded as a pre-trained model, was fine-tuned for 3D shape classification on ModelNet40 and ScanObjectNN datasets without voting. As shown in the table below, compared to Point-BERT and I2P-MAE, Point-Bind's multi-modal learning enhanced classification accuracy. This indicates its promising generalization ability for assisting downstream 3D applications.\n\n| Method       | ModelNet40 | ScanObjectNN |\n| -------------- | ---- | ------- |\n| Point-BERT  | 92.7 | 83.1 |\n| Point-Bind | 93.4 |84.8 |\n\n| Method       | ModelNet40 | ScanObjectNN |\n| -------------- | ---- | ------- |\n| I2P-MAE | 93.7 | 90.1 |\n| Point-Bind | 93.9   | 91.2   |", "category": "Experimental_Analysis"}, {"question": "How does Point-LLM differ from a simple utilization of ImageBind-LLM with visual cache, and what modifications were made to extend it for scene-level understanding in 3D domains?", "answer": "Point-LLM is built upon a pre-trained ImageBind-LLM by incorporating Point-Bind into it, emphasizing that it is not a new separate model but an emergent application without 3D instruction tuning. The original Point-LLM is designed for object-level question answering because Point-Bind was only pre-trained on single-object datasets. The authors further implemented a scene-level variant and showed qualitative examples from the ScanNet dataset in Figure 13 of the Appendix. Specifically, to extend it for scene-level understanding, the authors fine-tuned the object-level Point-LLM using a 3D question-answering dataset from ScanRefer. They appended three MLP layers with residual connections after Point-Bind’s 3D encoder to learn scene-level 3D geometries, only enabling the new MLP layers to be trainable while keeping other components frozen to preserve the pre-trained cross-modal knowledge. This experiment indicates that, although Point-LLM is extended from the existing ImageBind-LLM, the paper's approach can be utilized as a base model to be fine-tuned for more complex 3D scenarios, demonstrating promising generalizability in 3D domains.", "category": "Method_Comparison"}, {"question": "What steps were taken to ensure a fair comparison between Point-LLM, ImageBind-LLM, and PointLLM (Xu et al., 2023)[1] in terms of input data and training strategies?\n\n[1] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023b.", "answer": "To ensure a fairer evaluation, the authors compare the paper’s Point-LLM with the multi-view ImageBind-LLM and PointLLM (Xu et al., 2023) as follows. 1. ImageBind-LLM cannot take raw point clouds as input, so the authors try their best to transform them into a data form that ImageBind-LLM can process, i.e., rendered 2D images. Since the paper’s Point-LLM uses a 3D encoder, it cannot encode 2D images, the authors feed multi-view rendered images into ImageBind-LLM for GPT-4 evaluation. As shown in the table below, as the number of views increases, ImageBind-LLM achieves better 3D captioning capability by understanding more comprehensive 3D geometries. However, the paper’s Point-LLM still outperforms the others by directly encoding point clouds in 3D space with Point-Bind, indicating the superiority of the paper’s approach in 3D instruction following. \n\n| View Number | 1    | 2    | 4    | 8    |\n|-------------|------|------|------|------|\n| Tie         | 1.6% | 3.1% | 5.7% | 5.1% |\n| ImageBind-LLM Wins | 34.0% | 35.1% | 35.6% | 39.7% |\n| Point-LLM Wins    | 64.4% | 61.8% | 58.7% | 55.2% |\n\n2. PointLLM (Xu et al., 2023) requires a two-stage training strategy to inject 3D semantics into LLMs. After the pre-training stage for 3D-language alignment, PointLLM (Xu et al., 2023) collects a high-quality 3D instruction dataset for fine-tuning. To make a fair comparison, the authors also fine-tune the paper’s Point-LLM with the same 3D instruction dataset and present the comparison results in the table below, where the paper’s model demonstrates better 3D captioning accuracy. This demonstrates that the paper’s multi-modal embedding space of Point-Bind can enhance the 3D understanding performance of LLMs.\n\n| Method                     | Rate  |\n|----------------------------|-------|\n| PointLLM (Xu et al., 2023) Win | 44.8% |\n| Point-LLM Win              | 47.9% |\n| Tie                        | 7.3%  |\n", "category": "Method_Comparison"}, {"question": "How do the pre-training settings and image encoders of Point-Bind and ULIP differ, and what impact does this have on their performance in zero-shot classification on ModelNet40?", "answer": "Point-Bind’s teacher model, ImageBind, uses CLIP’s ViT-H image encoder and is trained with billions of image-text pairs, whereas ULIP’s teacher model, SLIP, uses ViT-L and is trained with 15 million image-text pairs. The authors reproduced a ULIP model using CLIP’s ViT-H image encoder, which is the same as ImageBind’s (ImageBind freezes CLIP’s image and text encoders for training other modalities). As shown in the table below, for zero-shot classification on ModelNet40, although the reproduced ULIP’s performance can be improved by the ViT-H image encoder, the proposed approach (Point-Bind) still performs better via a joint multi-modal embedding space.\n\n Method | Image Encoder | Accuracy \n|---------|---------|------------|\n| ULIP | ViT-L | 60.4 |\n| ULIP | ViT-H | 73.2 |\n| Point-Bind | ViT-H | 78.0 |\n", "category": "Method_Comparison"}, {"question": "What are the nine categories of audio clips obtained for the 3D-audio Pairs in Section 2.2, and how many audio samples are included in each category?", "answer": "The nine categories of audio clips sampled from ESC-50 are 'airplane', 'chirping_birds', 'can_opening', 'car_horn', 'clock_tick', 'keyboard_typing', 'crackling_fire', 'train', and 'washing_machine'. Each category contains 40 audio samples, totaling 360 samples. During training, a random audio sample from these categories is paired with a point cloud, and data augmentation techniques like random cropping and volume perturbation are applied for robust training.", "category": "Experimental_Setup"}, {"question": "Why does the Text+3D modality perform poorly in Table 4, and what is the impact of including the image modality on performance? Additionally, are there results available for the Image+3D combination?", "answer": "This is mainly because of the pre-training data scale. ImageBind only needs paired data between images and another modality, as there are sufficient image-paired data to ensure robust cross-modal alignment, for example, 2M for image-audio data and 510K for image-IMU data. In contrast, the 3D-paired dataset is much smaller in scale, i.e., 52.5K from ShapeNet. Therefore, using all four modalities (Text+Image+3D+Audio) provides more contrastive supervision signals, which can partially alleviate the influence of data scale. Results for the Image+3D combination show that incorporating more modalities for pre-training yields better performance.\n\n| Text | 3D | Image | Audio | Acc.  |\n| :---: | :---: | :---: | :---: | ---: |\n| ✓     | ✓     | -     | -     | 70.43 |\n| -     | ✓     | ✓     | -     | 68.72 |\n| ✓     | ✓     | ✓     | -     | 76.96 |\n| ✓     | ✓     | ✓     | ✓     | 78.00 |", "category": "Experimental_Analysis"}, {"question": "Does the final training data include only 9 categories from the public datasets, excluding other categories?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "2vAhX71UCL", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Dreamix: Video Diffusion Models are General Video Editors", "authors": ["Eyal Molad", "Eliahu Horwitz", "Dani Valevski", "Alex Rav-Acha", "Yossi Matias", "Yael Pritch", "Yaniv Leviathan", "Yedid Hoshen"], "abstract": "Text-driven image and video diffusion models have recently achieved unprecedented generation realism. While diffusion models have been successfully applied for image editing, none can edit motion in video. We present the first diffusion-based method that is able to perform text-based motion and appearance editing of general, real-world videos. Our approach uses a video diffusion model to combine, at inference time, the low-resolution spatio-temporal information from the original video with new, high resolution information that it synthesized to align with the guiding text prompt. As maintaining high-fidelity to the original video requires retaining some of its high-resolution information, we add a preliminary stage of finetuning the model on the original video, significantly boosting fidelity. We propose to improve motion editability by using a mixed objective that jointly finetunes with full temporal attention and with temporal attention masking. We extend our method for animating images, bringing them to life by adding motion to existing or new objects, and camera movements. Extensive experiments showcase our method's remarkable ability to edit motion in videos.", "keywords": "Video Diffusion;Video Motion Editing", "qa_pairs": [{"question": "Does the proposed method apply to other video diffusion models like ModelScope, and how do the results compare with those obtained using the primary VDM base model?", "answer": "The proposed method is applicable to other video diffusion models such as ModelScope. Preliminary experiments show that applying the method to the ModelScope backbone results in improved temporal consistency and editing capabilities compared to image-based methods, though it does not match the performance of the larger VDM backbone used in the main study.", "category": "Experimental_Analysis"}, {"question": "Why is the comparison between Tune-A-Video and Dreamix considered unfair, and what additional comparison has been conducted to address this concern?", "answer": "The comparison is considered unfair because Tune-A-Video is finetuned on an image diffusion model, while Dreamix is finetuned on a video diffusion model. To address this concern, the authors compared their method to the original version of ModelScope, which reinforced their claims of improvement over the original ModelScope.", "category": "Method_Comparison"}, {"question": "For text-based video editing, FateZero [ref-1] and TokenFlow [ref-2] are more advanced methods designed for video editing. For subject-driven video generation, Animatediff [ref-3] can learn a Lora model from several images and generate corresponding videos. For animating a single image, VideoComposer [ref-4] can generate video conditioned on the single image and text, which does not require additional transformation. How do these advanced methods compare with the paper's approach in terms of motion editing capabilities?\n\n[ref-1] Fatezero: Fusing attentions for zero-shot text-based video editing.\n\n[ref-2] TokenFlow: Consistent Diffusion Features for Consistent Video Editing.\n\n[ref-3] Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.\n\n[ref-4] VideoComposer: Compositional Video Synthesis with Motion Controllability.", "answer": "The authors conducted preliminary experiments on these methods, and while they excel in appearance editing with impressive temporal consistency, they are unable to perform motion edits, which is the core of the paper’s work. The authors tested AnimateDiff on subject-driven video generation, but the results were not satisfactory. AnimateDiff, designed for small motion generation, faced challenges with the large motions in the paper's examples (e.g., walking and drinking beer). The authors tested VideoComposer. In addition, the authors tested Gen-2, and PikaLabs (both commercial products). All methods failed to add significant motion or add objects to the input image and animate them.", "category": "Method_Comparison"}, {"question": "How does the proposed method perform when applied to an open-source video generation model like ModelScope, particularly in terms of temporal consistency and editing capabilities?", "answer": "Preliminary experiments show that applying the proposed method to the ModelScope backbone results in improved temporal consistency and editing capabilities compared to image-based methods, though it does not match the performance of the larger VDM backbone used in the paper.", "category": "Experimental_Exposition"}, {"question": "How does the method handle consistency between the input image and the first frame, and how does its quality compare to recent methods like AnimateDiff?", "answer": "The method experiences some resolution degradation between the input image and the generated results, but it uniquely adds motion and animates the scene, which no other current method can do. AnimateDiff is not applicable here as it requires a 'Dreambooth' model input, needing multiple images instead of one. Comparative analysis with VideoComposer, Gen-2, and PikaLabs shows that these methods also fail to add large motion or animate added objects.", "category": "Method_Comparison"}, {"question": "Why are the appearance features and details, such as the toy in Figure 2 and the bear in Figure 6, not fully maintained in the subject-driven video generation?", "answer": "Some fine details are not fully preserved due to the limitations of the method. A comparison with AnimateDiff, which is tailored for small motion generation, did not yield satisfactory results because AnimateDiff struggles with larger motions, such as those in the examples (e.g., walking and drinking beer).", "category": "Experimental_Analysis"}, {"question": "Did the application of the method to the ModelScope backbone show improved temporal consistency and editing capabilities compared to image-based methods?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the method's effectiveness on video generation tasks rely solely on the specific video generation model used, without testing on open-source alternatives?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the performance of the method on ModelScope match the performance of the larger VDM backbone used in the main experiments?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were preliminary experiments conducted to verify the method's performance on an open-source video generation model like ModelScope?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the ModelScope backbone in preliminary experiments match the performance power of the larger VDM backbone used in the paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the results from applying the method to ModelScope show improvements similar to those observed with the primary VDM base model used in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were preliminary experiments conducted to test the proposed method on other video diffusion models like ModelScope?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "PhJUd3mbhP", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "AutoAgents: A Framework for Automatic Agent Generation", "authors": ["Guangyao Chen", "Siwei Dong", "Yu Shu", "Ge Zhang", "Jaward Sesay", "Börje F. Karlsson", "Jie Fu", "Yemin Shi"], "abstract": "Large language models (LLMs) have enabled remarkable advances in automated task-solving with multi-agent systems. \nHowever, most existing LLM-based multi-agent approaches rely on predefined agents to handle simple tasks, limiting the adaptability of multi-agent collaboration to different scenarios.\nTherefore, we introduce AutoAgents, an innovative framework that adaptively generates and coordinates multiple specialized agents to build an AI team according to different tasks.\nSpecifically, AutoAgents couples the relationship between tasks and roles by dynamically generating multiple required agents based on task content and planning solutions for the current task based on the generated expert agents.\nMultiple specialized agents collaborate with each other to efficiently accomplish tasks. \nConcurrently, an observer role is incorporated into the framework to reflect on the designated plans and agents' responses and improve upon them.\nOur experiments on various benchmarks demonstrate that AutoAgents generates more coherent and accurate solutions than the existing multi-agent methods.\nThis underscores the significance of assigning different roles to different tasks and of team cooperation, offering new perspectives for tackling complex tasks.", "keywords": "Multi-Agent Framework;Large Language Models;Agent Generation", "qa_pairs": [{"question": "How are the roles, skills, and actions for specific tasks determined in the framework? Are they injected in the prompt or generated by the Planner agent?", "answer": "The roles, skills, and actions for specific tasks are determined by the Planner agent, which serves as the architect in the framework. This includes not only the creation of prompts, descriptions, and toolsets for each role but also the formulation of an intricate execution plan. This plan distinctly specifies whether self-refinement or collaborative refinement actions are required. Following this initial phase, both the list of roles and the execution plan are subject to further refinement, achieved through collaborative dialogues with Observer agents. It is imperative to note that the choice of specific tools and techniques employed by each agent to navigate the steps in the execution plan is made independently by the agents themselves. A crucial aspect of the paper’s methodology is that unique prompts are not created for each task. Rather, the primary prompts are methodically generated by the Planner agent, ensuring a consistent and streamlined approach across various tasks.", "category": "Method_Mechanics"}, {"question": "How is the decision-making process for including thoughts in the self-refinement process structured, and what is the role of the custom agent in this process?", "answer": "The custom agent, upon receiving an assigned task, analyzes and understands the task, plans future steps, and determines immediate actions. The agent outputs the results of completing the current step, and in subsequent feedback rounds, the history of executed steps and their outcomes (short-term memory) is fed back into the prompt under 'completed steps and responses' to guide future steps. ", "category": "Method_Mechanics"}, {"question": "What are the specific agent types compatible with the three modes of knowledge sharing (short-term memory, long-term memory, and dynamic memory), and how were they utilized in the experiments?", "answer": "The specific agent types that are compatible with the three modes of knowledge sharing as depicted in Figure 8 of the paper:\n\n- __Short-term Memory__: This is predominantly utilized by agents that emerge from self-refinement and collaborative-refinement processes. These agents leverage short-term memory to retain historical records throughout the refinement phase. Notably, every agent generated within the paper's framework actively employs short-term memory.\n\n- __Long-term Memory__: This component archives the execution results of each agent, primarily chronicling the cumulative progress of the execution plan. Access to long-term memory is exclusive to a pre-designated Action Observer. The Action Observer utilizes this memory to evaluate whether additional execution steps are necessary or if there's a need to generate dynamic memory content.\n\n- __Dynamic Memory__: This memory type is essential for distilling and summarizing pivotal information from long-term memory for diverse agents. Generated by the Action Observer, dynamic memory serves to amalgamate critical aspects of the overall execution plan, thereby facilitating the agents responsible for implementing subsequent tasks. Consequently, the dynamic memory content available to each generated agent may differ, contingent upon their specific roles and tasks.\n\nDue to the fundamental nature of Short-term Memory and Long-term Memory as indispensable components within the system, the paper's investigation was exclusively focused on understanding and analyzing the role of dynamic memory. The paper has included an ablation study of dynamic memory in the Appendix A. The following table shows the results without the dynamic memory. It is observed that the performance of AutoAgents decreases by 1% without the summarization from dynamic memory, which also proves the importance of dynamic memory.\n\n| Method | N (# triva questions) = 5|\n| --- | --- |\n| AutoAgents w/o dynamic memory | 89.0  |\n| AutoAgents | 90.0 |", "category": "Method_Disambiguation"}, {"question": "What was the rationale behind setting the number of discussions to 3 for collaborative planning and 5 for single-agent execution in the two stages?", "answer": "The number of discussions was set to 3 for collaborative planning because it has been observed that a detailed plan and a complete list of agents can generally be obtained after about three rounds. For single-agent execution, it was set to 5 because there is a high probability that the task can be completed within five rounds. Additionally, while more iterations typically lead to better outcomes, the setup aims to manage and control costs associated with GPT-4 usage.", "category": "Experimental_Analysis"}, {"question": "What are the specific details of the volunteer recruitment process for the open-ended question answering evaluation in the paper?", "answer": "For HumanEval, three independent volunteers were enlisted to evaluate two sets of responses—one generated by AutoAgents and the other by a different model—based on criteria such as helpfulness, reliability, accuracy, and comprehensiveness. The volunteers were blinded to the identity of the model that produced each response to ensure an unbiased assessment. They were instructed to adhere to these standards:\n\n```\nWe would like to request your feedback on the response to the user question\ndisplayed above. Please rate the helpfulness, relevance, accuracy, level of \ndetails of their responses. \n\nEach response receives an overall score on a scale of 1 to 10, where a higher \nscore indicates better overall performance. Please first provide a \ncomprehensive explanation of your evaluation, avoiding any potential bias and \nensuring that the order in which the responses were presented does not affect \nyour judgment. \n\nOutput with the following format:\nEvaluation evidence: <your evaluation explanation here>\nScore: <score>\n```", "category": "Experimental_Setup"}, {"question": "How does the AutoAgents framework expand the scope of collaborative applications and reduce resource consumption through its automated agent generation and collaborative refinement processes?", "answer": "Existing agent technologies are overly dependent on pre-defined prompts. This reliance poses a substantial user barrier and limits their practicality for broader applications. AutoAgents addresses this by automating the creation of new agent roles, which streamlines both the definition of roles and the execution of tasks, thereby enhancing the system's versatility. AutoAgents significantly extends its adaptability to diverse applications through the automatic generation of multiple intelligent agents, each tailored for specific tasks. This improvement in agent generation quality is achieved through the collaborative efforts of three predefined agents—the Planner, Agent Observer, and Plan Observer. These agents work together to discuss, refine, and finalize the list of agents and their corresponding execution plans. The paper delves deeper into this aspect with ablation experiments focusing on the Agent Observer and Plan Observer in Appendix A. The comparative analysis, as shown in the following table, reveals a marked decrease in AutoAgents' overall performance in the absence of cooperative dialogues among Observers. This outcome not only underscores the efficacy of collaborative refinement but also highlights the critical role of these discussions in enhancing task adaptability and resource optimization in the AutoAgents framework.\n\n| Method                     | N (# trivia questions) = 5 |\n|----------------------------|-----------------------------|\n| AutoAgents w/o observers   | 87.0                        |\n| AutoAgents                 | 90.0                        |\n", "category": "Method_Mechanics"}, {"question": "How is the number of agents determined during the automatic agent generation phase in the AutoAgents framework?", "answer": "AutoAgents fundamentally predefines only a Planner and three types of Observers. Throughout the automatic agent generation phase, neither the number nor the types of agents to be generated are pre-specified. Instead, the decision regarding both the number and types of agents is dynamically and autonomously determined through collaborative discussions among the Planner, Agent Observer, and Plan Observer. This process ensures that the agents generated are optimally aligned with the specific requirements and nuances of each task.", "category": "Method_Mechanics"}, {"question": "How does the AutoAgents framework differ from traditional frameworks in terms of adaptability and flexibility in agent generation?", "answer": " Traditional frameworks in agent generation typically hinge on the capability of a single agent to determine the list of agents. This approach, however, often results in limited adaptability across various tasks. In contrast, the AutoAgents framework places significant emphasis on the drafting stage for the auto-generation of agents. Here, the compilation of the agent list is a result of collaborative discussions between the Planner agent and the two Observers, enhancing adaptability and flexibility. Additionally, the prompts utilized in AutoAgents are designed with a universal approach, reflecting their ability to adapt to a wide range of tasks without the need for task-specific customization. While platforms like AgentVerse and SSP have opted for task-specific modifications, AutoAgents adopts a more holistic, unified prompt strategy. This uniform approach is adept at catering to diverse tasks, which is evident in the paper's system's notable performance in open-ended question answering and trivia creative writing tasks. Such results highlight the extensive applicability and adaptability of the paper's prompt design, making AutoAgents a versatile tool in the realm of agent-based systems.", "category": "Method_Comparison"}, {"question": "How does the inclusion of AgentVerse in the Table 2 demonstrate the superiority of AutoAgents in open-ended question answering tasks?", "answer": "The Table 2 includes results for AgentVerse, evaluated using both FairEval and HumanEval. The results show that AutoAgents outperforms AgentVerse in addressing a majority of open-ended questions. This superiority is attributed to AutoAgents' cooperative discussion-based agent generation, self-refinement, and collaborative refinement mechanisms, which collectively enhance its effectiveness in handling open-ended queries.\n\n| Method     | AutoAgents vs. AgentVerse |\n|------------|---------------------------|\n| FairEval   | 91.3%                     |\n| HumanEval  | 77.5%                     |\n", "category": "Experimental_Analysis"}, {"question": "Why does the performance on the trivia creative writing task improve at N=10 compared to N=5, as shown in Table 3?", "answer": "The performance improves at N=10 compared to N=5 because the wider array of available resources at N=10 helps the methods craft more cohesive and realistic narratives. In contrast, the constrained resource pool at N=5 means that the exclusion of even a single response category can result in significant fluctuations in performance. This is exemplified by a marked 20% performance drop due to the lack of one type of response. At N=10, however, such performance disparities are notably mitigated, evidenced by a more modest 10% change under similar conditions of response type absence.", "category": "Experimental_Analysis"}, {"question": "Why does the paper adopt the vertical communication paradigm, which assigns different responsibilities to agents according to their roles, during the execution stage?", "answer": "The paper's approach is grounded in the belief that specific tasks are best handled by specialized Agents to achieve the most effective results. This principle advocates for assigning a single Agent to address a distinct set of challenges within its expertise, rather than diluting its focus across multiple domains. In situations where a task requires a solitary response, an agent within the paper's vertical structure is specifically appointed to fulfill that task. This targeted allocation of responsibilities ensures that each agent operates within its area of specialization, thereby optimizing performance. This strategy, which is also employed in the AgentVerse framework, underscores the benefits of harnessing the strengths of specialized agents for specific domains, validating the paper's approach in the vertical communication model.", "category": "Motivation_Analysis"}, {"question": "How does the proposed framework AutoAgents perform in non-open-ended question-answering tasks, as demonstrated by the Trivia Creative Writing task?", "answer": "The Trivia Creative Writing task serves as a rigorous test for large language models, gauging their ability to retrieve and integrate a wide range of information from their internal knowledge bases. In this task, models are required to construct a cohesive narrative around a specified topic while seamlessly incorporating answers to N trivia questions. The evaluation criteria for this task consider any match to the answer variants of a question as a correct mention. Consequently, the Trivia Creative Writing task fits the description of a non-open-ended question-answering challenge. In this setup, each sample presents a series of either 5 or 10 common sense questions that necessitate accurate responses from diverse methodologies before the models proceed to narrative creation. The principal measure for assessing this task is the successful integration of these correct answers into the narratives. An analysis of the results in Table 3 clearly demonstrates that narratives generated by AutoAgents feature a notably higher proportion of accurate responses. This outcome robustly validates the capability of AutoAgents in the sphere of non-open-ended question-answering, underlining their adeptness at effectively amalgamating precise information within imaginative storylines.", "category": "Experimental_Analysis"}, {"question": "How can the main contributions of AutoAgents be identified, given that the key difference between the proposed framework and existing methods like Social Simulacra, Epidemic Modeling, SSP, and AgentVerse is its use of self-refinement and collaborative refinement? Isn’t this difference essentially a prompting technique, one that has already been employed in many existing works like [1, 2, 3]?\n\n[1] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflexion: an autonomous agent with dynamic memory and self-reflection. arXiv preprint arXiv:2303.11366, 2023.\n\n[2] Zhibin Gou, Zhihong Shao, Yeyun Gong, Yelong Shen, Yujiu Yang, Nan Duan, and Weizhu Chen. Critic: Large language models can self-correct with tool-interactive critiquing, 2023.\n\n[3] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv preprint arXiv:2207.05608, 2022.", "answer": "AutoAgents is a novel framework that addresses two salient issues: multi-agent generation and multi-agent cooperation. With respect to multi-agent generation, the paper's method surpasses four existing methods in terms of the completeness and adaptability of agent creation, by employing Agent Generalization by Multi-Agent Discussion and Prompt Generalization. Regarding multi-agent cooperation, the paper's framework exhibits the efficacy of Self-refinement Action and Cooperative Refinement Action. As illustrated in the table below, the distinctive features of AutoAgents encompass Agent Generalization by Multi-Agent Discussion, Prompt Generalization, Self-refinement Action, and Cooperative Refinement Action. AutoAgents is the first holistic framework that integrates multi-agent generation and collaboration. \n\n| Method | Task | Agent Generalization by Multi-Agent Discussion | Prompt Generalization | Self-refinement Action | Collaborative Refinement Action |\n| --- | --- | --- | --- | --- | --- |\n| Social Simulacra | Social Simulation | ✗ | ✗ | ✗ | ✗ |\n| Epidemic Modeling | Social Simulation | ✗ | ✗ | ✗ | ✗ |\n| SSP | General Autonomous Agents | ✗ | ✗ | ✗ | ✗ |\n| AgentVerse | General Autonomous Agents | ✗ | ✗ | ✗ | ✗ |\n| AutoAgents | General Autonomous Agents | ✓ | ✓ | ✓ | ✓ |\n\n- __Agent Generalization by Multi-Agent Discussion__: Existing agent generation frameworks rely more on a single agent to generate the list of agents, but this often limits their adaptability to different tasks. For AutoAgents, the drafting stage is crucial for the auto-generation of agents, as the list of agents to be generated is determined through cooperative discussions between the Planner agent and the two Observers. Additionally, the paper further supplements the discussion with ablation experiments concerning the Agent Observer and Plan Observer in Appendix A, with the following table providing a detailed comparison. In the absence of cooperative discussions among Observers, there is a significant decline in the overall performance of AutoAgents. This also validates the importance of generating reasonable agents for task adaptability.\n\n| Method | N (# trivia questions) = 5|\n| --- | --- |\n| AutoAgents w/o observers | 87.0  |\n| AutoAgents | 90.0 |\n\n- __Prompt Generalization__: The prompt employed by AutoAgents exhibits a more universal nature, signifying its capacity to acclimate to diverse tasks without necessitating bespoke customization. Both AgentVerse[1] and SSP[2] have implemented task-specific enhancements for varied task evaluations. In contrast, the paper's methodology leverages a singular, unified prompt format to accommodate an array of tasks. The commendable efficacy in open-ended question-answer and trivia creative writing tasks further corroborates the wide-ranging applicability and versatility of prompt design within the AutoAgents framework.\n\n[1] https://github.com/OpenBMB/AgentVerse/tree/minecraft/agentverse/tasks\n\n[2] https://github.com/MikeWangWZHL/Solo-Performance-Prompting/tree/main/prompts\n\n", "category": "Method_Comparison"}, {"question": "How is it determined when and which agents should engage in collaborative refinement, and what are the specific details of this process?", "answer": "The determination of when and which agents engage in collaborative refinement is made through discussions between the Planner and Observer. The principles of collaborative refinement are incorporated into the Planner's prompt, allowing the Planner discretion on when to employ this mode. Specifically, each step must assign at least one expert role, and if multiple expert roles are involved, their contributions and collaboration methods must be specified to produce integrated results.", "category": "Method_Mechanics"}, {"question": "How does the system ensure a clear identification of agents for each step in the plan if agent generation and plan generation are not parallel processes?", "answer": "The paper’s system does not operate in parallel. The creation process of agents always precedes the development of the plan. The detailed explanations of the underlying principles for each prompt are in the appendix, specifically outlining the process within the Planner.\n```\n1. Understanding and Breaking Down Tasks: The first step involves comprehensively understanding, analyzing, and deconstructing the given task or problem.\n\n2. Utilization of Existing Expert Roles:\n- Fully leverage existing expert roles suited to the problem.\n- Ensure that these roles have cooperative or dependent relationships.\n- Output the details of selected roles in JSON format, including their original information.\n\n3. Creation of New Expert Roles:\n- Avoid duplication of functions in new roles.\n- New roles should have clear names, detailed descriptions, domain expertise, available tools, execution suggestions, and prompt templates.\n- Ensure each new expert role has a distinct responsibility and domain of expertise.\n- Specify the goals, constraints, and toolset for each new role.\n- Provide execution suggestions and develop prompt templates for each new role.\n- Output details of new roles in JSON format, following a specific structure.\n\n4. Creation of Detailed Execution Plan:\n- Develop a comprehensive plan with multiple steps addressing the problem.\n- Assign at least one expert role to each step, detailing their contributions and collaborations.\n- Provide detailed descriptions for each step, including expected outputs and inputs for subsequent steps.\n- Include a final independent step for a language expert to provide a detailed response to the user's question.\n- Present the execution plan as a numbered list, indicating the expert roles involved in each step.\n```", "category": "Method_Mechanics"}, {"question": "How does the Action Observer interact and communicate with other agents to manage tasks, and under what conditions does it adapt the execution plan?", "answer": "The Action Observer adopts a vertical communication strategy for interfacing with agents. This entails assigning diverse tasks to different agents, guided by the specifics outlined in the execution plan, which includes both the identification of agents and their corresponding tasks. The Action Observer, in its role, meticulously reviews the results archived in long-term memory while also contemplating forthcoming execution strategies. This process is pivotal for assessing whether alterations to the plan are necessary. In instances where the extant outcomes are insufficient for the subsequent agent's requirements, the Action Observer takes the initiative to refine the execution plan, tailoring it based on these existing results. Presented below is a sample prompt that exemplifies the task modification process.\n```\n2.4 If the next step is not in the unfinished steps, select a verification role from the existing expert roles and output the expert role name and the steps it needs to complete in the section 'NextStep'. Please indicate the name of the expert role used at the beginning of the step.\n```", "category": "Method_Mechanics"}, {"question": "How are the types of knowledge (short-term, long-term, dynamic) determined for sharing, and which agents are designated to receive each type of knowledge in the Knowledge Sharing Mechanism described in Section 3.2?", "answer": "- __Short-term Memory__: This is predominantly utilized by agents that emerge from self-refinement and collaborative-refinement processes. These agents leverage short-term memory to retain historical records throughout the refinement phase. Notably, every agent generated within the paper's framework actively employs short-term memory.\n\n- __Long-term Memory__: This component archives the execution results of each agent, primarily chronicling the cumulative progress of the execution plan. Access to long-term memory is exclusive to a pre-designated Action Observer. The Action Observer utilizes this memory to evaluate whether additional execution steps are necessary or if there's a need to generate dynamic memory content.\n\n- __Dynamic Memory__: This memory type is essential for distilling and summarizing pivotal information from long-term memory for diverse agents. Generated by the Action Observer, dynamic memory serves to amalgamate critical aspects of the overall execution plan, thereby facilitating the agents responsible for implementing subsequent tasks. Consequently, the dynamic memory content available to each generated agent may differ, contingent upon their specific roles and tasks.", "category": "Method_Mechanics"}, {"question": "What is the detailed, step-by-step description of the self-refinement agent and collaborative refinement actions in Section 3.2’s ‘Execution Phase’?", "answer": "In order to streamline the implementation process of self-refinement and collaborative refinement, the paper has introduced the concept of a 'custom agent'. The intricate design of its prompt has been meticulously detailed in Appendix D.5. The core design principles of this custom agent are as follows: \n1. Understanding Previous Results: Begin by analyzing the results of previous agents to grasp the context and progress of the task. \n\n2. Task Analysis and Breakdown: Understand, analyze, and deconstruct the given task. Use available tools to assist in task completion. \n\n3. Current Step Identification and Execution: \n- 3.1 Examine completed steps and their outcomes to identify the current step that needs to be completed. \n- 3.2 In the absence of completed steps, analyze the task, design a plan for the necessary steps, and accomplish the first one. \n- 3.3 If steps have been completed, understand them to determine the next step to be completed. 4. Tool Selection and Action Execution: \n- 4.1 Choose the appropriate tool from the given list ({tool}) to complete the current step. \n- 4.2 Follow specific format guidelines when using tools like 'Write File'. \n- 4.3 Once all steps are completed, use the 'Final Output' action to summarize each step's outputs, ensuring the final output is detailed, comprehensive, and solves the task. \n\n5. Maintaining Format Consistency: Ensure that the final output adheres to the given format example, prioritizing helpfulness, relevance, accuracy, and detail. Building upon the custom agent's framework, a critical aspect of the refinement process is the configuration of 'completed steps' in step 3.1. \n\nThe specific procedural steps for self-refinement and collaborative refinement are outlined as follows: \n__Self-refinement Action__: During its first execution, each custom agent omits the outlined step 3.1 and proceeds to complete the remaining steps. If the task's criteria are not met or the step limit has not been reached, the outcomes from this execution are incorporated as 'completed steps' in the prompt for the next iteration of the task. In subsequent executions, the custom agent will execute all steps, continuing this process until the task is deemed successfully completed. \n__Collaborative Refinement Action__: This mirrors the self-refinement action, yet involves active collaboration between multiple agents. For instance, when agents A and B collaborate on a task, A initially bypasses step 3.1 in its first execution. Upon completion, B incorporates A's results as 'completed steps' and then executes all steps. In later cycles, A and B alternate their roles in the task, perpetuating this collaborative cycle until a joint conclusion is reached.", "category": "Method_Mechanics"}, {"question": "What specific qualitative insights demonstrate the impact of Self-Refinement agents and Collaborative Refinement Action on the performance of AutoAgents compared to SSP, and how do these components contribute to the observed results?", "answer": "Appendix A of the paper has incorporated an ablation study specifically focused on AutoAgents within the context of trivia creative writing task. The table provided below delineates the performance fluctuations of AutoAgents when certain key components are omitted, thereby illustrating the impact of each individual element on the overall efficacy of the system. \n- __AutoAgents w/o observers__: During the Drafting Stage, the Planner in AutoAgents engages in collaborative discussions with two Observers to determine the optimal list of agents and the execution plan. Table below elucidates that in the absence of observers, there is a marked 3% reduction in the overall performance of AutoAgents. This substantiates the imperative role of collaborative discussions in agent generation. \n- __AutoAgents w/o self-refinement__: The performance of AutoAgents decreases by 3% in the absence of the self-refinement agent. This observation corroborates the assertion that self-refinement is instrumental in augmenting proficiency in trivia creative writing tasks. Furthermore, the enhancement of single agents via self-refinement plays a pivotal role in fortifying the integrity of the overarching multi-agent framework. \n- __AutoAgents w/o collaborative refinement__: It's observable that compared to the scenario with AutoAgents, there is a decline of 2% without collaborative refinement. Since the necessity for multiple agents to collaborate on a single task is entirely dependent on the decision made by the agent in the drafting phase, not all problems necessarily involve tasks that require collaborative refinement. However, it's evident that when this principle is omitted from the prompt's design, there's a noticeable performance decrease in AutoAgents. This also corroborates the beneficial impact of collaborative refinement in tackling complex tasks. \n- __AutoAgents w/o dynamic memory__: Dynamic memory predominantly addresses the requisites of specialized agents. The Action Observer amalgamates pivotal data for forthcoming tasks, utilizing the historical action records archived in long-term memory. The below table elucidates a 1% diminution in the efficacy of AutoAgents bereft of dynamic memory. Quintessential insights derived from dynamic memory are primarily assimilated into the prompt, thereby augmenting the comprehension of critical information and bolstering the operational proficiency of actions. \n\n| Method                      | N (# trivia questions) | 5    |\n|-----------------------------|------------------------|------|\n| SSP                         |                        | 84.4 |\n| AutoAgents w/o observers    |                        | 87.0 |\n| AutoAgents w/o self-refinement |                      | 87.0 |\n| AutoAgents w/o collaborative refinement |             | 88.0 |\n| AutoAgents w/o dynamic memory |                      | 89.0 |\n| AutoAgents                  |                        | 90.0 |\n", "category": "Experimental_Exposition"}, {"question": "Why was the ablation study limited to the exclusion of the Agent Observer and Plan Observer, and how does the study address the roles and importance of Short-term Memory, Long-term Memory, and Dynamic Memory within the AutoAgents framework?", "answer": "AutoAgents are fundamentally engineered to create unique agents tailored for varied tasks. Within this design, only a select group of agents – the Planner, Agent Observer, Plan Observer, and Action Observer – are pre-established. The Planner and Action Observer, playing pivotal roles in the drafting and execution phases, respectively, are essential components of the framework and are indispensable. Consequently, the paper's ablation study focused explicitly on examining the impact resulting from the exclusion of the Agent Observer and Plan Observer. \n\nAs depicted in Figure 8 of the paper, the specific agent types compatible with the three modes of knowledge sharing are as follows:\n - __Short-term Memory__: This is predominantly utilized by agents that emerge from self-refinement and collaborative-refinement processes. These agents leverage short-term memory to retain historical records throughout the refinement phase. Notably, every agent generated within the paper's framework uses short-term memory. \n- __Long-term Memory__: This component archives the execution results of each agent, primarily chronicling the cumulative progress of the execution plan. Access to long-term memory is exclusive to a pre-designated Action Observer. The Action Observer utilizes this memory to evaluate whether additional execution steps are necessary or if there's a need to generate dynamic memory content. \n- __Dynamic Memory__: This memory type is essential for distilling and summarizing pivotal information from long-term memory for diverse agents. Generated by the Action Observer, dynamic memory serves to amalgamate critical aspects of the overall execution plan, thereby facilitating the agents responsible for implementing subsequent tasks. Consequently, the dynamic memory content available to each generated agent may differ, contingent upon their specific roles and tasks. Due to the fundamental nature of Short-term Memory and Long-term Memory as indispensable components within the system, the paper's investigation focused exclusively on understanding and analyzing the role of dynamic memory. An ablation study of dynamic memory is included in Appendix A of the paper. The following table shows the results without the dynamic memory. It is observed that the performance of AutoAgents decreases by 1% without the summarization from dynamic memory, which also proves the importance of dynamic memory. \n\n| Method | N (# triva questions) = 5| \n| --- | --- | \n| AutoAgents w/o dynamic memory | 89.0  | \n| AutoAgents | 90.0 |", "category": "Experimental_Analysis"}, {"question": "How do the Self-Refinement agents and Collaborative Refinement Action influence the experiment results, particularly in comparison to SSP, and what are the qualitative insights into their effects?", "answer": "\n| Method | N (# triva questions) = 5|\n| --- | --- |\n| SSP | 84.4 | \n| AutoAgents w/o self-refinement | 87.0 |\n| AutoAgents w/o collaborative refinement | 88.0  |\n| AutoAgents | 90.0 |\n\n__Enhancing single-agent through self-refinement__. Self-Refinement is a technique that enables LLMs to “converse” with themselves, evaluate their own generation, and iteratively improve their answers. Although AutoAgents is a framework for multi-agent collaboration, it also requires self-refinement agents to perform specialized roles for individual tasks. Figure 6 in Appendix A depicts the self-refinement process of programmers’ coding. They first write a pseudo code file and then generate the corresponding program files based on it. This refinement process significantly enhances the validity of the output file. Additionally, as shown in the results in the above table, the performance of AutoAgents decreases by 3% in the absence of the self-refinement action. This observation corroborates the assertion that self-refinement is instrumental in augmenting the proficiency in performing trivia creative writing tasks. Furthermore, the enhancement of single agents via self-refinement plays a pivotal role in fortifying the integrity of the overarching multi-agent framework.\n__Enhancing multi-agent collaboration through collaborative refinement action__. For collaborative refinement, the process is similar to the collaborative dialogue mentioned above, which involves integrating knowledge from different domains to accomplish tasks that demand cross-domain knowledge fusion. The results in the above table demonstrate a decline in performance when collaborative refinement is absent. It is observed that compared to the scenario with AutoAgents, there is a decline of 2%. As the necessity for multiple agents to collaborate on a single task is entirely dependent on the decision made by the agent in the drafting phase, not all problems necessarily involve tasks that require collaborative refinement. However, it is evident that when this principle is omitted from the prompt’s design, there is a noticeable performance decrease in AutoAgents. This also corroborates the beneficial impact of collaborative refinement in tackling complex tasks effectively.", "category": "Experimental_Analysis"}, {"question": "Why are the OPEN-ENDED QUESTION ANSWER and The Trivia Creative Writing tasks well-suited for assessing the impact of the number of agents, self-refinement, and Collaborative Refinement Action?", "answer": "The Open-Ended Question Answering Task and Trivia Creative Writing tasks are well-suited for assessing these factors because they demonstrate the adaptability and versatility of dynamic agent generation. For open-ended questions, AutoAgents can generate agents from three distinct domains, providing more elaborate answers compared to GPT-4 (Figure 10, Appendix C). For the trivia creative writing task in Figures 11 and 12 in Appendix D, AutoAgents decomposed the task into four steps: first, it retrieved the answer to the question through a domain-specific agent and then constructed the story. Meanwhile, it is evident that the Language Expert Agent performed multiple checks on the coherence between the story content and the question, ensuring the accuracy of the story content. These examples all demonstrate the versatility of building a dynamic agent generation framework for complex tasks.", "category": "Experimental_Analysis"}, {"question": "Were details provided about the recruitment process and ethics approval for the volunteers in the Open-ended Q&A evaluation?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "1ikK0kHjvj", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "A Generalist Agent", "authors": ["Scott Reed", "Konrad Zolna", "Emilio Parisotto", "Sergio Gómez Colmenarejo", "Alexander Novikov", "Gabriel Barth-maron", "Mai Giménez", "Yury Sulsky", "Jackie Kay", "Jost Tobias Springenberg", "Tom Eccles", "Jake Bruce", "Ali Razavi", "Ashley Edwards", "Nicolas Heess", "Yutian Chen", "Raia Hadsell", "Oriol Vinyals", "Mahyar Bordbar", "Nando de Freitas"], "abstract": "Inspired by progress in large-scale language modeling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy. The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens. In this report we describe the model and the data, and document the current capabilities of Gato.", "keywords": "", "qa_pairs": [{"question": "What is the conclusion that Skill Mastery, involving the object shapes used for evaluation, aims to embody compared to Skill Generalization?", "answer": "The Skill Mastery challenge and the Skill Generalization task are both from the RGB Stacking benchmark, and their agent baselines were previously published by Lee, Devin (2021)[1]. Skill Mastery is considered an easier challenge because it evaluates the agent’s ability to grasp objects that were present in the training data. However, the agent still needs to generalize to new initial conditions.\n\n[1]Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdolmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In Conference on Robot Learning, 2021", "category": "Concept_Understanding"}, {"question": "What methodology was used to determine the sample weights for different datasets?", "answer": "The domain sampling weight was initially set to the square root of the number of tasks in the domain, followed by manual tuning to ensure no domain was overrepresented.", "category": "Experimental_Setup"}, {"question": "How can the individual dataset sizes be estimated in terms of bytes?", "answer": "The size in bytes can be estimated by considering that each token is 4 bytes, with an additional 768 bytes per token for environments with image observations. Based on this estimate, the total size of the tokenized control datasets is approximately 150TB.", "category": "Experimental_Setup"}, {"question": "Why was there limited evaluation of frozen weight zero-shot performance on new tasks in Gato, and what are the implications of this limitation?", "answer": "Gato is prompted with an expert demonstration, which helps the agent output actions corresponding to the given task. This is particularly useful because, unlike many multi-task reinforcement learning (RL) settings, there is otherwise no task identifier available to the agent. Gato infers the relevant task from the observations and actions in the prompt. However, the context length of the agent described in the paper is limited to 1024 tokens. This limitation means the agent sometimes processes only a few environment timesteps in total. This is especially true for environments with image observations, where each observation can result in over one hundred tokens depending on its resolution. Consequently, for certain environments, only a short segment of a demonstration episode fits within the transformer’s memory. Due to this limited prompt context, preliminary experiments with various prompt structures yielded very similar performance. Likewise, early evaluations of the model using prompt-based in-context learning on new environments did not demonstrate a significant performance improvement compared to evaluations without prompts in the same setting. Therefore, the limited context length is a current limitation of the architecture described in the paper, primarily due to the quadratic scaling of self-attention. ", "category": "Experimental_Analysis"}, {"question": "How is the information from different tasks stored in the distributed representations, and what does the visualization of T-SNE embeddings reveal about the clustering or superimposition of this information?", "answer": "The visualization of T-SNE embeddings, colorized per task in Section 5.7, shows that information within a task is clustered together in embedding space, and the language task clusters are closer together.", "category": "Experimental_Analysis"}, {"question": "What experiments and visualizations have been conducted to analyze the nature of representations in the attention layers, and what limitations exist in performing additional ablations over hyperparameters?", "answer": "Attention visualizations are provided in Section 5.6 and Appendix J, and per-task embeddings are also visualized. Training models to completion requires significant resources, such as 256 TPUs for approximately 5 days, which limits the extent of hyperparameter ablations that can be performed.", "category": "Experimental_Analysis"}, {"question": "Can you provide detailed information on how the finetuning data and settings were prepared for each task, particularly for the non-open-sourced task?", "answer": "The authors generated the fine-tuning data similarly to other tasks, as detailed in Section 3.1. The authors created cascading subsets by randomly selecting episodes in decreasing order (1000, 100, 10, 5, 3, 1) and repeated this 3 times for each task. Each subset was used for a separate fine-tuning experiment, reported in Section 5.2. The authors used the canonical versions of the tasks, with 3 out of 4 being open-sourced. For the fourth task, DMLab ‘Order of apples forage simple,’ the goal is to collect apples in a specific order, detailed in the Appendix E.", "category": "Experimental_Setup"}, {"question": "How can the Gato approach achieve the scale required for effective agents, particularly given the challenges of obtaining web-scale datasets for control tasks? While Gato’s primary contribution is its ability to handle multi-modal, multi-task data within a single framework, this alone may not be sufficient. The data used includes thousands, possibly millions, of episodes across various tasks, which is largely achievable only through simulations and simulated agents. For the current tasks, the authors likely had access to previously trained checkpoints from existing papers on the existing benchmarks. However, for novel tasks, one would need to build new simulations and develop a useful control agent that still provides enough diverse data.", "answer": "Gato is a data-driven approach, as it is derived from imitation learning. While natural language or image datasets are relatively easy to obtain from the web, a web-scale dataset for control tasks is not currently available. This may seem problematic at first, especially when scaling Gato to a higher number of parameters. That being said, extensive investigation has already been conducted into this issue. Offline RL aims at leveraging existing control datasets, and its increasing popularity has already resulted in the availability of more diverse and larger datasets. Richer environments and simulations are being built, such as the Metaverse. Additionally, an increasing number of users already interact with them, especially among the thousands of already deployed online games, like StarCraft 2, which has a large dataset. Real-life data has also been stored for ML research purposes; for example, data for training self-driving cars is acquired from recording human driver data. Finally, while Gato uses data consisting of both observations and corresponding actions, the possibility of using large-scale observation-only data to enhance agents has already been studied [Baker et al. 2022][1]. Thanks to online video sharing and streaming platforms such as YouTube and Twitch, observation-only datasets are not significantly more difficult to collect than natural language datasets, motivating a future research direction to extend Gato to learn from web data. While the previous paragraph focuses on alleviating drawbacks of data collection from RL agents, it is important to note that this approach presents a different set of tradeoffs compared to scraping web data and is actually more practical in some situations. Once the simulation is set up and a near state-of-the-art agent is trained, it can be used to generate massive amounts of high-quality data. This contrasts with the quality of web data, which is notorious for its low quality. In short, acquiring suitable data is another research question on its own, and this is an active area of research with growing momentum and importance.\n\n[1]Bowen Baker, Ilge Akkaya, Peter Zhokhov, Joost Huizinga, Jie Tang, Adrien Ecoffet, Brandon Houghton, Raul Sampedro, and Jeff Clune. Video pretraining (vpt): Learning to act by watching unlabeled online videos. Preprint arXiv::2206.11795, 2022.\n", "category": "Method_Disambiguation"}, {"question": "Does the 'generalist agent' demonstrate improved generalization capabilities through finetuning, particularly in comparison to training from scratch, and what are the results across different domains?", "answer": "Finetuning of Gato leads to clearly better agents compared to training from scratch for three domains, with consistent improvements across different numbers of demonstrations used for finetuning. Only one domain (ALE Atari) does not benefit from finetuning.", "category": "Experimental_Analysis"}, {"question": "How does Gato benefit from being prompted with expert demonstrations, and what are the limitations of this approach due to the context length of the agent?", "answer": "Gato benefits from being prompted with expert demonstrations by inferring the relevant task from the observations and actions in the prompt, which aids the agent in outputting actions corresponding to the given task. This is particularly useful because there is otherwise no task identifier available to the agent. However, the context length of the agent is limited to 1024 tokens, which means the agent sometimes attends to only a few environment timesteps, especially in environments with image observations where each observation can result in more than one hundred tokens. Hence for certain environments only a short chunk of a demonstration episode fits in the transformer memory. Due to this limited prompt context, preliminary experiments with different prompt structures resulted in very similar performance. Similarly, early evaluations of the model using prompt-based in-context learning on new environments did not show a significant performance improvement compared to prompt-less evaluation in the same setting. Context-length is therefore a current limitation of the paper's architecture, mainly due to the quadratic scaling of self-attention.", "category": "Method_Disambiguation"}, {"question": "What are the challenges and solutions for obtaining suitable datasets for imitation learning in RL tasks, particularly when compared to natural language or image datasets?", "answer": " Gato is a data-driven approach, as it is derived from imitation learning. While natural language or image datasets are relatively easy to obtain from the web, a web-scale dataset for control tasks is not currently available. This may seem at first to be problematic, especially when scaling Gato to a higher number of parameters. That being said, there has already been extensive investigation into this issue. Offline RL aims at leveraging existing control datasets, and its increasing popularity has already resulted in the availability of more diverse and larger datasets. Richer environments and simulations are being built (e.g. Metaverse), and increasing numbers of users already interact with them among thousands of already deployed online games (e.g. there exists a large dataset of Starcraft 2 games). Real-life data has also been already stored for ML research purposes; for example, data for training self-driving cars is acquired from recording human driver data. Finally, while Gato uses data consisting of both observations and corresponding actions, the possibility of using large scale observation-only data to enhance agents has been already studied [Baker et al. 2022][1]. Thanks to online video sharing and streaming platforms such as Youtube and Twitch, observation-only datasets are not significantly more difficult to collect than natural language datasets, motivating a future research direction to extend Gato to learn from web data. While the previous paragraph focuses on alleviating drawbacks of data collection from RL agents, it is important to note that this approach presents a different set of tradeoffs compared to scraping web data and can be actually more practical in some situations. Once the simulation is set up and near SOTA agent trained, it can be used to generate massive amounts of high quality data. That is in contrast to the quality of web data which is notorious for its low quality. In short, acquiring suitable data is another research question on its own, and this is an active area of research with growing momentum and importance.\n\n[1]Bowen Baker, Ilge Akkaya, Peter Zhokhov, Joost Huizinga, Jie Tang, Adrien Ecoffet, Brandon Houghton, Raul Sampedro, and Jeff Clune. Video pretraining (vpt): Learning to act by watching unlabeled online videos. Preprint arXiv::2206.11795, 2022.\n", "category": "Method_Disambiguation"}, {"question": "How does the paper justify Gato’s focus on multi-task training, given that although Gato is able to achieve multi-modal multi-task training, it does not bring significant improvement in a single task compared with large models in the NLP field?", "answer": "Gato focuses on multi-task training and achieves near expert/SOTA scores for a few hundreds of tasks. The specialist Meta-World agent approaches 100% success rate, and it is the best model known for this purpose.", "category": "Experimental_Analysis"}, {"question": "Did the paper train the model for a fixed number of updates (steps), always one million steps?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the version of Meta-world used in the study specified as July 23rd 2021 in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was a 50% success rate considered among the best for the robot task setting, even when compared to specialist agents trained with the same data volume using Offline RL (CRR)?", "answer": "True", "category": "Claim_Verification"}, {"question": "Can the total size of the tokenized control datasets be estimated to be approximately 150TB using the given token size information?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the individual dataset size explicitly provided in terms of 'bytes' in the paper?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "8s8K2UZGTZ", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "Teaching Models to Express Their Uncertainty in Words", "authors": ["Stephanie Lin", "Jacob Hilton", "Owain Evans"], "abstract": "We show that a GPT-3 model can learn to express uncertainty about its own answers in natural language -- without use of model logits. When given a question, the model generates both an answer and a level of confidence (e.g. \"90% confidence\" or \"high confidence\"). These levels map to probabilities that are well calibrated. The model also remains moderately calibrated under distribution shift, and is sensitive to uncertainty in its own answers, rather than imitating human examples. For testing calibration, we introduce the CalibratedMath suite of tasks. We compare the calibration of uncertainty expressed in words (\"verbalized probability\") to uncertainty extracted from model logits. Both kinds of uncertainty are capable of generalizing calibration under distribution shift. We also provide evidence that GPT-3's ability to generalize calibration depends on pre-trained latent representations that correlate with epistemic uncertainty over its answers.", "keywords": "", "qa_pairs": [{"question": "Why was verbalized probability calibration chosen instead of Platt scaling on model logits?", "answer": "Verbalized probability calibration can sometimes generalize better than logit-based calibration, as demonstrated in the Multi-Answer evaluation set. It also offers more flexible interactions with non-technical users and is applicable in scenarios where logits are not available, such as in information retrieval models.", "category": "Motivation_Analysis"}, {"question": "Can language models differentiate between cases where they are confident and right versus cases where they are confident and wrong, particularly in scenarios involving common misconceptions, outdated information, fiction, conspiracy theories, and social biases?", "answer": "Yes, it is plausible that language models can learn to differentiate between 'confident and right' and 'confident and wrong' through fine-tuning. For example, models can recognize features  (e.g. phrases and ideas that are more common on conspiracy-theory forums than on Wikipedia or BBC) associated with conspiracy theories, common misconceptions, and time-sensitive questions, and adjust their confidence levels accordingly  like 'Humans use only 10% of their brain'. While models cannot reliably answer all time-sensitive questions without external databases, they can still be well-calibrated by distinguishing between questions with quickly changing answers ('Did it rain in London today?' or 'Who is the PM of Israel?')  and those with slow changing answers ('What are the two largest political parties in the U.S?').", "category": "Experimental_Analysis"}, {"question": "Has GPT-3 demonstrated the ability to express its own uncertainty, and how does this relate to the calibration of its logits in tasks like Add-subtract and Multiply-divide?", "answer": "GPT-3’s logits are moderately calibrated zero-shot on both the Add-subtract and Multiply-divide task. The logits contain information about GPT3’s own uncertainty about the next token, not human uncertainty or the uncertainty of another model. During the paper's finetuning for verbalized probability, GPT3 could learn to output verbalized probabilities for an answer that depend on its zero-shot logit for the answer. (To be more precise, both the logit and the verbalized probability could depend on the same feature or pre-trained representations.) And if the verbalized probability depends on the logits, then it expresses the model’s own uncertainty. This is a fairly simple mechanism and doesn’t imply that GPT-3 would be capable of more sophisticated kinds of metacognition and self-knowledge.", "category": "Experimental_Analysis"}, {"question": "What are the advantages of using verbalized probability over other calibration methods?", "answer": "Verbalized probability offers several advantages: 1. Models using information retrieval may lack logit probabilities over their answers. 2. Models can use natural language to express continuous distributions, such as uncertainty over questions like 'How does wealth vary in this country?' or 'When will a large asteroid next hit the Earth?', allowing the model to decide which continuous distribution to use instead of requiring human users to choose the distribution and bounds on the parameters. 3. Suppose a non-technical human user Bob wants to get a language model to answer some questions and include uncertainty estimates (e.g. possible strategies for a company and estimates of how likely they are to work). Bob can give some examples of these strategies for few-shot prompting. If the model can produce calibrated verbalized probabilities, then it could usefully generalize the patterns in Bob’s examples. This setup (where Bob can teach the model using natural language) seems more flexible for non-technical users than trying to work with the logits. 4. Verbalized probability calibration may generalize better than logit-based calibration, as demonstrated by the example of generalizing to the Multi-Answer evaluation set.", "category": "Method_Comparison"}, {"question": "What evidence supports the claim that GPT-3 can express its own uncertainty through verbalized probabilities, and how does this relate to its calibration estimates?", "answer": "The evidence that GPT-3 can express its own uncertainty is tentative. GPT-3’s logits are moderately calibrated zero-shot on tasks like Add-subtract and Multiply-divide, containing information about its own uncertainty about the next token, not human uncertainty or the uncertainty of another model. During finetuning for verbalized probability, GPT-3 could learn to output probabilities dependent on its zero-shot logits, thus expressing its own uncertainty(To be more precise, both the logit and the verbalized probability could depend on the same feature or pre-trained representations). This mechanism is simple and does not imply sophisticated metacognition or self-knowledge.", "category": "Experimental_Analysis"}, {"question": "How does the paper address the distinction between model confidence and user confidence, and why is the focus on model confidence appropriate for achieving the paper's goals?", "answer": "The paper aims to fine-tune a model to provide well-calibrated confidences, which serve as a guide for users to determine how much they can trust the model. For example, if a calibrated model assigns low probability to its own answer to a question, then humans would know to not trust the answer. A calibrated model's confidence levels indicate to humans when to trust or not trust the model's answers, thereby aligning model confidence with user confidence. Focusing on model confidence is appropriate because it directly influences user trust in the model's outputs.", "category": "Motivation_Analysis"}, {"question": "How does the finetuning process in the paper improve the model's confidence calibration for individual examples, given that confidence is initially presented as more reliable for groups of examples?", "answer": "Finetuning improves calibration on the evaluation sets (both MSE and MAD metrics), making the model’s confidences for individual examples more informative about whether the answers are true or false. Calibrated models assign similar confidences to similar questions/answers. For example, if a model was totally ignorant about quantum mechanics, then it should assign 50% probability to all true/false exam questions about quantum mechanics. (The only way to avoid assigning similar confidence to similar questions/answers is if a model says 100% confidence for every true answer and 0% for every false answer). The finetuning uses fixed confidence labels for ~100 categories of questions(which vary based on the math concept involved, # of digits, and formatting), which may lead to less informative confidences in terms of MSE but still remain informative for individual examples, especially in off-distribution evaluation settings, Concretely, there are ~100 categories of question on the addition/subtraction training set. The questions on evaluation sets all involve different mathematical concepts (e.g. prime numbers) and may allow for multiple correct answers ('Name a prime number below 100'). Model cannot predict the categories of questions due to lack of labeled examples. ", "category": "Experimental_Setup"}]}
{"id": "DwgRm72GQF", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "Inverse Scaling: When Bigger Isn't Better", "authors": ["Ian R. McKenzie", "Alexander Lyzhov", "Michael Martin Pieler", "Alicia Parrish", "Aaron Mueller", "Ameya Prabhu", "Euan McLean", "Xudong Shen", "Joe Cavanagh", "Andrew Gritsevskiy", "Derik Kauffman", "Aaron T. Kirtland", "Zhengping Zhou", "Yuhui Zhang", "Sicong Huang", "Daniel Wurgaft", "Max Weiss", "Alexis Ross", "Gabriel Recchia", "Alisa Liu", "Jiacheng Liu", "Tom Tseng", "Tomek Korbak", "Najoung Kim", "Samuel R. Bowman", "Ethan Perez"], "abstract": "Work on scaling laws has found that large language models (LMs) show predictable improvements to overall loss with increased scale (model size, training data, and compute). Here, we present evidence for the claim that LMs may show inverse scaling, or worse task performance with increased scale, e.g., due to flaws in the training objective and data. We present empirical evidence of inverse scaling on 11 datasets collected by running a public contest, the Inverse Scaling Prize, with a substantial prize pool. Through analysis of the datasets, along with other examples found in the literature, we identify four potential causes of inverse scaling:\n(i) preference to repeat memorized sequences over following in-context instructions,\n(ii) imitation of undesirable patterns in the training data,\n(iii) tasks containing an easy distractor task which LMs could focus on, rather than the harder real task, and\n(iv) correct but misleading few-shot demonstrations of the task.\nWe release the winning datasets at inversescaling.com/data to allow for further investigation of inverse scaling. Our tasks have helped drive the discovery of U-shaped and inverted-U scaling trends, where an initial trend reverses, suggesting that scaling trends are less reliable at predicting the behavior of larger-scale models than previously understood. Overall, our results suggest that there are tasks for which increased model scale alone may not lead to progress, and that more careful thought needs to go into the data and objectives for training language models.", "keywords": "", "qa_pairs": [{"question": "Why is the use of FLOPS as the sole metric for measuring scaling considered appropriate within individual model families, and what are the limitations of using FLOPS for comparisons across different model families?", "answer": "FLOPS are considered appropriate within individual model families because they typically either keep the dataset size fixed or increase it as the model size increases, making them a good proxy for scale within a model family. However, FLOPS comparisons across models from different model families are not recommended due to potential inaccuracies and other issues.", "category": "Method_Disambiguation"}, {"question": "What are the specific technical contributions of this paper regarding the phenomenon of inverse scaling in language models?", "answer": "The specific technical contributions include discovering 11 cases of inverse scaling in language models, analyzing these cases to identify possible causes of inverse scaling behavior, providing the first systematic examination and explanation of the phenomenon, and collecting instances of inverse scaling from the literature to show how they fit into the explanation of why inverse scaling occurs in language models. ", "category": "Method_Disambiguation"}, {"question": "Why do the authors believe that LLMs fail at the modus tollens task despite human contractors showing good performance in Table 1?", "answer": "The authors hypothesize that the internet contains many examples of poor modus tollens reasoning from humans, which LLMs might imitate. (one extreme example of this is 'modus schmollens', where 'reasoners conclude that “the soup tastes like garlic” from the premises “If a soup tastes like garlic, then there is garlic in the soup; Carole tells Didier that there is no garlic in the soup they are eating”) This is not inconsistent with the good performance of human contractors who are specifically asked about the validity of modus tollens inferences and can use external resources.", "category": "Experimental_Analysis"}, {"question": "What are the key differences between BIG-Bench and the current work, particularly in terms of their goals and focus on inverse scaling behavior in language models?", "answer": "The key difference is that BIG-Bench aimed to crowdsource general language model evaluations, while the current work specifically targets revealing inverse scaling behavior in language models. BIG-Bench includes some tasks that demonstrate inverse scaling but does not actively solicit such tasks or discuss their potential causes. Additionally, some BIG-Bench tasks show U-shaped scaling.", "category": "Method_Comparison"}, {"question": "What is the purpose of the human agreement scores, and how were these scores collected?", "answer": "The purpose of the human agreement scores is to ensure that the labels given by task authors represent the answer that humans would agree, upon reflection, was correct for that task. This prevents the construction of inverse scaling tasks based on incorrect answers. The scores were collected by allowing contractors to use the internet to help them come to a final answer to each question, reducing the likelihood of mistakes or a lack of knowledge affecting the scores.", "category": "Method_Mechanics"}, {"question": "How is normalization applied in the description of Classification Accuracy?", "answer": "Normalization is performed over the multi-choice options given, rather than over all sequences.", "category": "Method_Mechanics"}, {"question": "What are the limitations of the non-apples-to-apples comparison between LLMs and humans in classification tasks, and how does the human comparison validate the task labels?", "answer": "The human comparison is intended to validate that the task labels given by the task author are correct, rather than to assess the performance an average human would receive. While the comparison is not exactly one-to-one (e.g., contractors could see multiple-choice options while models could not), this does not undermine the scores because the goal is to verify whether humans would agree that the provided task labels are correct, not to assess their ability to perform the task. Contractors were allowed to use the internet to ensure that mistakes or a lack of knowledge did not affect the scores, as the purpose of the human agreement scores is to confirm that the labels represent the answer humans would agree upon as correct. This is because it is easy to construct an inverse scaling task if the answers were allowed to be incorrect: a dataset of simple True/False questions with the True and False labels switched would show inverse scaling. Contractors were allowed to use the internet to help them come to a final answer to each question, since this meant it was less likely that mistakes or a lack of knowledge would affect the scores.'", "category": "Experimental_Setup"}]}
{"id": "YfZ4ZPt8zd", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks", "authors": ["Wenhu Chen", "Xueguang Ma", "Xinyi Wang", "William W. Cohen"], "abstract": "Recently, there has been significant progress in teaching language models to perform step-by-step reasoning to solve complex numerical reasoning tasks. Chain-of-thoughts prompting (CoT) is the state-of-art method for many of these tasks. CoT uses language models to produce text describing reasoning, and computation, and finally the answer to a question. Here we propose `Program of Thoughts' (PoT), which uses language models (mainly Codex) to generate text and programming language statements, and finally an answer. In PoT, the computation can be \ndelegated to a program interpreter, which is used to execute the generated program, thus decoupling complex computation from reasoning and language understanding. We evaluate PoT on five math word problem datasets and three financial-QA datasets in both few-shot and zero-shot settings.  We find that PoT has an average performance gain over CoT of around 12% across all datasets.\nBy combining PoT with self-consistency decoding, we can achieve extremely strong performance on all the math datasets and financial datasets. All of our data and code will be released.", "keywords": "", "qa_pairs": [{"question": "What does the term 'bias' refer to in the context of '-2 as the bias' mentioned on page #6, first line?", "answer": "The term 'bias' refers to the `logit_bias` argument in the `openai.Completion.create` function, where the vocab index is set to 1303 and the argument is set to -2 to achieve suppression.", "category": "Concept_Understanding"}, {"question": "How was the value -2 determined for the bias parameter in the experiment?", "answer": "The value -2 was determined by testing bias values of -1, -2, -3, and -5 on held-out 10 examples and observing the proportion of outputs that were executable code. -2 was found to perform roughly the best.", "category": "Experimental_Setup"}, {"question": "Has the study of GPT-4 been conducted with its exterior applications, and what are the results?", "answer": "Yes, GPT-4 experiments with CoT have been conducted on the GSM, AQuA, and FinQA datasets. The results show that GPT-4 (CoT) achieved 92.0 on GSM, 72.4 on AQuA, and 58.2 on FinQA, while GPT-4 (PoT) achieved 97.2 on GSM, 84.4 on AQuA, and 74.0 on FinQA, indicating significant improvements with PoT.\n| Method         | GSM    | AQuA  | FinQA |\n|----------------|---------|---------|---------|\n| GPT4 (CoT)  |  92.0   |   72.4    |  58.2   |\n| GPT4 (PoT)  |  97.2   |   84.4    |  74.0   |", "category": "Experimental_Exposition"}, {"question": "How does the model decide when to stop or use the intermediate result and continue prompting, particularly in the AQuA QA dataset, and why is this approach not necessary for other datasets?", "answer": "In the AQuA QA dataset, the model learns when to stop or require additional Chain-of-Thought (CoT) prompting through demonstrations. These demonstrations include examples with markers such as 'PoT program. **This is the answer**' and 'PoT program. **Please continue to **prompt**'. During prompting, the model detects these markers to decide the next step. If '**This is the answer**' is detected, the answer is compared with all options in AQuA using string match for prediction. If '**Please continue to **prompt**' is detected, the program output is executed and fed into CoT to choose the best answer. This approach is specific to AQuA QA because other datasets do not require such prompting. Additionally, while a program could convert floating-point numbers into HH:MM format, this task is better handled by CoT to leverage the strengths of both Program-of-Thought (PoT) and CoT, breaking down questions into computation and text reasoning parts for optimal performance.", "category": "Experimental_Setup"}, {"question": "What explains the performance difference between CoT + calc and PoT, and were the same prompts used for both experiments?", "answer": "The performance difference is due to CoT + calc using regular expressions for post-processing, It might be able to correct CoT output 'Alice has 20 apples and 33 pears, therefore she has 20+33=63 fruits' by finding the '=', and then use a calculator to fix '63' into '53'. However, if the CoT output is 'Alice has 20 apples and 33 pears, therefore she has 63 fruits', the regular expression won't detect where the addition happens, thus not able to fix it through post-processing.  Regarding prompts, the current experiments use the same 'demonstration question' but in different formats (language for CoT and program for PoT).", "category": "Experimental_Analysis"}, {"question": "What happens if the assumption that the problem is decomposable into programs is not satisfied in the proposed PoT technique?", "answer": "The proposed PoT technique assumes that the problem can be solved at least partially by a program. For AQuA QA, the model first performs calculations with programs and then conditions on the executed results to prompt further chain of thoughts. Clarifications have been added in the discussion section.", "category": "Experimental_Exposition"}, {"question": "How does the performance of PoT compare when using GPT-3.5-turbo and CodeGen as backend LLMs, given that Codex has been officially shut down?", "answer": "GPT-3.5-turbo shows reasonable improvement over Codex, while CodeGen's results are significantly worse than GPT-*.", "category": "Method_Comparison"}, {"question": "Did the results show that PoT significantly improved GPT-4's performance on all the tested datasets (GSM, AQuA, and FinQA)?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were experiments conducted using GPT-4 with CoT on the GSM, AQuA, and FinQA datasets to study its exterior applications?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "bx24KpJ4Eb", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback", "authors": ["Stephen Casper", "Xander Davies", "Claudia Shi", "Thomas Krendl Gilbert", "Jérémy Scheurer", "Javier Rando", "Rachel Freedman", "Tomek Korbak", "David Lindner", "Pedro Freire", "Tony Tong Wang", "Samuel Marks", "Charbel-Raphael Segerie", "Micah Carroll", "Andi Peng", "Phillip J.K. Christoffersen", "Mehul Damani", "Stewart Slocum", "Usman Anwar", "Anand Siththaranjan", "Max Nadeau", "Eric J Michaud", "Jacob Pfau", "Dmitrii Krasheninnikov", "Xin Chen", "Lauro Langosco", "Peter Hase", "Erdem Biyik", "Anca Dragan", "David Krueger", "Dorsa Sadigh", "Dylan Hadfield-Menell"], "abstract": "Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work emphasizes the limitations of RLHF and highlights the importance of a multi-layered approach to the development of safer AI systems.", "keywords": "", "qa_pairs": [{"question": "How does the method remain effective when applied to a dataset that is not exchangeable?", "answer": "The method can remain effective by adjusting the dataset to address the source of non-exchangeability. For example, ascending IDs(e.g. as in SQuAD and HumanEval), a common source of non-exchangeability, can be removed or permuted while keeping IDs constant to retain the test's applicability.", "category": "Method_Mechanics"}, {"question": "Was the separation of the fundamental limitation in Section 3.1.3 from Section 3.1.4 intentional, or was it an editing mistake?", "answer": "The separation was intentional, as there are multiple types of tradeoffs. Some involve the feedback type (e.g., preference feedback is easy but gives limited info; language feedback is rich but high effort) and some involve the overall process (sourcing data from product users versus trained experts). Authors added to Section 3.1.4 to explain this contrast.", "category": "Motivation_Analysis"}, {"question": "How does the paper address the concern that it overly focuses on LLMs rather than clearly distinguishing between RLHF and LLMs?", "answer": "The paper addresses this concern by explaining that LLMs are the most prominent application of RLHF today, making it challenging and undesirable to separate the study of RLHF challenges from LLMs. Additionally, all bold-headed paragraphs in section 3 about RLHF problems apply to non-LLM RLHF, and an extra mention was added to the introduction to clarify the focus on LLMs due to their current state-of-the-art applications.", "category": "Method_Mechanics"}, {"question": "What is the rationale behind the use of 'tractable' and 'fundamental' flags in distinguishing problems addressed by RLHF and non-RLHF approaches, and how have you addressed concerns about their subjectivity and potential for confusion?", "answer": "The 'tractable' and 'fundamental' flags are used to distinguish between problems that can be addressed with improvements to RLHF and those that require non-RLHF approaches. This distinction is important because RLHF is not equipped to handle certain problems. Authors have updated the paragraph in section 3 to explain the rationale behind this classification, focusing on why authors make this distinction rather than just what it is. Authors also provide detailed explanations in the Appendix and are open to reconsidering specific classifications or using alternative terminology.", "category": "Motivation_Analysis"}, {"question": "How does the authors' discussion of Inverse Reinforcement Learning (IRL) as a supplement to RLHF address the perceived inefficiency of using demonstrations in RLHF?", "answer": "The authors updated their discussion in section 4 to clarify that IRL is best viewed as an old technique to supplement RLHF rather than replace it. They argue that it is an open question how useful IRL might be with LLMs, but given the efficacy of supervised finetuning on good demonstrations, it seems plausible that reward models could efficiently learn from demonstrations. They also referenced recent work from [Li et al. (2023)] to support this perspective.", "category": "Method_Mechanics"}, {"question": "Does the rebuttal clarify that the tradeoffs involving feedback type and overall process were explained in section 3.1.4?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the fundamental limitation mentioned in section 3.1.3 intentionally separated from the content in section 3.1.4, or was it an editing mistake?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "jKN1pXi7b0", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "Unsupervised Dense Information Retrieval with Contrastive Learning", "authors": ["Gautier Izacard", "Mathilde Caron", "Lucas Hosseini", "Sebastian Riedel", "Piotr Bojanowski", "Armand Joulin", "Edouard Grave"], "abstract": "Recently, information retrieval has seen the emergence of dense retrievers, using neural networks, as an alternative to classical sparse methods based on term-frequency. These models have obtained state-of-the-art results on datasets and tasks where large training sets are available. However, they do not transfer well to new applications with no training data, and are outperformed by unsupervised term-frequency methods such as BM25. In this work, we explore the limits of contrastive learning as a way to train unsupervised dense retrievers and show that it leads to strong performance in various retrieval settings. On the BEIR benchmark our unsupervised model outperforms BM25 on 11 out of 15 datasets for the Recall@100. When used as pre-training before fine-tuning, either on a few thousands in-domain examples or on the large MS~MARCO dataset, our contrastive model leads to improvements on the BEIR benchmark. Finally, we evaluate our approach for multi-lingual retrieval, where training data is even scarcer than for English, and show that our approach leads to strong unsupervised performance. Our model also exhibits strong cross-lingual transfer when fine-tuned on supervised English data only and evaluated on low resources language such as Swahili. We show that our unsupervised models can perform cross-lingual retrieval between different scripts, such as retrieving English documents from Arabic queries, which would not be possible with term matching methods.", "keywords": "", "qa_pairs": [{"question": "What are the parameter counts for each method used in the finetuning experiments, and how do they compare?", "answer": "All neural network-based methods use the BERT base model, except ANCE which uses RoBERTa-base, and docT5query which includes an additional T5 model. Therefore, the parameter counts for the underlying models are roughly the same across all fine-tuned models.", "category": "Experimental_Setup"}, {"question": "What are the novel aspects of the proposed method compared to prior work such as Lee et al. (2019), and how do they contribute to the improvement in performance?", "answer": "The improvement in performance over ICT mainly comes from three factors: using MoCo to handle negatives, which allows scaling to a large number of negatives; the sampling procedure to generate pairs of (query, key); and using data from both CC-net and Wikipedia for training. The paper will be updated to highlight the origin of these gains.", "category": "Method_Comparison"}, {"question": "Is the mBERT baseline pre-trained on the same data as the proposed method, and if not, how does the performance of the proposed method compare to the mBERT and XLM-R baselines on the Mr.TyDi benchmark?", "answer": "The mBERT baseline is pre-trained on Wikipedia data only, while the proposed model is obtained by further training mBERT on Wikipedia and the CC100 data from XLM-R. The proposed model outperforms both the mBERT and XLM-R baselines on the Mr.TyDi benchmark, as shown in the provided table. The strong performance of the proposed model is attributed to the training objective rather than the data. The results for mBERT + MSMARCO differ due to a change in the temperature used for fine-tuning. The original paper used a temperature of 0.05, which led to poor performance for XLM-R. A similar temperature works well for both XLM-R and mBERT on MSMARCO before testing on Mr. TyDi. The pre-trained model, mContriever, is not sensitive to the fine-tuning temperature. The paper will be updated to reflect this change.\n| Model | ar | bn | en | fi | id | ja | ko | ru | sw | te | th | avg |\n| -----------  | ----------- | -----------  | ----------- | -----------  | ----------- | -----------  | ----------- | -----------  | ----------- | -----------  | ----------- | -----------  |\n| MRR@100 |\n| mBERT + MSMARCO | 33.9 | 36.2 | 23.9 | 30.0 | 35.2 | 26.3 | 28.6 | 30.3 | 36.9 | 38.9 | 18.7 | 30.8 |\n| XLM-R + MSMARCO  | 25.0 | 35.2 | 23.9 | 26.3 | 28.4 | 20.1 | 26.8 | 25.7 | 29.7 | 26.2 | 29.5 | 27.0 |\n| mContriever + MSMARCO  |  43.4 | 42.3 | 27.1 | 35.1 | 42.6 | 32.4 | 34.2 | 36.1 | 51.2 | 37.4 | 40.2 | 38.4 |\n| R@100 |\n| mBERT + MSMARCO | 81.1 | 88.7 | 78.3 | 74.2 | 81.0 | 76.1 | 66.7 | 77.6 | 74.1 | 89.5 | 57.8 | 76.8 |\n| XLM-R + MSMARCO | 80.6 |  87.4 | 75.8 | 80.8 | 87.9 | 71.0 | 69.6 | 76.2 | 72.7 | 89.9 | 90.4 | 80.2 |\n| mContriever + MSMARCO  | 88.7 | 91.4 | 77.2 | 88.1 | 89.8 | 81.7 | 78.2 | 83.8 | 91.4 | 96.6 | 90.5 | 87.0 |", "category": "Method_Comparison"}, {"question": "What was the training data construction setting for the Contriever model, including the use of Crop + delete, ICT, and the sampling process for the final positive training set?", "answer": "For the Contriever model, random cropping with 10% of token deletion was used. ICT was not used except for ablations. To sample a positive pair (query, key), a list of 256 tokens was obtained using the BERT BPE tokenizer. The length of the query and the key were sampled uniformly between 12 and 128 tokens, and then the query and key were uniformly sampled as contiguous segments of tokens according to their respective length among the 256 tokens.\nBelow is an example of (key, query) pair.\n\nQuery: *[CLS] lighting designer – billy mawer an early review of a technical rehearsal, published three days prior to the opening of the show to the public was negative – \" not quite the next rocky horror \", however subsequent media responses based on the actual performances before paying audiences were all positive. the script and playing time was reduced by 22 minutes prior to opening night and production was tightened in the first week of the season. the show went on to receive positive responses from many sources. these include : alan jones - \" there ’ s a rock musical on [SEP]*\n\nKey: *[CLS] prior to opening night and production was tightened in the first week of the season. the show went on to receive positive responses from many sources. these include : alan jones - \" there ’ s a rock musical on at the moment called the island of doctor moron … the singing dancing is unbelievable. \" ; \" the island of doctor moron has come [SEP]*", "category": "Experimental_Setup"}, {"question": "What are the computational costs of the proposed method compared to other dense retrieval models at inference time, and how do they benefit from techniques to reduce memory or computational costs?", "answer": "The majority of the methods compared to are retrievers using dense bi-encoders based on variants of BERT base, so embedding the collection of documents will have the same cost. At inference, these methods benefit from techniques like product quantization, dimension reduction, and approximate nearest neighbour search to reduce memory or computational costs. Some methods, like Colbert, require more memory and lead to slower inference, while others like Splade and BM25 are relatively fast but do not benefit from approximate nearest neighbor search or quantization. DocT5query uses a T5 encoder-decoder to augment the query. Additionally, the contrastive pre-training cost is similar to BERT and is amortised if the model is fine-tuned on multiple datasets.", "category": "Method_Comparison"}, {"question": "How does the amount of training data used for the inverse cloze task (ICT) and masked salient spans (MSS) baselines compare to the training data used in the proposed approach?", "answer": "The ICT/MSS baselines are trained for 100k steps with a batch size of 4096 and a maximum sentence length of 256 tokens, leading to a rough upper bound of 210B tokens. The proposed approach uses 500k steps with a batch size of 2048 and an average maximum sentence length of 140 tokens, roughly corresponding to 143B tokens. Additionally, when the proposed model is trained for 100k steps, it outperforms the ICT/MSS baselines while using less computational cost.\n| Model | NQ R@5 | NQ R@20 | NQ R@100 | TQA R@5 | TQA R@20 | TQA R@100 |\n| -----------  | :---: | :---: |:---: |:---: |:---: |:---: |\n| Inverse Cloze Task (Sachan et al.) | 32.3 | 50.9 | 66.8 | 40.2 | 57.5 | 73.6 |\n| Masked salient spans (Sachan et al.) | 41.7 | 59.8 | 74.9 | 53.3 | 68.2 | 79.4|\n| BM25 (Ma et al.) | - | 62.9 | 78.3 | - | 76.4 | 83.2 |\n| Contriever | 47.8 | 67.8 | 82.1 | 59.4 | 74.2 |83.2 |\n| 100k training steps on CC-net + Wikipedia | 46.0 | 67.1 | 81.7 | 60.0 | 74.2 |83.1 |\n| 100k training steps on Wikipedia | 42.7 | 63.4 | 77.7 | 55.9 | 70.9  | 81.8 |", "category": "Method_Comparison"}, {"question": "What are the main factors behind the improvements in performance over previous work, as analyzed in the ablation section?", "answer": "The improvement in performance over previous work mainly comes from three factors: using MoCo to handle negatives, which allows scaling to a large number of negatives; the sampling procedure to generate pairs of (query, key); and using data from both CC-net and Wikipedia for training.", "category": "Experimental_Analysis"}, {"question": "How does the performance of ICT with in-batch negatives compare to ICT with MoCo in the context of data augmentation methods?", "answer": "ICT with in-batch negatives is slightly worse than ICT with MoCo. The results are as follows: MoCo ICT has an overall score of 25.9, while in-batch ICT has an overall score of 24.1.\n | Model | NFCorpus | NQ | ArguAna | Quora | DBPedia | SciDocs | FEVER | Overall | \n| -----------  | :---: | :---: |:---: |:---: |:---: |:---: |:---: |:---: |\n | MoCo ICT  | 23.2 | 19.4 | 31.6 | 27.6 | 21.3 | 10.6 | 55.6 | 25.9 | \n| In-batch ICT | 20.3 | 13.3 | 23.8 | 69.0 | 20.1 | 9.0 | 38.7 | 24.1 |", "category": "Method_Comparison"}, {"question": "What are the key differences and learnings between the proposed work and Gao & Callan 2021, particularly in terms of methodology and results?", "answer": "The work by Gao & Callan primarily emphasizes the advantages of contrastive learning as an in-domain pre-training before fine-tuning. While the proposed work shares similar observations regarding the benefits of contrastive learning, it differs technically by using MoCo to handle negatives, enabling scaling to a large number of negatives. This approach results in unsupervised performance comparable to BM25 for Recall@100 on question answering datasets. Additionally, the proposed work includes training a multilingual model with the same contrastive learning approach, achieving state-of-the-art results after fine-tuning. The authors also note that their submission has been public since early October of the previous year.", "category": "Method_Comparison"}, {"question": "Are the results for mBERT + MSMARCO reported in the rebuttal identical to those reported in the original paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the rebuttal provide a comparison between the proposed model and the XLM-R base model fine-tuned on MS MARCO on the Mr.TyDi benchmark?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the rebuttal explain that the difference in results for mBERT + MSMARCO is due to a change in the temperature used for fine-tuning?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the mBERT baseline pre-trained on the same data as the proposed method in the multilingual experiment?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "TQeT7UupFk", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "One-for-All: Generalized LoRA for Parameter-Efficient Fine-tuning", "authors": ["Zeyu Han", "Chao Gao", "Jinyang Liu", "Jeff Zhang", "Sai Qian Zhang"], "abstract": "We present Generalized LoRA (GLoRA), a flexible approach for universal parameter-efficient fine-tuning tasks. Enhancing Low-Rank Adaptation (LoRA), GLoRA employs a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations, providing more flexibility and capability across diverse tasks and datasets. Moreover, GLoRA facilitates efficient parameter adaptation by employing a scalable, modular, and layer-wise structure search that learns the individual adapter of each layer. Originating from a unified mathematical formulation, GLoRA exhibits strong transfer learning, few-shot learning and domain generalization abilities, as it adapts to new tasks through not only weights but also additional dimensions like activations. Comprehensive experiments demonstrate that GLoRA outperforms all previous methods in natural, specialized, and structured benchmarks in the field of vision, achieving superior accuracy with fewer parameters and computations. To demonstrate the applicability in the language domain, we perform GLoRA on LLaMA-1/2 models, which also achieve considerable enhancements compared to the original LoRA. Furthermore, our structural re-parameterization design ensures that GLoRA incurs no extra inference cost, rendering it a practical solution for resource-limited applications.", "keywords": "", "qa_pairs": [{"question": "How does the increased complexity of GLoRA, which leads to a larger search space and longer training times, justify its development compared to simpler and more scalable PEFT methods?", "answer": "GLoRA addresses a crucial gap in existing PEFT methods' ability to generalize across diverse datasets by incorporating adapter-architecture search and additional training steps, enhancing performance and stability. Despite increased complexity and longer training times, GLoRA's superior performance and stability across datasets justify the additional computational costs. Comparative analysis shows GLoRA's training time is 6.6 folds over LoRA, but it outperforms the top five PEFT methods in aggregate, demonstrating its computational efficiency and effectiveness.\n| Method   | Training Time | Inference Time | Natural | Specialized | Structured | Average |\n|----------|---------------|----------------|---------|-------------|------------|---------|\n| NOAH     | $6 \\mathrm{x}$ | $\\uparrow$     | 80.3    | 84.9        | 61.3       | 75.5    |\n| Best-5   | $10.1 \\mathrm{x}$ | $\\uparrow$ | 82.5    | 86.7        | 63.3       | 77.5    |\n| GLoRA    | $6.6 \\mathrm{x}$  | -             | 83.6    | 87.0        | 63.3       | $\\mathbf{78.0}$ |", "category": "Experimental_Exposition"}, {"question": "What are the advantages of GLoRA over LoRA given the increased training time and memory consumption, and how do these trade-offs justify the performance improvements?", "answer": "GLoRA was designed to address a broader set of challenges, particularly in generalization across diverse datasets. The additional training time and memory consumption are justified by the significant performance improvements GLoRA achieves, as shown in Table 2 of the paper, where GLoRA achieves an average accuracy of 78.0, outperforming LoRA's 74.5. The architecture search in GLoRA is a one-time cost, and the identified subnetworks can be reused across similar tasks, amortizing the computational cost over multiple use cases. The additional computational effort is a worthwhile investment for the performance gains and generalization capabilities GLoRA provides.", "category": "Method_Comparison"}, {"question": "How does GLoRA compare to adaptive LoRA methods like IA3 and AdaLoRA in terms of performance and applicability?", "answer": "GLoRA has been compared to AdaLoRA, showing an average score of 50.8 for GLoRA versus 50.1 for AdaLoRA, indicating GLoRA's superiority. IA3 was not included in the baselines because it is inferior to AdaLoRA and is limited to generative language models, whereas GLoRA is applicable to both vision and language tasks.\n| Adaptation | ARC (25-s) | HellaSwag (10-s) | MMLU (5-s) | TruthfulQA (0-s) | Average |\n|------------|--------------------|------------------|------------|------------------|---------|\n| Base       | 51.0               | 77.8             | 35.7       | 34.3             | 49.7    |\n| LoRA       | 53.5               | 77.3             | 33.8       | 34.8             | 49.8    |\n| AdaLoRA    | 53.1               | 77.4             | 35.6       | 34.4             | 50.1    |\n| GLoRA      | 52.9               | 78.1             | 34.5       | 37.8             | 50.8    |", "category": "Method_Comparison"}, {"question": "How does the training time and computational burden change as the model size increases, particularly in the context of the evolutionary search procedure?", "answer": "The training time for larger models using the evolutionary search procedure does not significantly increase. Authors require only a single 40GB A100 for the procedure, with memory requirements similar to other PEFT methods. Around 20 epochs are sufficient for subnet search in general, and 5 epochs for LLM experiments. The top 50 subnets perform equally well due to weight entanglement, minimizing the need for a fine search strategy. This allows well-performing models to be obtained without a significant increase in computational burden by controlling the sparsity of the search approach.\n|           | Natural | Specialized | Structured |\n|-----------|---------|-------------|------------|\n| Mean      | 83.56   | 86.94       | 63.18      |\n| Std       | 0.0285  | 0.0178      | 0.0278     |", "category": "Experimental_Exposition"}, {"question": "How does the computational cost and training time of GLoRA compare to other PEFT methods like LoRA, considering both the evolutionary search and hyperparameter tuning requirements?", "answer": "GLoRA requires an additional 5.6 folds of training time compared to a single run of LoRA, totaling ~142 minutes for each VTAB-1k task. GLoRA's GPU memory consumption is 13 GB, compared to 9 GB for LoRA. It needs roughly 5 times more epochs than LoRA for convergence, with extra time spent on evolutionary search. This leads to a 4.5% accuracy increase across 19 vision tasks. When comparing the combined training time of the 5 best existing methods (LoRA, NOAH, FacT, SSF, RepAdapter), GLoRA's total training time (6.6x) is still lower and delivers superior performance. Additionally, GLoRA requires minimal hyperparameter search, unlike some other methods that need thorough data-specific tuning, training time required by GLoRA would be significantly lower in a relative sense.\n| Method   | Training Time | Inference Time | Natural | Specialized | Structured | Average |\n|----------|---------------|----------------|---------|-------------|------------|---------|\n| NOAH     | $6 \\mathrm{x}$ | $\\uparrow$     | 80.3    | 84.9        | 61.3       | 75.5    |\n| Best-5   | $10.1 \\mathrm{x}$ | $\\uparrow$ | 82.5    | 86.7        | 63.3       | 77.5    |\n| GLoRA    | $6.6 \\mathrm{x}$  | -             | 83.6    | 87.0        | 63.3       | $\\mathbf{78.0}$ |", "category": "Method_Comparison"}, {"question": "What are the exact number of parameters for each method listed in Table 4?", "answer": "The exact number of parameters for each method in Table 4 are as follows: Adapter - 0.16M, VPT - 0.53M, LoRA - 0.29M, NOAH - 0.36M, GLoRA - 0.29M.\n|                            |  ImageNet               | -Sketch | -V2 | -A  | -R  | Param(M)                  |\n|----------------------------|-----------------|-------|-----|-----|-----|-------------------|\n| Adapter | 70.5            | 16.4  | 59.1 | 5.5 | 22.1 | 0.16              |\n| VPT         | 70.5            | 18.3  | 58.0 | 4.6 | 23.2 | 0.53              |\n| LoRA        | 70.8            | 20.0  | 59.3 | 6.9 | 23.3 | 0.29              |\n| NOAH     | 71.5            | 24.8  | 66.1 | 11.9| 28.5 | 0.36              |\n| GLoRA        | 78.3            | 30.6  | 67.5 | 13.3| 31.0 | 0.29              |", "category": "Experimental_Setup"}, {"question": "What are the definitions and dimensions of all variables used in the equations in Section 2.1 for VPT, AdaptFormer, FacT, and RepAdaptor?", "answer": "All variables and their respective dimensions have been defined in the updated draft. For instance, in FacT, $\\mathbf{U} \\in \\mathbb{R}^{d \\times r_1}$, $\\mathbf{V} \\in \\mathbb{R}^{d \\times r_2}$, $ \\Sigma \\in \\mathbb{R}^{12L \\times r_1 \\times r_2}$, and $L$ represents the total number of layers in the network. Similar definitions are provided for VPT, AdaptFormer, and RepAdaptor.", "category": "Concept_Understanding"}, {"question": "What is the specific limitation of LoRA that prevents it from scaling up automatically for larger matrices, and can the rank be increased to address this issue?", "answer": "LoRA can be re-parameterized at inference but requires manual rank search based on the matrix size, network position, and layer type, which prevents it from scaling up automatically for larger matrices. Increasing the rank alone does not resolve this limitation.\nOther studies [1] also mentioned and discussed this in their work.\n\n[1] ADALORA: ADAPTIVE BUDGET ALLOCATION FOR PARAMETER-EFFICIENT FINE-TUNING", "category": "Method_Disambiguation"}, {"question": "What is the process for selecting the final number of parameters in GLoRA, and what are the resulting parameter counts for LLaMA-7B with Alpaca and ShareGPT datasets?", "answer": "The final number of parameters in GLoRA is selected through an evolutionary search procedure that identifies the optimal layer-wise configuration for each layer. Only the parameters corresponding to the identified optimal configuration are retained. For instance, for LLaMA-7B with the Alpaca dataset, this results in 1.9M parameters, and for LLaMA-7B with the ShareGPT dataset, it results in 2.2M parameters.", "category": "Experimental_Setup"}, {"question": "Why is GLoRA slower than other PEFT algorithms, and how is this additional training time justified?", "answer": "GLoRA is slower than other PEFT algorithms because it adopts evolutionary search, which increases the associated training time. Specifically, GLoRA requires an additional 5.6 folds of training time compared to LoRA, amounting to ~142 minutes per VTAB-1k task, and consumes 13 GB of GPU memory compared to LoRA's 9 GB. This is primarily due to GLoRA requiring roughly 5 times more epochs for convergence and the additional time spent on evolutionary search. However, the additional training time is justified by GLoRA's average 4.5% accuracy increase across 19 vision tasks compared to LoRA, its superior generalization across datasets, and its overall performance improvements. Furthermore, GLoRA's total training time (6.6x) is less than the combined training time of the 5 best-performing methods on the VTAB-1k dataset (LoRA, NOAH, FacT, SSF, and RepAdapter), while delivering superior performance. Unlike other methods, GLoRA requires minimal hyperparameter search, which would further reduce its relative training time if considered.", "category": "Experimental_Analysis"}, {"question": "How sensitive is GLoRA to the choice of hyperparameters in the evolutionary algorithm, and why was an ablation study on hyperparameter optimization not conducted?", "answer": "The paper did not include an exhaustive exploration of hyperparameters for the evolutionary search process. Instead, the authors adopted the approach used by NOAH, as detailed in the Appendix, without additional hyperparameter optimization (HPO). The default hyperparameters used with GLoRA produced highly competitive results, demonstrating robustness and effectiveness. The authors chose this strategy to highlight the inherent capabilities of GLoRA in facilitating effective evolutionary search, yielding state-of-the-art results without extensive hyperparameter tuning. Sensitivity to different HPO is acknowledged as an important subject and will be explored in follow-up research.", "category": "Motivation_Analysis"}, {"question": "Did the method using a fixed set of hyperparameters with GLoRA achieve state-of-the-art (SOTA) results?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper include an ablation study on the choice of hyper-parameters in the evolutionary algorithm used with GLoRA?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the study conduct additional hyperparameter optimization (HPO) beyond the approach used by NOAH for the evolutionary search?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "p6xslUyvka", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Detecting Anomalies within Time Series using Local Neural Transformations", "authors": ["Hari Sowrirajan", "Jingbo Yang", "Andrew Y. Ng", "Pranav Rajpurkar"], "abstract": "We develop a new method to detect anomalies within time series, which is essential in many\napplication domains, reaching from self-driving cars, finance, and marketing to medical\ndiagnosis and epidemiology. The method is based on self-supervised deep learning that has\nplayed a key role in facilitating deep anomaly detection on images, where powerful image\ntransformations are available. However, such transformations are widely unavailable for time\nseries. Addressing this, we develop Local Neural Transformations (LNT), a method learning\nlocal transformations of time series from data. The method produces an anomaly score for\neach time step and thus can be used to detect anomalies within time series. We prove in\na theoretical analysis that our novel training objective is more suitable for transformation\nlearning than previous deep Anomaly detection (AD) methods. Our experiments demonstrate\nthat LNT can find anomalies in speech segments from the LibriSpeech data set and better\ndetect interruptions to cyber-physical systems than previous work. Visualization of the\nlearned transformations gives insight into the type of transformations that LNT learns.", "keywords": "", "qa_pairs": [{"question": "How did the authors address the concerns raised about the empirical evaluation, including comparisons with prior work and baselines?", "answer": "The authors addressed the concerns by adding the performance of their own run of GDN to the results table for comparison. Additionally, they included an additional baseline with an adjusted threshold to match the precision level of LNT and compared the recall of LNT against this baseline.", "category": "Method_Mechanics"}, {"question": "What is the importance of hyperparameter selection in the proposed LNT approach, and how are these hyperparameters determined in the absence of labeled anomalous data for tuning?", "answer": "The crucial aspect of LNT's hyperparameter selection lies in the representation learning with CPC, which depends on the frequency of observations and sequence lengths in the time series data. These parameters can be determined similarly to other representation learning methods. Selecting incorrect hyperparameters can lead to performance collapse, but good hyperparameters can be found without requiring anomalies in the validation data. The sizes of embedding vectors z and c vary based on dataset size and inherent variations, and neural transformation sizes are scaled proportionally to these embeddings and validated with smaller validation sets containing anomalies. Suboptimal transformation sizes do not cause complete collapse if the underlying representation learning is converging.", "category": "Method_Mechanics"}, {"question": "Why were simple baseline methods, such as those proposed by Wu & Keogh, not included in the experiments, and how do the current experiments address the concerns raised by Wu & Keogh?", "answer": "The work by Wu/Keogh addresses issues like triviality, unrealistic anomaly density, mislabeled ground truth, and run-to-failure bias in anomaly detection benchmarks. Authors' synthetic experiments on LibriSpeech control the generation of artificial anomalies to address concerns (ii)-(iv), ensuring a realistic anomaly density and avoiding mislabeled ground truth and positional bias. Specifically, authors set only 10% of a time series to be an anomaly to have a realistic density. Mislabeled ground truth can be excluded for synthetic anomalies as well and since authors equally and randomly place anomalies within a time series in order to not have any bias in the position. Authors chose other datasets not listed by Wu/Keogh. The remaining concern, triviality, is addressed by authors' self-supervised approach, which learns a robust profile of normality without specific anomaly knowledge, unlike the supervised approach of Wu/Keogh that requires tuning for specific anomalies. Adding a simple baseline like moving mean for LibriSpeech could be considered, as the moving mean should remain zero regardless of added sine tones.", "category": "Experimental_Analysis"}, {"question": "How does the proposed method compare with self-supervised anomaly detection works for time series, such as those referenced in [2] and [4]?\n[2] TimeAutoAD: Autonomous Anomaly Detection with Self-Supervised Contrastive Loss for Multivariate Time Series. IEEE Transactions on Network Science and Engineering.\n[4] Self-Supervised Learning for Time-Series Anomaly Detection in Industrial Internet of Things.", "answer": "The empirical ablation in Section 5.5 demonstrates that DDCL enhances anomaly detection performance beyond what is achieved by pure CPC. Furthermore, the related work section has been updated to include comparisons with self-supervised anomaly detection methods for time series.", "category": "Method_Comparison"}, {"question": "How do the authors' contributions in combining self-supervised methods with deep learning architectures for time series anomaly detection differ from existing approaches, and what specific value does their method add?", "answer": "The authors address the open research question of extending data transformation methods, which show superior performance in images and tabular data, to time series anomaly detection. Their theoretical analysis indicates the necessity of combining these methods with temporal representation learning (CPC) to avoid trivial collapse. Empirical ablation studies demonstrate that DDCL enhances anomaly detection performance beyond what pure CPC achieves.", "category": "Method_Comparison"}, {"question": "What is the specific definition of a temporal anomaly in the context of time-series data, and how does the proposed method ensure that change points or non-stationary dynamics are not incorrectly identified as anomalies?", "answer": "A temporal anomaly is defined as an 'out of distribution' observation conditioned on the observed history at all time steps. The method ensures that change points or non-stationary dynamics are not incorrectly identified as anomalies by incorporating normal change points into the model's learned representations during training, as demonstrated with the LibriSpeech data where only synthetic anomalies are alerted.", "category": "Method_Mechanics"}, {"question": "Did the experiments include ablation studies comparing different versions of the proposed method, particularly given its multiple loss components and architectural choices?", "answer": "Yes, the paper includes an ablation study in Section 5.5, where different approaches to using CPC (with DDCL ablated) for anomaly detection were considered. Additionally, the reverse ablation (DDCL without CPC) is discussed in the theoretical analysis in Section 4.1, showing that the solution becomes impractical for anomaly detection.", "category": "Experimental_Exposition"}, {"question": "Was the performance of GDN included in the comparison table for SWaT to address the need for stronger competitors?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were experiments with other methods of injecting synthetic anomalies into LibreSpeech incorporated to strengthen the paper?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "hUgBU9wgSf", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "A Kernel Two-sample Test for Dynamical Systems", "authors": ["Hari Sowrirajan", "Jingbo Yang", "Andrew Y. Ng", "Pranav Rajpurkar"], "abstract": "Evaluating whether data streams are drawn from the same distribution is at the heart of various machine learning problems. This is articularly relevant for data generated by dynamical systems since such systems are essential for many real-world processes in biomedical, economic, or engineering systems. While kernel two-sample tests are powerful for comparing independent and identically distributed random variables, no established method exists for comparing dynamical systems. The main problem is the inherently violated independence assumption. We propose a two-sample test for dynamical systems by addressing three core challenges: we (i) introduce a novel notion of mixing that captures autocorrelations in a relevant metric, (ii) propose an efficient way to estimate the speed of mixing relying purely\non data, and (iii) integrate these into established kernel two-sample tests. The result is a data-driven method that is straightforward to use in practice and comes with sound theoretical guarantees. In an example application to anomaly detection from human walking data, we show that the test is readily applicable without any human expert knowledge and feature engineering.", "keywords": "", "qa_pairs": [{"question": "How is the time shift selected in practice, specifically in relation to MMD-mixing as defined in Def. 7?", "answer": "In practice, the time shift is selected by picking the first time shift that yields an HSIC value smaller than the test threshold. Clearly, authors cannot guarantee that it will not increase again afterward. Nevertheless, data that is sampled according to this shift exhibits (almost) independent samples from the stationary distribution and hence, is suitable for subsequent kernel two-sample testing. This is based on the assumption that a system is mixing, as discussed in Appendix A.1.1. Additionally, an explanation of this process has been added to the main corpus of the manuscript.", "category": "Method_Mechanics"}, {"question": "What does it mean to be 'not able to estimate an a*' in the context of the manuscript, and how is this condition diagnosed?", "answer": "The inability to estimate an appropriate $a^*$ occurs when the data remains heavily correlated even after applying the maximal possible time shift, indicating insufficient mixing. An intuitive example are LTI systems $x_{k+1} = Ax_k + \\epsilon_k$, where dynamics $A$ are close to the identity matrix and $\\epsilon$ has zero mean and tiny variance. This is diagnosed by observing that the kernel two-sample test should not be applied in such cases, as the null hypothesis of independent data is rejected. The proposed method detects this by estimating mixing properties from the data rather than assuming it beforehand, as demonstrated in numerical experiments where randomized dynamics and noise matrices were used. In our numerical experiments, authors have randomized the dynamics and noise matrix. By sampling often enough, authors will also pick systems that mix slowly. As described above, authors could detect the slow mixing properties. For the numerical examples, authors rerolled the dynamics if the time shift exceeded an a priori defined maximal value. ", "category": "Concept_Understanding"}, {"question": "How does the proposed method compare to relevant baselines, such as naive iid two-sample tests or advanced methods like two-sample tests for functional data?", "answer": "The proposed method is tailored for stationary systems with long trajectories and generalizes state-of-the-art methods requiring i.i.d. assumptions for dynamical systems. It focuses on comparing underlying stationary distributions, unlike methods like Wynne & Duncan (2022), which compare joint distributions of trajectories. The MMD, a well-established method for i.i.d. data, is adapted here for non-i.i.d. dynamical systems by estimating mixing times from data, a unique feature not found in other methods. Additionally, the method is computationally efficient since authors do not require sophisticated kernels and aim for maximizing the information content of each data point, the method makes it suitable for applications in embedded and battery-powered devices.", "category": "Method_Comparison"}, {"question": "How does the proposed work relate to the methods and ideas presented in (Chérief-Abdellatif & Alquier, 2022) and (Wynne and Duncan, 'A Kernel Two-Sample Test for Functional Data, 2020)?", "answer": "The work by Chérief-Abdellatif & Alquier (2022) introduces a similar idea of mixing in RKHS but focuses on parameter estimation using MMD as a loss function. Their notion of mixing differs from authors, and they do not estimate mixing from data. The work by Wynne & Duncan (2022) investigates whether two samples of functions have the same underlying distribution by embedding functional data into an RKHS. Their kernel design could be useful for extending the method to compare multiple correlated joint distributions, as authors’ current work focuses on stationary distributions and decorrelates data via mixing.", "category": "Method_Comparison"}, {"question": "What are the details of the kernel family and hyperparameters (e.g., kernel width) used in MMD computation, and how sensitive is the approach to these modeling choices?", "answer": "The paper uses the squared exponential kernel with the median heuristic to determine the kernel length scale, following Gretton et al. The approach is robust with these settings, though hyperparameter optimization can further improve the test's power. This optimization was excluded as it is orthogonal to the paper's main contributions.", "category": "Method_Disambiguation"}, {"question": "Why are precision, recall, and AUC not suitable evaluation metrics for the classification algorithm used in this study, and what alternative approach is employed for classification?", "answer": "Precision, recall, and AUC are not suitable because the classification algorithm does not involve a gradually increased decision boundary. Instead, a fixed trajectory is labeled by comparing it against all known labeled trajectories using the Maximum Mean Discrepancy (MMD) metric. The closest trajectory is selected to determine the label. This approach leverages the MMD as a metric to handle more complicated problems, especially when there is substantial deviation within each class. Then authors are technically considering probability distributions over probability distributions. However, the illustrated approach works well. Alternatively, authors could correct the test threshold and retrieve all trajectories that are below the threshold. This is particularly relevant when there are mistakes in the labels. Here, however, authors are certain that all trajectories are labeled correctly.", "category": "Experimental_Analysis"}, {"question": "Where is MMD used with samples from one trajectory versus with $m$ trajectories from $X$ and $Y$? Specifically, is Section 5.1 or 5.2 used in Experiment 8.2.1, and how is $n$ in the hypothesis test threshold related to the number of trajectories $m$?", "answer": "MMD is used in two components of the approach: a) Estimating mixing times using $m$ independent trajectories to determine the appropriate time shift for subsampling, and b) kernel two-sample testing applied to subsampled data from two long trajectories. In Experiment 8.2.1, $m = 70$ (35 subjects and access to two different trajectories for each subject. One with an orthosis restricting the movement of the knee and one without the orthosis), and the approach from Section 5.1 is used. $n$ represents the number of points extracted from each long trajectory for kernel two-sample testing, which is determined based on the estimated time shift from the $m$ trajectories.", "category": "Experimental_Setup"}, {"question": "Why were baselines with learned feature representations, such as another kernel method or a neural network, not included in the experimental evaluation?", "answer": "The main goal of the paper is to apply the standard kernel-two sample test to dynamical systems, demonstrating the relevance of dynamical systems, the suitability of the new mixing assumption for real-world problems, and the effectiveness of subsampling via the estimated mixing speed to treat data as independent. The selected baselines are those commonly used in practice, and the performance of statistical tests for dynamical systems is highly system-dependent. Additionally, the paper introduces a new method to analyze mixing behavior, which is crucial for dynamical systems.", "category": "Experimental_Analysis"}, {"question": "How does the running time of computing confidence intervals for HSIC across different time shifts compare to the competitor method mentioned in Remark 1, and what is the practical approach for determining an adequate time shift $a^*$?", "answer": "The running time of HSIC is quite fast, scaling with the number of independent trajectories, and evaluations for different time shifts can be computed in parallel. The goal is to find a large enough shift $a^*$ for approximately independent data, without needing to compute for all possible shifts. Finding one shift that is large enough is sufficient. In the paper, authors have computed the HSIC in a high resolution (with confidence bounds) to illustrate the decrease in dependencies. The competitor method in Remark 1 does not estimate mixing from data and is used for parameter estimation, a different problem. Their technical formulation could aid in deriving sharper bounds for the paper's Theorem 3 but requires further evaluation beyond this paper's scope.", "category": "Method_Comparison"}, {"question": "What are some theoretical examples of data that provably satisfy the assumption of being epsilon-independent, and how does the data in Section 8 achieve epsilon-independence for a sufficient time shift?", "answer": "Theoretical examples include independent and identically distributed (i.i.d.) data, which naturally satisfies the epsilon-independence property when $\\epsilon=\\alpha$ and $\\kappa$ is chosen appropriately for a level-$\\alpha$ kernel two-sample test. In practice, while authors cannot guarantee data independence or identical distribution, the HSIC over time shifts can determine with high probability that data is too correlated for tasks like two-sample testing. The examples in Figures 1 and 2 are empirical, as deriving theoretical mixing times is nontrivial. Further, these results can render useless in the real world since there is usually a substantial discrepancy between simulation models and the real world behavior of a system. Thus, in this paper, authors present an algorithm that can estimate mixing times directly from real-world data and show that this enables reliable two-sample testing for dynamical systems. Authors have actually discussed the walking results with colleagues that work on bipedal locomotion and in particular, model predictive control. For the data in Section 8, the results align with the heuristic in bipedal locomotion research that predictions become unreliable after three walking steps, indicating that data is mixed after this time shift. The oscillatory behavior of the HSIC in Figure 3 corresponds to steps, with certain parts of the step being more predictable than others.", "category": "Experimental_Analysis"}, {"question": "How does the proposed classification method for time series compare to nearest neighbor techniques on the UCR public library, especially when the data is not assumed to be iid or stationary?", "answer": "The proposed method is specifically designed for systems with many independent and long trajectories and performs well under these conditions. When tested on the earthquake dataset from the UCR public library, which is non-iid and non-stationary, the method classified 51.08% of trajectories correctly, 22.30% as unknown, and 26.62% incorrectly. Performance varied by label: 65.38% correct for 'no major event' and 8.57% correct for 'earthquake' trajectories. The weaker performance compared to nearest neighbor methods is partly due to the non-stationarity of the data and the method's design as a two-sample test, which assumes samples with the same label follow the same distribution—an assumption that does not hold for this dataset. For the earthquake dataset, this is not a reasonable assumption. The labels are determined by a value on the Richter scale, where everything above 4 is classified as an earthquake. However, trajectories with different values on the Richter scale, even if all of them are either above or below 4, may very well follow completely different distributions. ", "category": "Method_Comparison"}]}
{"id": "aRtjVZvbpK", "venue": "TMLR 2024", "paper_decision": "Withdraw", "title": "A Unified Survey on Anomaly, Novelty, Open-Set, and Out of-Distribution Detection: Solutions and Future Challenges", "authors": ["Mohammadreza Salehi", "Hossein Mirzaei", "Dan Hendrycks", "Yixuan Li", "Mohammad Hossein Rohban", "Mohammad Sabokrou"], "abstract": "Machine learning models often encounter samples that are diverged from the training distribution. Failure to recognize an out-of-distribution (OOD) sample, and consequently assign that sample to an in-class label, significantly compromises the reliability of a model. The problem has gained significant attention due to its importance for safety deploying models in open-world settings. Detecting OOD samples is challenging due to the intractability of modeling all possible unknown distributions. To date, several research domains tackle\nthe problem of detecting unfamiliar samples, including anomaly detection, novelty detection, one-class learning, open set recognition, and out-of-distribution detection. Despite having similar and shared concepts, out-of-distribution, open-set, and anomaly detection\nhave been investigated independently. Accordingly, these research avenues have not crosspollinated, creating research barriers. While some surveys intend to provide an overview of these approaches, they seem to only focus on a specific domain without examining the\nrelationship between different domains. This survey aims to provide a cross-domain and comprehensive review of numerous eminent works in respective areas while identifying their commonalities. Researchers can benefit from the overview of research advances in different fields and develop future methodology synergistically. Furthermore, to the best of our knowledge, while there are surveys in anomaly detection or one-class learning, there is no comprehensive or up-to-date survey on out-of-distribution detection, which this survey covers extensively. Finally, having a unified cross-domain perspective, this study discusses and sheds light on future lines of research, intending to bring these fields closer together.", "keywords": "", "qa_pairs": [{"question": "Why is there a lack of extensive quantitative benchmarking for AD and ND tasks in the paper, and how has this been addressed?", "answer": "The lack of extensive quantitative benchmarking for AD and ND tasks has been addressed by expanding Table 7 to include per-class results for commonly used datasets such as MNIST, CIFAR-10, and MVTecAD. Additionally, a GitHub repository has been created, containing OOD baselines and all evaluation metrics discussed in the paper, enabling consistent evaluation and benchmarking of SOTA pre-trained models on different datasets.", "category": "Method_Mechanics"}, {"question": "How do the authors address the lack of clear connections between different methods discussed in the paper, and what steps have been taken to provide a unified perspective on the relationships of these methods?", "answer": "Section 2 provides a unified perspective on the relationships of different tasks. The authors have reviewed critical methods in anomaly detection, open set recognition, and novelty detection, explaining each technique and clarifying their connections to others. For instance, the explanation of the 'Redefining the Adversarially Learned One-Class Classifier Training Paradigm (Old-is-Gold) Zaheer et al. (2020)' paper discusses its relation to the ALLOC method. While summarizing core ideas is beneficial, it would require omitting method details due to space constraints. The main differences between these methods, particularly in their evaluation settings, are discussed in Section 7.", "category": "Method_Mechanics"}, {"question": "Did the paper include extensive quantitative benchmarking for AD and ND tasks similar to the OOD detection tasks?", "answer": "True", "category": "Claim_Verification"}, {"question": "Has the paper expanded Table 7 to include per-class results for datasets like MNIST, CIFAR-10, and MVTecAD?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "tBRNC6YemY", "venue": "NEURIPS 2024", "paper_decision": "Accept (poster)", "title": "Gorilla: Large Language Model Connected with Massive APIs", "authors": ["Shishir G Patil", "Tianjun Zhang", "Xin Wang", "Joseph E. Gonzalez"], "abstract": "Large Language Models (LLMs) have seen an impressive wave of advances, with\nmodels now excelling in a variety of tasks, such as mathematical reasoning and\nprogram synthesis. However, their potential to effectively use tools via API calls\nremains unfulfilled. This is a challenging task even for today’s state-of-the-art\nLLMs such as GPT-4 largely due to their unawareness of what APIs are available\nand how to use them in a frequently updated tool set. We develop Gorilla, a\nfinetuned LLaMA model that surpasses the performance of GPT-4 on writing API\ncalls. Trained with the novel Retriever Aware Training (RAT), when combined\nwith a document retriever, Gorilla demonstrates a strong capability to adapt to\ntest-time document changes, allowing flexible user updates or version changes.\nIt also substantially mitigates the issue of hallucination, commonly encountered\nwhen prompting LLMs directly. To evaluate the model’s ability, we introduce\nAPIBench, a comprehensive dataset consisting of HuggingFace, TorchHub, and\nTensorHub APIs. The successful integration of the retrieval system with Gorilla\ndemonstrates the potential for LLMs to use tools more accurately, keep up with\nfrequently updated documentation, and consequently increase the reliability and\napplicability of their outputs. Gorilla’s code, model, data, and demo are available\nat: https://gorilla.cs.berkeley.edu", "keywords": "LLM;Tool Use;APIs;Function Calling", "qa_pairs": [{"question": "How do authors ensure that the hand-generated instruction seeds used for GPT-4's self-instruct generation meet real-world scenarios and address real-world difficulties?", "answer": "For self-instruct generation, authors provide three in-context examples and reference API documentation, tasking the model with generating real-world use cases that call upon the API. Authors instruct the model to avoid using API names or hints. Authors constructed 6 Instruction-API pairs for each of the 3 model hubs, then manually verified and audited each generation to ensure a high-quality dataset. Additionally, the Gorilla models have been downloaded over 10,000 times, indicating their real-world applicability.\n\nHere are some examples randomly chosen to demonstrate the diversity and versatility:\n\nExample 1 (q 638): 'I am an illustrator, I want to create an appealing image based on a text description for commercial purposes.'\n\nExample 2 (q 704): 'Assist me in finding the accurate information in a table related to the NYSE stock market.'\n\nExample 3 (q 867): 'We want to communicate product information to online customers in France. 'Introducing the new eco-friendly water bottle made of high-quality stainless steel with double-wall insulation to keep your drinks cool for 24 hours or hot for 12 hours.''", "category": "Method_Mechanics"}, {"question": "What are the details of the training data for Gorilla, including data statistics, construction methods, and how the training data differs from the evaluation sets of APIBench to avoid data contamination?", "answer": "The training data for Gorilla is constructed as follows: (1) From HuggingFace, 925 models were selected after filtering for quality and documentation. (2) From TensorFlow Hub, 626 models were selected after filtering. (3) From Torch Hub, all 95 models were included. Each API was manually verified for quality and correctness. Synthetic instruction data was generated using GPT-4-0613, guided by the self-instruct paradigm, with 18 hand-generated examples and 10 instruction-API pairs generated for each of the 1,645 APIs. The data was split uniformly into train and test sets. Gorilla was trained for 5 epochs with a learning rate of 2e-5, cosine decay, batch size 64, warm-up ratio 0.03, and a max sequence length of 2048. The evaluation metric ensures high-quality code by checking against ground truth, and human evaluation confirms the relationship between the metric and execution accuracy.", "category": "Experimental_Setup"}, {"question": "How does the performance vary across different domains, and what factors influence this variance?", "answer": "The performance varies across domains, with 'Computer Vision Object Detection' having an accuracy of 0.77 and 'NLP Text2Text Generation' having an accuracy of 0.68. The variance is impacted by the number of models in each category.", "category": "Experimental_Exposition"}, {"question": "How are the evaluation metrics (accuracy, hallucination, and error by selecting the wrong API call) calculated in the experimental results?", "answer": "Accuracy is calculated as the number of correct outputs divided by the total number of outputs. A correct output requires the model’s output to match an API in the API pool (using Abstract Syntax Tree subtree matching, including functional names and non-optional arguments) and the matched API’s function description to be equivalent to the ground truth. Hallucination is calculated as the number of outputs that do not match any API call in the pool divided by the total number of outputs. The sum of accuracy, hallucination, and other errors (e.g., syntactic errors) equals 1.", "category": "Experimental_Exposition"}, {"question": "How was the human evaluation of 78% accuracy in the AST as a Hallucination Metric section (line 251) conducted, and what were the key findings?", "answer": "The human evaluation was conducted by directly executing the code. The authors manually evaluated 100 randomly chosen LLM generations from their evaluation set. The accuracy using AST subtree matching was 78%, which matched the human evaluation accuracy in calling the correct API. All generations flagged as incorrect by AST were also flagged as incorrect in the manual evaluation. Additionally, 72% of the supporting code generated by Gorilla (e.g., installing dependencies(e.g., `pip install transformers[sentencepiece]`), setting environment variables) was successfully executed. The 6% discrepancy was due to errors external to the API in the supporting code, not semantic errors. This manual validation reinforced the efficiency of using AST as a robust offline metric.", "category": "Experimental_Exposition"}, {"question": "What are the licenses for the crawled APIs, TensorFlow, code samples, Pytorch Hub, and HuggingFace hubdocs?", "answer": "The licenses for the crawled APIs are Apache 2.0. TensorFlow is licensed under Creative Commons Attribution 4.0, code samples are under Apache 2.0, Pytorch Hub follows the Linux Foundation Policies, and HuggingFace hubdocs are on Apache 2.0.", "category": "Method_Disambiguation"}, {"question": "Can authors clarify the process of filtering and selecting models for evaluation from HuggingFace, TensorFlow Hub, and Torch Hub, and explain why the dataset for HuggingFace is not exhaustive?", "answer": "For HuggingFace, authors filtered models based on criteria such as poor documentation, lack of dependencies, and incomplete model card information, selecting the top 20 models from each domain, resulting in 925 models. For TensorFlow Hub, authors focused on the latest version (v2), which has 801 models, and after filtering, authors retained 626 models. Torch Hub provided an exhaustive set of 95 models. Each dataset was manually verified for quality and correctness, ensuring executable code. The HuggingFace dataset is not exhaustive because it represents a curated subset of the total models available on the platform.", "category": "Method_Mechanics"}, {"question": "What is the reason for the difficulty in using TorchHub compared to HuggingFace and TensorFlowHub, as indicated in Table 1?", "answer": "TorchHub APIs require pre-processing that is not templated like in HuggingFace, leading to confusion. For instance, loading Tacotron 2 from Nvidia involves a specific API call (`tacotron2_model = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_tacotron2', model_math='fp16')`), which can be confused with a similar pre-processing step (`utils = torch.hub.load('NVIDIA/DeepLearningExamples:torchhub', 'nvidia_tts_utils')`).", "category": "Method_Disambiguation"}, {"question": "How does the Gorilla approach differ from existing knowledge augmentation techniques used in other domains, and what novel contributions does it make to the field of connecting LLMs with APIs?", "answer": "The Gorilla approach differs from existing knowledge augmentation techniques by focusing on connecting LLMs with APIs. Its novel contributions include: 1. Developing and manually curating a high-quality dataset valuable to the community. 2. Introducing Retriever-Aware Training (RAT) to teach LLMs to use or ignore retrieved context, enhancing performance. 3. Proposing evaluation metrics, including the first technique to define and measure hallucination for APIs using Abstract Syntax Tree (AST), validated through human evaluation. 4. Studying realistic API calls with constraints, requiring reasoning about memory and accuracy budgets. 5. Measuring hallucination for popular API models.", "category": "Method_Comparison"}, {"question": "How does Gorilla's performance compare to DocPrompting, a specialized API-focused model, in terms of accuracy and hallucination reduction?", "answer": "Gorilla outperforms DocPrompting in both accuracy and hallucination reduction. Specifically, Gorilla achieves an accuracy of 71.68 compared to DocPrompting's 61.72, and reduces hallucination to 10.95 compared to DocPrompting's 17.36.\n |Accuracy ↑ | Accuracy ↑ | Hallucination ↓ | Hallucination ↓ |\n |---|---|---|---|\n |DocPrompting | **Gorilla** | DocPrompting | **Gorilla**| \n|61.72 |**71.68**| 17.36| **10.95**|", "category": "Method_Comparison"}, {"question": "What is the performance impact of varying the base model size, and how does the Retrieval Aware Training (RAT) fine-tuning recipe perform across different base models such as MPT-7B, Falcon-7B, and LLAMA-7B?", "answer": "The performance of the model was evaluated across three base models: MPT-7B (0.70), Falcon-7B (0.66), and LLAMA-7B (0.71) on the HF dataset. The results, included in Appendix Figure 13, demonstrate that the Retrieval Aware Training (RAT) fine-tuning recipe is robust to the underlying base model when all other factors are kept constant.", "category": "Experimental_Exposition"}, {"question": "How does Gorilla's Retriever Aware Training (RAT) enable it to generalize to entirely new APIs or domains not covered in the training data?", "answer": "Gorilla's Retriever Aware Training (RAT) enables it to generalize well to out-of-domain APIs. For example, when evaluated on 2000 APIs from diverse sets including hyperscalers (GCP, Azure, AWS), Postman, and RapidApi, Gorilla achieved an accuracy of 0.89, outperforming gpt-3.5-turbo-0125 (accuracy: 0.75). Examples include creating a charge in Stripe and retrieving dividend information from Yahoo Finance, which are entirely out-of-domain to ML APIs, demonstrating strong generalization capabilities.", "category": "Method_Mechanics"}, {"question": "How does the performance of fine-tuned models compare to Gorilla, particularly in terms of accuracy and hallucination, and why is Gorilla's approach considered more robust?", "answer": "Fine-tuned models like DocPrompting show lower accuracy (61.72) and higher hallucination (17.36) compared to Gorilla, which achieves higher accuracy (71.68) and lower hallucination (10.95). Gorilla's robustness stems from its Retrieval Aware Training (RAT) fine-tuning recipe, which is effective across different base models and adaptable to changes in API documentation without requiring complete re-training, unlike Toolformer.\n|Accuracy ↑ | Accuracy ↑ | Hallucination ↓ | Hallucination ↓ |\n|---|---|---|---|\n|DocPrompting | **Gorilla** | DocPrompting | **Gorilla**|\n|61.72 |**71.68**| 17.36| **10.95**|", "category": "Method_Comparison"}, {"question": "What evidence supports the statistical significance of Gorilla's Retriever Aware Training (RAT) in generalizing to out-of-domain APIs, and how does it compare to gpt-3.5-turbo-0125?", "answer": "To demonstrate the statistical significance of Gorilla's Retriever Aware Training (RAT), authors evaluated RAT-trained Gorilla on 2000 APIs from a diverse set, including hyperscalers (GCP, Azure, AWS), Postman, and RapidApi. Gorilla achieved an accuracy of 0.89, outperforming gpt-3.5-turbo-0125, which had an accuracy of 0.75. Examples of out-of-domain APIs tested include `stripe.Charge.create(amount={amount}, currency={currency}, source={source}, description={description})` for creating a charge in Stripe and `yfinance.Ticker({ticker}).dividends` for retrieving dividend information from Yahoo Finance, showcasing Gorilla's ability to generalize well.", "category": "Method_Comparison"}, {"question": "What are the most common types of errors encountered in the model's API call generations, and how do they vary across different retrieval settings?", "answer": "Common types of errors include model-name hallucination (e.g., `resnet-201` instead of `resnet-101`), hallucination in API call format (e.g., `torch.hub.load` instead of `torch.load`), and changes in directory structure (e.g., `'facebookresearch/pytorch_GAN_zoo:hub` to `facebook/pytorch_GAN_zoo:hub`). These error categories vary in frequency across different retrieval settings (0-shot, BM-25, oracle-retriever) but remain consistent across datasets. A collection of examples will be included in the appendix for further qualitative analysis.", "category": "Experimental_Exposition"}, {"question": "What are the common types of errors that differ between the categories of APIs, and how do their frequencies vary across different retrieval methods and datasets?", "answer": "Common types of errors include model-name hallucination (e.g., `resnet-201` instead of `resnet-101`), hallucination in API call format (e.g., `torch.hub.load` to `torch.load`), and changes in directory structure (e.g., `'facebookresearch/pytorch_GAN_zoo:hub` to `facebook/pytorch_GAN_zoo:hub`). The frequency of these error categories varies across retrieval methods (0-shot, BM-25, oracle-retriever) but remains steady across datasets. A collection of examples for qualitative analysis will be included in the appendix.", "category": "Experimental_Exposition"}, {"question": "How does Gorilla's performance compare to specific methods like DocPrompting and prompting techniques (0-shot and 3-shot) in terms of accuracy and hallucination when building software pipelines using LLM for tool calling?", "answer": "Gorilla demonstrates improved accuracy and reduced hallucination compared to DocPrompting, with accuracy increasing from 61.72 to 71.68 and hallucination decreasing from 17.36 to 10.95. Additionally, Gorilla shows enhanced performance across both 0-shot and 3-shot prompting techniques for GPT-3.5 and GPT-4, as detailed in Table 6 of the Appendix.\n|Accuracy ↑ | Accuracy ↑ | Hallucination ↓ | Hallucination ↓ |\n|---|---|---|---|\n|DocPrompting | **Gorilla** | DocPrompting | **Gorilla**|\n|61.72 |**71.68**| 17.36| **10.95**|", "category": "Method_Comparison"}, {"question": "What are the specific contributions of Gorilla, and how do they impact the community?", "answer": "Gorilla's contributions include: 1. Developing and manually curating a high-quality dataset valuable to the community. 2. Introducing Retriever-Aware Training (RAT), a novel approach that teaches the LLM to use or ignore retrieved context, enabling superior performance. 3. Proposing evaluation metrics, including the first technique to define and measure hallucination for APIs, validated through human evaluation. 4. Studying realistic API calls with constraints, requiring reasoning about memory and accuracy budgets. 5. Measuring hallucination for popular models for APIs. These contributions advance the field by improving LLM-API integration and providing tools for better evaluation and reasoning.", "category": "Method_Mechanics"}, {"question": "How do authors determine which API is better among similar ones with overlapping functions, and how do authors ensure consistent results for the same request?", "answer": "Authors determine the better API based on user-specified constraints, such as performance metrics (e.g., top-1 accuracy on ImageNet). If constraints are not specified, all suitable APIs are considered equally valid.  For example, if the user requests an object detection model, any and all object detection APIs are the right answers. However, if a user specifies a constraint either explicitly or implicitly (e.g., they want the model to have a top-1 accuracy of xx on ImageNet), then the set of APIs narrows down and a subset of them are 'better' than other similar ones. Consistent results are ensured by controlling the dataset so that multiple valid APIs are marked as accurate for the same request.", "category": "Method_Mechanics"}, {"question": "What is the reason for Gorilla consistently producing large errors in selecting incorrect APIs, as shown in Table 1, and do the results align with authors' hypothesis?", "answer": "The large errors occur because the model, despite being taught domain knowledge through Retriever Aware Training, still struggles to accurately identify user intent, leading to the generation of APIs for the wrong task (not 0, but lesser than baselines). This aligns with the hypothesis that the training reduces but does not eliminate hallucinations compared to baselines.", "category": "Experimental_Analysis"}, {"question": "Did the rebuttal specify the number of hand-generated instruction-API pairs and the method of data division into train and test splits?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the Gorilla training data explicitly compared to the evaluation sets of APIBench to address potential data contamination concerns?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is Llama-3 used as the foundation model in the study?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the licenses of the crawled APIs mentioned and confirmed to be permissible under specific terms such as Apache 2.0?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the rebuttal provide detailed statistics and construction methods for the Gorilla training data?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the rebuttal clarify that each API's dataset was manually checked for quality and correctness to ensure high-quality code?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is Llama-2 used as the foundation model in the study?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the study evaluate the performance of the proposed method across different base models, including Llama-2?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "4NJBV6Wp0h", "venue": "NEURIPS 2024", "paper_decision": "Accept (oral)", "title": "LLM Evaluators Recognize and Favor Their Own Generations", "authors": ["Arjun Panickssery", "Samuel R. Bowman", "Shi Feng"], "abstract": "Self-evaluation using large language models (LLMs) has proven valuable not only in benchmarking but also methods like reward modeling, constitutional AI, and self-refinement. But new biases are introduced due to the same LLM acting as both the evaluator and the evaluatee. One such bias is self-preference, where an LLM evaluator scores its own outputs higher than others’ while human annotators consider them of equal quality. But do LLMs actually recognize their own outputs when they give those texts higher scores, or is it just a coincidence? In this paper, we investigate if self-recognition capability contributes to self-preference. We discover that, out of the box, LLMs such as GPT-4 and Llama 2 have non-trivial accuracy at distinguishing themselves from other LLMs and humans. By finetuning LLMs, we discover a linear correlation between self-recognition capability and the strength of self-preference bias; using controlled experiments, we show that the causal explanation resists straightforward confounders. We discuss how self-recognition can interfere with unbiased evaluations and AI safety more generally.", "keywords": "LLMs;evaluations;benchmarking;situational-awareness", "qa_pairs": [{"question": "What statistical test was used to determine the significance of the difference between LLMs' self-preference and human annotators' preferences, and what were the results for GPT-4 and GPT-3.5?", "answer": "A Chi-squared test of statistical significance was used to compare the difference between self-preference (Figure 4) and human preference (Section 2.5). The results showed a significant difference with a p-value much less than 0.001 for GPT-4 (compared to GPT-3.5 and Llama 2) even before fine-tuning, and a significant difference with a p-value of approximately 0.007 for GPT-3.5 after fine-tuning on 10 examples for self-recognition ability.", "category": "Experimental_Exposition"}, {"question": "What statistical methods were used to validate the claims in the paper, and what are the results?", "answer": "Chi-Squared tests were performed to confirm the statistical significance of the following claims (p-value << 0.001): 1. LLMs demonstrate preference for their own generations disproportionately compared to humans 2. LLMs demonstrate significantly higher self-preference after fine-tuning for self-recognition 3. There is a significantly higher increase in self-recognition and self-preference from fine-tuning for self-recognition compared to fine-tuning on the control tasks.", "category": "Experimental_Exposition"}, {"question": "Could the fine-tuning process allow the model to learn a shortcut that preferentially selects the labeled option, thereby acting as a confounder in the self-preference evaluation?", "answer": "No, the fine-tuning examples are distinct from the evaluation examples, and the ordering of options is randomized and balanced, preventing trivial explanations for increased self-preference. The phenomenon of the model recognizing patterns from its own generation in fine-tuning aligns with authors' causal hypothesis that self-preference is due to self-recognition. Authors control for confounders between self-recognition and self-preference but not for shortcuts used by the model. The suggested experiment is already included in Section 3.4 and did not significantly affect self-recognition or self-preference strength.", "category": "Experimental_Analysis"}, {"question": "Did authors experiment with labeling the source of each piece of text and re-evaluating self-preference, and what were the results?", "answer": "Yes, the paper conducted this experiment in Section 3.5, which confirmed an increase in self-preference. the paper also used intentionally incorrect labels to test the model's ability to recognize its own generation from textual features, and found that GPT-4 still strongly preferred its own generation even when lied to about the source.", "category": "Experimental_Exposition"}, {"question": "Do larger language models exhibit better self-recognition capabilities, and what hypotheses can explain this trend as observed in Figure 1?", "answer": "Larger and more capable models appear better at self-recognition. One hypothesis is that their output probabilities are more focused around their own generations. Evidence suggests that detection is not more difficult for larger LLMs, and their perplexity is more sensitive to whether the text is generated by the LLM. If self-recognition relies on similar mechanisms as detection methods, this might explain the improved self-recognition in larger LLMs.", "category": "Experimental_Exposition"}, {"question": "How did the authors address the concern regarding the lack of variety and breadth in the experiments to thoroughly explore the randomness of the models?", "answer": "The authors maximized the coverage in the experiments by including task diversity (extractive and abstractive summarization), various evaluation formats (pairwise and individual), different fine-tuning setups (in- and out-of-domain, number of training examples), control tasks to rule out confounders, and varied ordering of options in pairwise evaluations, within the constraints of the compute budget.", "category": "Experimental_Setup"}]}
{"id": "Vi8AepAXGy", "venue": "NEURIPS 2024", "paper_decision": "Accept (oral)", "title": "Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs", "authors": ["Shengbang Tong", "Ellis L Brown II", "Penghao Wu", "Sanghyun Woo", "ADITHYA JAIRAM IYER", "Sai Charitha Akula", "Shusheng Yang", "Jihan Yang", "Manoj Middepogu", "Ziteng Wang", "Xichen Pan", "Rob Fergus", "Yann LeCun", "Saining Xie"], "abstract": "We introduce Cambrian-1, a family of multimodal LLMs (MLLMs) designed with a vision-centric approach. While stronger language models can enhance multimodal capabilities, the design choices for vision components are often insufficiently explored and disconnected from visual representation learning research. This gap hinders accurate sensory grounding in real-world scenarios. Our study uses LLMs and visual instruction tuning as an interface to evaluate various visual representations, offering new insights into different models and architectures—self-supervised, strongly supervised, or combinations thereof—based on experiments with over 15 vision models. We critically examine existing MLLM benchmarks, addressing the difficulties involved in consolidating and interpreting results from various tasks. To further improve visual grounding, we propose spatial vision aggregator (SVA), a dynamic and spatially-aware connector that integrates vision features with LLMs while reducing the number of tokens. Additionally, we discuss the curation of high-quality visual instruction-tuning data from publicly available sources, emphasizing the importance of distribution balancing. Collectively, Cambrian-1 not only achieves state-of-the-art performances but also serves as a comprehensive, open cookbook for instruction-tuned MLLMs. We provide model weights, code, supporting tools, datasets, and detailed instruction-tuning and evaluation recipes. We hope our release will inspire and accelerate advancements in multimodal systems and visual representation learning.", "keywords": "Multimodal LLM;Visual Representation Learning;Evaluation Protocol;Data Mix;Open Science", "qa_pairs": [{"question": "How does the performance of SVA change with increasing values of D and G, and at what point does it saturate?", "answer": "Performance improves with increasing values of D and G, and it saturates when D exceeds 4 or G exceeds 3.\n|D|G|OCR & Chart|\n|:-:|:-:|:-:|\n|2|1|52.1|\n|3|1|52.4|\n|4|1|52.8|\n|5|1|52.1|\n|3|1|52.4|\n|3|2|52.6|\n|3|3|53.1|\n|3|4|52.8|", "category": "Experimental_Exposition"}, {"question": "Why does increasing the value t not continuously enhance performance, and how does data quality influence the results shown in Table 2 and Table 3?", "answer": "Data quality matters more than quantity, as shown in §5.2. Intermediate t values result in better performance, consistent with observations in CLIP and MetaCLIP data curation. The higher-quality 7M subset outperforms the 10M dataset due to better balancing of dataset sources and adjusting their relative sizes.", "category": "Experimental_Exposition"}, {"question": "What statistical analysis has been conducted on the response length, difficulty, and diversity distribution of the instruction data for Cambrian-7M and -10M?", "answer": "Further analysis has been conducted on Cambrian-7M and -10M regarding data composition, length of questions, length of responses, and number of rounds in the instruction tuning data.  Cambrian-7M data are distributed similarly to the best data ratio found in the data ratio experiment (Fig. 10).", "category": "Experimental_Exposition"}, {"question": "How does the performance of LLaVA trained with LLaMA-3-8b and Cambrian-7M data compare to LLaVA-Next and Cambrian-8B across various tasks using 576 visual tokens?", "answer": "LLaVA trained with LLaMA-3-8b and Cambrian-7M data, using 576 visual tokens, performs comparably or better than LLaVA-Next in general, knowledge, and vision tasks. It achieves a general score of 72.0, a knowledge score of 58.1, an OCR & Chart score of 54.3, and a Vision-Centric score of 55.6. Cambrian-8B, also using 576 tokens, outperforms both with scores of 74.4, 60.1, 66.2, and 60.3 respectively in the same tasks. Adding SVA further improves LLaVA's performance, especially in OCR & Chart tasks.\n|Data|# Vis Token|General|Knowledge|OCR & Chart|Vision-Centric|\n|:--|--:|--:|--:|--:|--:|\n|LLaVA-Next|2880|72.5|55.6|61.0|55.0|\n|LLaVA w/ Cambrian Data|576|72.0|58.1|54.3|55.6|\n|Cambrian-8B|576|74.4|60.1|66.2|60.3|", "category": "Method_Comparison"}, {"question": "Does training the visual encoder (unfreezing) outperform freezing it in all tasks, and how does the convergence speed of end-to-end training compare to two-stage training?", "answer": "Unfreezing most visual encoders outperforms freezing in most tasks, as shown in Fig. 13, which visualizes the % change from Frozen to Unfrozen for each model on each benchmark (full results are in Tables 9 & 10 &11). Regarding convergence speed, unfreezing is approximately 50-55% slower during fine-tuning with fixed compute.", "category": "Experimental_Analysis"}, {"question": "Why does integrating more vision encoders not always result in higher performance improvements, as observed in Table 11 with models like MMB, VStar, and MMEP?", "answer": "Integrating more vision encoders does not always lead to higher performance on every benchmark because different vision encoders have varying strengths and weaknesses. These differences result in combinations of vision encoders inheriting a mix of these strengths and weaknesses, which may not uniformly enhance performance across all benchmarks.", "category": "Experimental_Analysis"}, {"question": "How transferable are the findings in the paper to 'native' multimodal LLMs, and what are the authors' thoughts on these models?", "answer": "Authors find native Multimodal LLMs that do not use a pre-trained vision encoder, like GPT-4o or Reka, to be an intriguing and promising approach. These native models have only recently been explored and are predominantly developed by proprietary companies such as OpenAI. As a result, their actual implementation, architecture, and training methods remain largely undisclosed. Additionally, there is insufficient evidence to assert that native MLLMs can overcome the limitations of current MLLMs, such as their visual deficiencies. On the other hand, vision-only representation learning itself is a significant and meaningful objective. Authors' study connects this goal with multimodal learning, providing scientific insights that complement future advancements in multimodal systems, whether they are native or not.\nAuthors note that one major downside of native MLLMs is that they require much more data and computational resources, as they do not rely on the knowledge embedded in pretrained encoders.\n\nThe findings in the paper are highly transferable to 'native' multimodal LLMs. The authors believe that their findings regarding Data, Connectors, Evaluation, and Vision Backbones will continue to hold and guide the development of future models. For example, the data curation studies can be useful for supervised fine-tuning native MLLMs, the SVA module could be a competitive candidate for resolving high-resolution feature conflicts, the evaluation study can help native MLLMs better assess their visual capability, and the insights from comparing visual representations can guide developers in designing architecture, data, and methods for training native MLLMs, training data (e.g., CLIP vs. SSL), training methods (Encoding vs. Generative), network architecture (ViT vs. ConvNext), and image resolutions. The authors find native multimodal LLMs, which do not use pre-trained vision encoders, to be a promising approach, though they note that these models require much more data and computational resources and their actual implementation and architecture are largely undisclosed due to their proprietary nature.", "category": "Experimental_Analysis"}, {"question": "Why does the paper limit its experiments to a LLaVa-like formulation of MLLMs built upon pretrained LLMs, rather than exploring native multimodal LLMs like GPT-4o and Reka, which have different architectures and training methodologies?", "answer": "Native multimodal LLMs, such as GPT-4o and Reka, are an intriguing approach but are predominantly developed by proprietary companies, making their implementation, architecture, and training methods largely undisclosed. Additionally, there is insufficient evidence that native MLLMs can overcome the limitations of current MLLMs, such as visual deficiencies. Furthermore, native MLLMs require significantly more data and computational resources, as they do not rely on pretrained encoders. Authors' study focuses on connecting vision-only representation learning with multimodal learning, providing insights that complement future advancements in multimodal systems, whether native or not. Authors' findings on Data, Connectors, Evaluation, and Vision Backbones can be transferred to guide the development of native MLLMs.\n\nData: The pool of instruction tuning data and authors' data curation studies can be very useful for supervised fine-tuning “native” MLLMs. Training native MLLMs is likely to still consist of pretraining, supervised fine-tuning, and RLHF. The data collection and curation insights can play an important role in both the pretraining and supervised fine-tuning stages.\nConnector: The SVA module authors proposed can be part of, or inspiration for, future native MLLM designs. The conflict between high-resolution features and a constrained number of tokens is likely to continue in native MLLMs. Therefore, the SVA module could be a competitive candidate for resolving this issue in such native MLLMs.\nEvaluation: Authors' study on the “Multimodality” of benchmarks can help native MLLMs better assess their visual capability. The categorization of benchmarks also provides more organized and interpretable evaluation protocols for future works, especially in vision-centric domains.\nVision Backbones: Authors' study compares current visual representations, uncovering insights about training data (e.g., CLIP vs. SSL), training methods (Encoding vs. Generative), network architecture (ViT vs. ConvNext), and image resolutions. These insights can better guide developers when designing architecture, data, and methods for training native MLLMs.", "category": "Experimental_Analysis"}, {"question": "How does the proposed SVA module compare to C/D-Abstractor in terms of performance and functionality, particularly in tasks involving token reduction and spatial inductive bias?", "answer": "The SVA module is distinct from C/D-Abstractor in its ability to dynamically combine visual features from multiple vision models with varying resolutions. To isolate the effect of spatial inductive bias, experiments were conducted using OpenAI CLIP as the vision model, compressing its original 576 tokens to 36 tokens with SVA and other connectors. Three baselines were included: Direct interpolation + MLP, C-Abstractor, and LDPNetV2Projector. Compared with the simple MLP baseline, C-Abstractor performs better on General and Vision-Centric tasks but inferior on Knowledge and OCR & Chart tasks. LDPNetV2 performs similarly to the MLP baseline. Results show that SVA consistently outperforms the baselines across all categories, particularly in OCR & Chart and Vision-Centric tasks. This superior performance is attributed to SVA's local attention mechanism, which applies the same parameters to all positions in the grid, resulting in more supervision and data efficiency compared to C-Abstractor.\n\n | Method            | General | Knowledge | OCR & Chart | Vision-Centric | \n|-------------------|:-------:|:---------:|:-----------:|:--------------:|\n | Interpolate + MLP |  63.4   |   43.8    |    28.1     |      43.7      |\n | C-Abstractor      |  64.4   |   42.8    |    26.1     |      44.3      |\n | LDPNetV2          |  62.5   |   43.9    |    28.7     |      43.9      | \n| SVA               |**65.5** | **44.5**  |  **31.4**   |    **46.9**    | ", "category": "Method_Comparison"}, {"question": "How does the systematic study of Multimodal Large Language Models (MLLMs) in this work address the contradictory results and overlapping findings from previous isolated studies?", "answer": "The systematic study in this work combines different components of MLLMs, aiming to draw more robust and reliable conclusions and clarify contradictory results from previous studies. For example, it compares 15+ vision models and curates the largest open-source instruction tuning datasets, providing insights into both MLLM and visual representation learning communities. These efforts aim to narrow the gap between open-source and proprietary models.", "category": "Method_Mechanics"}, {"question": "How do the results in Table 11 support the claim in Finding 7 that performance improves with the vision encoder ensemble, given that SigLIP+DINOv2 performs worse than sole SigLIP?", "answer": "The conclusion in Finding 7 is based on multiple experiments, not just a single data point. Compared to ensembling only CLIP models, adding the DINOv2 model improves performance, particularly on vision-centric benchmarks like RealWorldQA and CV-Bench.", "category": "Experimental_Analysis"}, {"question": "How does the study address privacy issues related to the construction of the new dataset based on web search?", "answer": "The study addresses privacy issues by collecting data from licensed websites, specifically Wikipedia（https://en.wikipedia.org/wiki/Wikipedia:Copyrights）, which is licensed under CC BY-SA, and attributing the data source in §4.1 and Appx. E. Additionally, the data collection pipeline is fully open-sourced in Appx. E, allowing for thorough inspection and verification to ensure compliance with copyright and data privacy regulations. ", "category": "Method_Mechanics"}, {"question": "What are the key differences between the proposed work and previous studies in the field of Multimodal Large Language Models (MLLMs), particularly in terms of technical novelty, experimental scope, and vision model aggregation strategies?", "answer": "The proposed work differs from previous studies in the following ways: 1) More Comprehensive Experiments: The study considers more vision backbones and organized benchmarks, yielding new insights such as the benefits of high-resolution ConvNets for OCR & Chart tasks and the potential of self-supervised learning models to compete with language-supervised ones. It also reveals properties of different CLIP models beyond ImageNet accuracy, providing valuable insights for MLLM and Vision model developers. 2) Findings on High-Resolution Encoders: The work pinpoints the potential of using ConvNets, such as the ConvNext OpenCLIP model, to efficiently process high-resolution images. 3) New Vision Model Aggregation Strategy: The study explores both which models to combine and strategies for combining them, introducing the SVA approach that preserves resolution, maintains spatial inductive bias, and uses a fixed number of visual tokens. ", "category": "Method_Comparison"}, {"question": "How does the performance of the proposed method, using 576 fixed tokens and the SVA module, compare to LLaVA-1.6 (LLaVA-Next) with 2880 dynamic tokens, particularly in General, Knowledge, OCR & Chart, and Vision-Centric tasks?", "answer": "The proposed method, using 576 fixed tokens and the SVA module, matches or outperforms LLaVA-1.6 (LLaVA-Next) with 2880 dynamic tokens in General, Knowledge, and Vision-Centric tasks. Specifically, LLaVA with Cambrian data (576 tokens) achieves 72.0 in General, 58.1 in Knowledge, 54.3 in OCR & Chart, and 55.6 in Vision-Centric tasks, compared to LLaVA-Next's 72.5, 55.6, 61.0, and 55.0, respectively. Adding the SVA module further improves performance, especially in OCR & Chart tasks, with Cambrian-8B achieving 74.4 in General, 60.1 in Knowledge, 66.2 in OCR & Chart, and 60.3 in Vision-Centric tasks.\n\n |Data|# of Vis Tokens|General|Knowledge|OCR & Chart|Vision-Centric| \n|:--|--:|--:|--:|--:|--:|\n |LLaVA-Next|2880|72.5|55.6|61.0|55.0|\n |LLaVA w/ Cambrian Data|576|72.0|58.1|54.3|55.6|\n |Cambrian-8B|576|74.4|60.1|66.2|60.3|", "category": "Method_Comparison"}, {"question": "How does authors’ work differ from previous studies that have already verified findings such as the use of high-resolution visual encoders, and what new insights does authors' systematic study provide?", "answer": "Authors' work differs from previous studies by undertaking a systematic study that isolates and examines the numerous moving parts of Multimodal Large Language Models, leading to more long-standing conclusions. Specifically,authors conduct more comprehensive experiments, considering more vision backbones and organized benchmarks, which yield new insights. For example, authors find that high-resolution ConvNets benefit OCR & Chart tasks, and self-supervised learning models can potentially compete with language-supervised ones. Additionally, authors' experiments reveal properties of different CLIP models (e.g., SigLIP, EVA-CLIP) beyond ImageNet accuracy, providing valuable insights for both MLLM and Vision model developers. For instance, authors observe that EVA-CLIP performs well in most domains but struggles with Chart tasks, highlighting the need for CLIP developers to focus more on OCR & Chart data collection during training. Furthermore, while concurrent works emphasize the importance of high-resolution features, authors take a step further to pinpoint the potential of using ConvNets, such as the ConvNext OpenCLIP model, to efficiently and effectively process high-resolution images.", "category": "Method_Disambiguation"}, {"question": "What is the importance of different tokens across datasets (GQA, DocVQA, ScienceQA) in the SVA module, and how does increasing the size of instruction tuning data affect model performance?", "answer": "The importance of visual features from different vision models varies across datasets. For GQA, the contribution of SigLIP, CLIP, DINOv2, and ConvNext is relatively uniform, in part due to the similar characteristics of SigLIP and CLIP. For DocVQA, SigLIP and ConvNext have increased influence due to text-heavy and high-resolution images. For ScienceQA, SigLIP's contribution increases further while DINOv2's decreases.\n|Model|GQA|DocVQA|ScienceQA| \n|:--|--:|--:|--:|\n |SigLIP|29.7%|31.1%|35.2%|\n |CLIP|18.5%|13.4%|16.3%|\n |DINOv2|24.1%|11.0%|17.6%|\n |ConvNext|27.7%|44.5%|30.9%| \n\n Increasing alignment data size improves performance across all benchmarks, and increasing instruction tuning data size leads to notable overall improvement, particularly for Knowledge, OCR & Chart, and Vision-Centric tasks.\n\n|                         |  General | Knowledge | OCR & Chart | Vision-Centric |\n| --------------------------------- | :------: | :-------: | :---------: | :------------: |\n| 1.2M alignment + 0.7M instruction |   72.3   |    54.8   |     58.3    |      57.2      |\n| 2.5M alignment + 0.7M instruction |   72.7   |    55.8   |     58.9    |      58.3      |\n| 2.5M alignment + 7M instruction   | **74.4** |  **60.1** |   **66.2**  |    **60.3**    |\n", "category": "Experimental_Exposition"}, {"question": "How was potential data leakage prevented in the instruction tuning data collected from the open web, and what statistical analysis of the data categories will be provided?", "answer": "Data leakage was prevented by carefully selecting only the training set of data sources from existing open-source works, focusing exclusively on Wikipedia, which does not overlap with the benchmarks. section 5 reports the number of test images in each benchmark present in the final Cambrian-10M dataset by checking image hashes.", "category": "Method_Mechanics"}, {"question": "Is the Internet Data Engine in the project limited to using Wikipedia to avoid overlap with benchmarks?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the instruction tuning data collected from the open web, potentially raising data leakage concerns in the paper?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "k0qTnbQxzR", "venue": "NEURIPS 2024", "paper_decision": "Reject", "title": "CogCoM: Train Large Vision-Language Models Diving into Details  through\u000b Chain of Manipulations", "authors": ["Ji Qi", "Ming Ding", "Weihan Wang", "Yushi Bai", "Qingsong Lv", "Wenyi Hong", "Bin Xu", "Lei Hou", "Juanzi Li", "Yuxiao Dong", "Jie Tang"], "abstract": "Vision-Language Models (VLMs) have demonstrated their broad effectiveness thanks to extensive training in aligning visual instructions to responses. However, such training of conclusive alignment leads models to ignore essential visual reasoning, further resulting in failures in meticulous visual problems and unfaithful responses. Drawing inspiration from human cognition in solving visual problems (e.g., marking, zoom in), this paper introduces Chain of Manipulations, a mechanism that enables VLMs to solve problems step-by-step with evidence. After training, models can solve various visual problems by eliciting intrinsic manipulations (e.g., grounding, zoom in) and results (e.g., boxes, image) actively without involving external tools, while also allowing users to trace error causes. We study the roadmap to implement this mechanism, including (1) a flexible design of manipulations upon extensive analysis, (2) an efficient automated data generation pipeline, (3) a compatible VLM architecture capable of multi-turn multi-image, and (4) a model training process for versatile capabilities.  With the design, we also manually annotate 6K high-quality samples for the challenging graphical mathematical problems. Our trained model, CogCoM, equipped with this mechanism with 17B parameters achieves state-of-the-art performance across 9 benchmarks from 4 categories, demonstrating the effectiveness while preserving the interpretability. Our code, model weights, and collected data will be publicly available.", "keywords": "Vision-Language Models;Multimodal Large Language Models;Visual Reasoning;Chain of Manipulations", "qa_pairs": [{"question": "Have the authors evaluated their model on more modern and challenging VQA benchmarks such as MMBench, MathVista, and SeedBench, and if so, what were the performance results compared to LLaVA-1.5?", "answer": "Yes, the authors have evaluated their model on MMBench, SEDD-Bench, and MathVista. The model achieved stronger performance, outperforming LLaVA-1.5 by 10.2, 4.9, and 8.1 percentage points on these benchmarks, respectively. The results are listed in Table 2 of the supplementary PDF file.", "category": "Method_Comparison"}, {"question": "How does authors’ work compare to ViperGPT and V* in terms of approach and focus, particularly in addressing complex visual problems?", "answer": "ViperGPT shares the basic idea of decomposing complex visual problems into reasoning steps but focuses on building a training-free framework combining external VLMs using a code LLM. In contrast, the paper's work trains an end-to-end VLM to enhance multimodal reasoning ability. V* also aims to solve VQA problems by progressively acquiring visual cues using a two-part framework: a VQALLM model lists visual objects needed, followed by a dedicated search model to acquire them. The paper's approach, however, uses one model to actively perform reasoning and identify the most useful visual information, potentially enabling it to solve more complex reasoning tasks, such as challenging geometric math problems.", "category": "Method_Comparison"}, {"question": "How does the proposed CoM training method differ from the DualFocus approach in terms of methodology and emphasis on reasoning processes?", "answer": "The CoM training method differs from DualFocus by emphasizing a single reasoning process for answering questions and marking images to solve complex problems, whereas DualFocus constructs a training dataset with intermediate cues (bounding boxes) and uses a two-stage training based on accurate localization.", "category": "Method_Comparison"}, {"question": "Why are the performance improvements of CogCom compared to CogVLM relatively limited, and could the low success rate (around 55% in Figure 6) be a contributing factor?", "answer": "The limited improvements of CogCom over CogVLM on Grounding and General Multimodal benchmarks are due to the paper's focus on enhancing reasoning capabilities in VLMs. However, on evaluation sets requiring logical reasoning (GQA), complex counting (TallyVQA), and detailed recognition (TxtVQA, ST-VQA), CogCom achieves improvements of 6.5, 10.9, 1.04, and 9.0 percentage points, respectively. Mistakes in reasoning paths are a major factor affecting performance on general multimodal benchmarks.", "category": "Experimental_Analysis"}, {"question": "What strategies are proposed or discussed to mitigate the negative effects of planned pathway failures in reasoning paths, and how do they compare to the approach used in DualFocus?", "answer": "The paper proposes several strategies to mitigate reasoning path failures: 1) Incorporating DualFocus's PPL-based method to switch between direct-answering and reasoning-before-answering modes. 2) Using explicit launching prompts and training randomness to allow user-initiated and model-initiated mode switching. 3) Employing Reinforcement Learning (RL) to penalize negative reasoning paths during training, although RL introduces unstable training and low efficiency. 4) Humans solve difficult questions with a period of thinking before outputting the answers. As LLMs/VLMs generate outputs immediately following the question prompt, the CoT mechanism serves as an effective substitute for thinking. Exploring backtracking mechanisms to correct reasoning paths, though initial experiments show high time-complexity and marginal benefits. These strategies are compared to DualFocus, which relies on PPL-based mode switching without explicit prompts or training randomness.", "category": "Method_Comparison"}, {"question": "What is the purpose of including 6K high-quality manually annotated math samples in the CoM dataset, and are there any test results demonstrating their impact on the model's performance?", "answer": "The 6K annotated math samples in the CoM dataset were included to advance the development of VLMs in solving geometry math problems, a challenging task for current models. Experiments conducted on MathVista showed that the model trained on these samples achieved an improvement of 0.90 percentage points. The data aims to enhance multimodal reasoning processes similar to human methods, such as drawing auxiliary lines, contributing to the field's potential development.", "category": "Experimental_Exposition"}, {"question": "What evidence or test results (qualitative or quantitative) support the claim that CogCoM is capable of multi-image multi-turn understanding?", "answer": "CogCoM solves visual reasoning problems through multiple rounds of interaction with an input image, such as marking auxiliary lines, and re-inputting the resulting image at each new round in an end-to-end manner (e.g., a marked image with auxiliary lines). For example, in geometry math problem-solving, the model undergoes multiple rounds of reasoning, drawing auxiliary lines, and cropping or zooming in on specific regions, which exemplifies its multi-turn, multi-image capability. This process is analogous to the human thinking process and differentiates it from existing methods.", "category": "Experimental_Exposition"}, {"question": "What are the detailed statistics regarding the maximum number of turns the model can accept and the average number of turns involving multiple images in the training and testing datasets?", "answer": "The training data from TextVQA and ST-VQA has an average of 1.42 turns and a maximum of 7 turns, with a restriction to 4 turns during training to prevent OOM. The test set of TextVQA produced an average of 1.54 turns involving multiple images. Not all images require manipulations like zooming; some can be answered through reasoning or direct observation.", "category": "Experimental_Setup"}, {"question": "How does this work compare to LLaVA-Plus and other agentic LMMs in terms of methodology and focus?", "answer": "In comparison to LLaVA-Plus[1], which focuses on training VLMs to invoke external tools to solve tasks, this work emphasizes stimulating the model's inherent abilities to solve problems in an end-to-end manner through active reasoning. While LLaVA-Plus constructs an instruction-tuning dataset with tool use examples, this work offers advantages in enhancing the model's reasoning abilities and reducing the time complexity of reasoning, despite being disadvantaged in producing pixel-level masks.\n[1] https://arxiv.org/abs/2311.05437", "category": "Method_Comparison"}, {"question": "What mechanisms are in place to ensure the quality of the generated data and address inaccuracies introduced by the visual models (GroundingDino, PaddleOCR) during data generation?", "answer": "During the collection of the 70K CoM training data, wrongly recognized data (referred to as negative paths) are discarded because they cannot terminate at the golden answer node during DFS traversal. This filtering strategy ensures that only correct data capable of reaching the golden answer (positive paths) is included in the training dataset.", "category": "Experimental_Setup"}, {"question": "What is the impact of incorporating CoM data on the model's performance?", "answer": "The model benefits from the CoM corpus, which contains informative visual reasoning evidence, achieving improvements of 6.6, 0.2, and 0.9 percentage points on TextVQA, MMVet, and MathVista benchmarks, respectively, as shown in Table 3 of the supplementary PDF file.", "category": "Experimental_Exposition"}]}
{"id": "aKkAwZB6JV", "venue": "COLM 2024", "paper_decision": "Accept", "title": "Zephyr: Direct Distillation of LM Alignment", "authors": ["Lewis Tunstall", "Edward Emanuel Beeching", "Nathan Lambert", "Nazneen Rajani", "Kashif Rasul", "Younes Belkada", "Shengyi Huang", "Leandro Von Werra", "Clémentine Fourrier", "Nathan Habib", "Nathan Sarrazin", "Omar Sanseviero", "Alexander M Rush", "Thomas Wolf"], "abstract": "We aim to produce a smaller language model that is aligned to user intent. Previous research has shown that applying distilled supervised fine-tuning (dSFT) on larger models significantly improves task accuracy; however, these models are unaligned, i.e. they do not respond well to natural prompts. To distill this property, we experiment with the use of preference data from AI Feedback (AIF). Starting from a dataset of outputs ranked by a teacher model, we apply distilled direct preference optimization (dDPO) to learn a chat model with significantly improved intent alignment. The approach requires only a few hours of training without any additional sampling during fine-tuning. The final result, Zephyr-7B, set a new state-of-the-art on chat benchmarks for 7B parameter models, and requires no human annotation. In particular, results on MT-Bench show that Zephyr-7B surpassed Llama2-Chat-70B, at the time the best open-access RLHF-based model.", "keywords": "A;I; ;f;e;e;d;b;a;c;k;,; ;o;p;e;n; ;L;L;M;s;,; ;R;L;H;F", "qa_pairs": [{"question": "What explains the substantial increase in academic benchmark performance shown in Table 2, given that the paper primarily focuses on chat performance through imitation learning?", "answer": "The distinction between chat and academic tasks is less clear-cut than the terminology suggests. The UltraFeedback dataset includes completions from prompts that are more academic than chat. By prioritizing GPT-4's preferred answers over less accurate ones, the model receives relevant signals for academic tasks as well.", "category": "Experimental_Analysis"}, {"question": "What is the main technical contribution of this paper, and how does it differ from existing methods like SFT and DPO?", "answer": "The main technical contribution of this paper is to integrate existing components like Mistral, UltraChat, UltraFeedback, and DPO to create the first open-model with intent aligned behavior. While the DPO algorithm was known, it had not been proven on large datasets. This paper demonstrates that DPO is sufficient for achieving intent alignment by comparing it across multiple benchmarks, showing the necessity of the DPO stage for this outcome.", "category": "Method_Comparison"}, {"question": "What are the estimated costs involved in creating Zephyr, specifically for dataset construction using LLMs in dSFT, generating answers with multiple models in AIF, and ranking responses using GPT-4?", "answer": "The estimated cost of dataset construction is 1.5k for judging with GPT-4, and the cost of constructing roughly 250k alternative completions is approximately 300 for 7 hours on 8xH100. These aspects are less expensive than collecting high-quality annotations at scale. Imitation learning / Academic benchmarks. The answer seems to be that the distinction between chat / academic is less clear cut than this terminology suggests. For example the UltraFeedback dataset includes completions from many prompts that are more academic than chat. By weighting the preference of GPT-4 preferred answer over a most likely uncorrect response, the model is getting signal for these tasks as well.", "category": "Experimental_Setup"}, {"question": "Does the paper provide detailed cost information for creating Zephyr, including the costs of using LLMs in dSFT, generating answers in AIF, and ranking responses using GPT-4?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "tEYskw1VY2", "venue": "COLM 2024", "paper_decision": "Accept", "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", "authors": ["Albert Gu", "Tri Dao"], "abstract": "Foundation models, now powering most of the exciting applications in deep learning, are almost universally based on the Transformer architecture and its core attention module. Many subquadratic-time architectures such as linear attention, gated convolution and recurrent models, and structured state space models (SSMs) have been developed to address Transformers' computational inefficiency on long sequences, but they have not performed as well as attention on important modalities such as language. We identify that a key weakness of such models is their inability to perform content-based reasoning, and make several improvements. First, simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities, allowing the model to selectively propagate or forget information along the sequence length dimension depending on the current token. Second, even though this change prevents the use of efficient convolutions, we design a hardware-aware parallel algorithm in recurrent mode. We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (Mamba). Mamba enjoys fast inference (5x higher throughput than Transformers) and linear scaling in sequence length, and its performance improves on real data up to million-length sequences. As a general sequence model backbone, Mamba achieves state-of-the-art performance across several modalities such as language, audio, and genomics. On language modeling, our Mamba-3B model outperforms Transformers of the same size and matches Transformers twice its size, both in pretraining and downstream evaluation.", "keywords": "s;e;q;u;e;n;c;e; ;m;o;d;e;l;,; ;d;e;e;p; ;l;e;a;r;n;i;n;g;,; ;s;t;a;t;e; ;s;p;a;c;e; ;m;o;d;e;l;,; ;S;4;,; ;M;a;m;b;a", "qa_pairs": [{"question": "Why are there no results on long-context text tasks to justify that the models can replace transformers?", "answer": "There are no standard and widely agreed-upon high-quality benchmarks to measure long context in text. Therefore, the focus was on standard LM pretraining objectives and other modalities like DNA where long context matters. Follow-up studies, such as [LongMamba] and [Jamba], have further investigated long-context tasks where Mamba or its variants perform well.", "category": "Experimental_Analysis"}, {"question": "How does the training cost in terms of wall clock time and hardware efficiency compare between Mamba and Transformers, considering the differences in matrix multiplication utilization and learning rate tuning?", "answer": "Mamba is faster than Transformers in wall-clock time for longer sequences but less hardware-efficient on a FLOP-for-FLOP basis due to fewer matrix multiplications. However, recent follow-up works have developed related models that leverage matrix multiplications and are comparable to Transformers in wall-clock time even for shorter sequences.", "category": "Method_Comparison"}, {"question": "How do the fixed size memory/hidden state limitations of the model affect its performance in scenarios where users have different queries over the same encoded context in large language models (LLMs)?", "answer": "The tension between state size and recall ability is a fundamental issue in sequence models. SSMs and Transformers have different tradeoffs in inductive bias and capabilities, with SSMs having more trouble[1] with rote memorization but better ability to learn other types of formal languages[2]. Hybrid architectures combining SSMs and attention have shown strong performance in addressing these limitations.\n[1] https://arxiv.org/abs/2402.01032 [2] https://arxiv.org/abs/2405.17394", "category": "Method_Mechanics"}, {"question": "What is the hardware efficiency of Mamba compared to Transformers in terms of parallelization and utilization of machine compute?", "answer": "Mamba is faster than Transformers in wall-clock time for longer sequences but is less hardware-efficient on a *FLOP-for-FLOP* basis because it does not utilize as many matrix multiplications, which GPUs are specialized for. Active research is improving this, with follow-up works like [GLA, HGRN2, Mamba-2] achieving faster performance comparable to Transformers even for shorter sequences.", "category": "Method_Comparison"}, {"question": "Why are only B and C required to be context-aware, while A remains static in the SSM recurrence governed by equation (2a)?", "answer": "The discrete dynamics $\\bar{A}$, which govern the SSM recurrence in equation (2a), depend on both $A$ and $\\Delta$. Since $\\Delta$ is context-aware, $\\bar{A}$ inherently becomes context-aware. Early explorations showed that making $A$ input-dependent did not improve performance.", "category": "Concept_Understanding"}, {"question": "What is the comparison of Mamba's training wall time with that of other models of similar sizes?", "answer": "Mamba's training cost is comparable to Pythia, with Pythia 2.8B taking 14240 A100 hours and Mamba-2.7B taking 14370 A100 hours.", "category": "Experimental_Setup"}, {"question": "How does the performance of the Mamba architecture compare to the Transformer++ and other Transformer variants on downstream NLP tasks, as shown in the scaling law study?", "answer": "The Mamba architecture matches Transformer++ in performance on perplexity, as demonstrated in the scaling law study. For models up to 3B trained for 300B tokens, comparison with public models trained with similar token counts shows that Mamba, an SSM (linear in sequence length), can match the performance of the strongest Transformer (quadratic in sequence length) and outperform other Transformer variants.", "category": "Method_Comparison"}, {"question": "How critical is the short convolution module to Mamba's performance, and what are the effects of removing it?", "answer": "The short convolution module is a simplification of H3’s 'shift SSM' and is equivalent to RWKV’s 'token shift' with width=2. Removing it results in a degradation of model quality, as evidenced by a higher perplexity (ppl 8.97 without conv vs. 8.71 with conv at 350M with 7B tokens). The primary improvement due to the short conv is in selectivity, as shown in Table 6. Detailed ablation experiments are provided in Appendix E.3 (Tables 5-9).", "category": "Experimental_Exposition"}, {"question": "What are the formulas used to calculate FLOPs for Transformers and Mamba in the scaling law experiments, and why was the Chinchilla protocol, originally tuned for Transformers, used instead of fitting a new curve for Mamba?", "answer": "For Transformer FLOPs: 6 * nparams * ntokens + attention FLOPs (12 * nlayers * hdim * ctx_len * ntokens). For Mamba: 6 * nparams * ntokens + scan FLOPs (27 * nlayers * hdim * state_dim * ntokens). The Chinchilla protocol (20 tokens per params) was used because (1) it was originally tuned for Transformers, (2) changing it would exceed the compute budget, and (3) it provides a fair comparison to Transformer baselines.", "category": "Method_Disambiguation"}, {"question": "What is the upper-bound of Mamba's memory capacity in terms of memorizing history information, and how does it compare to other models like attention-based models?", "answer": "Recent studies, including the Based paper [1], show that finite-state models like SSMs have comparable memorization capabilities to attention-based models when plotted against effective state size. Pure softmax attention achieves larger effective state sizes, but recent SSM follow-ups (e.g., GLA [2], HGRN2 [3], Mamba-2 [4], RWKV-6 [5], xLSTM [6]) have significantly improved state sizes and performance on synthetic memorization tasks. An interesting theoretical direction is to characterize memory capacity in terms of effective state size, such as proving a lower bound. Another direction is exploring whether a model’s state size can grow with context length, potentially balancing the linear growth of Transformers and the constant size of SSMs. For example, the linear growth (i.e. KV cache) of a pure Transformer might be wasteful, whereas the constant size state of a pure SSM might be too restrictive.\n[1] Simran Arora et al. “Zoology: Measuring and Improving Recall in Efficient Language Models” ICML 2024.\n\n[2] Songlin Yang et al. “Gated Linear Attention Transformers with Hardware-Efficient Training” ICML 2024.\n\n[3] Zhen Qin et al. “HGRN2: Gated Linear RNNs with State Expansion” https://arxiv.org/abs2404.07904\n\n[4] Tri Dao and Albert Gu. “Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality” ICML 2024.\n\n[5] Bo Peng et al. “Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence” https://arxiv.org/abs/2404.05892\n\n[6] Maximilian Beck et al. “xLSTM: Extended Long Short-Term Memory” https://arxiv.org/abs/2405.04517\n", "category": "Method_Comparison"}]}
{"id": "IW1PR7vEBf", "venue": "COLM 2024", "paper_decision": "Accept", "title": "LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders", "authors": ["Parishad BehnamGhader", "Vaibhav Adlakha", "Marius Mosbach", "Dzmitry Bahdanau", "Nicolas Chapados", "Siva Reddy"], "abstract": "Large decoder-only language models (LLMs) are the state-of-the-art models on most of today's NLP tasks and benchmarks. Yet, the community is only slowly adopting these models for text embedding tasks, which require rich contextualized representations. In this work, we introduce LLM2Vec, a simple unsupervised approach that can transform any decoder-only LLM into a strong text encoder. LLM2Vec consists of three simple steps: 1) enabling bidirectional attention, 2) masked next token prediction, and 3) unsupervised contrastive learning. We demonstrate the effectiveness of LLM2Vec by applying it to 4 popular LLMs ranging from 1.3B to 8B parameters and evaluate the transformed models on English word- and sequence-level tasks. We outperform encoder-only models by a large margin on word-level tasks and reach a new unsupervised state-of-the-art performance on the Massive Text Embeddings Benchmark (MTEB). Moreover, when combining LLM2Vec with supervised contrastive learning, we achieve state-of-the-art performance on MTEB among models that train only on publicly available data (as of May 24, 2024). Our strong empirical results and extensive analysis demonstrate that LLMs can be effectively transformed into universal text encoders in a parameter-efficient manner without the need for expensive adaptation or synthetic GPT-4 generated data.", "keywords": "t;e;x;t; ;e;m;b;e;d;d;i;n;g;s;,; ;l;a;r;g;e; ;l;a;n;g;u;a;g;e; ;m;o;d;e;l;s;,; ;a;d;a;p;t;a;t;i;o;n", "qa_pairs": [{"question": "Did the authors compare the performance of the proposed encoder with LLMs using different prompts, and what were the results?", "answer": "Yes, the authors conducted experiments using the best-performing instructions (E5 inst) from previous work for all models, including Echo-embeddings, E5, and GritLM. Additionally, they compared the performance of their best unsupervised approach (Bi + MNTP + Mean) with Echo using two sets of instructions: E5 inst and Instructor inst. The results showed a 2-3% performance drop for both approaches when switching from E5 inst to Instructor inst, indicating similar behavior across prompt variations. The prompts used in the paper (E5 inst) are the best-performing ones in the literature.", "category": "Method_Comparison"}, {"question": "Where in the paper can the ablation studies demonstrating the effectiveness of MNTP be found, and what specific improvements does MNTP provide?", "answer": "The ablation studies demonstrating the effectiveness of MNTP can be found in Table 5 in the appendix. Specifically, MNTP's impact is observed by comparing the Bi vs. Bi+MNTP and Bi+SimCSE vs. Bi+MNTP+SimCSE models with mean pooling. For example, applying MNTP before SimCSE improves Llama-2’s performance by 34% in the unsupervised setting.", "category": "Experimental_Exposition"}, {"question": "What specific adaptations or optimizations were made to existing techniques like bidirectional attention, masked next token prediction (MNTP), and unsupervised contrastive learning in the LLM2Vec framework to enhance their novelty?", "answer": "The adaptations include substituting the masked token prediction of MLM with masked next token prediction to align with the nature of auto-regressive language models, i.e., predicting the next token instead of the current token. ", "category": "Method_Mechanics"}, {"question": "Why were enhancements like MNTP not applied to the unidirectional setting for a more equitable comparison between unidirectional and bidirectional attention mechanisms in the paper?", "answer": "MNTP is specifically designed for bidirectional models to utilize future tokens, which is not possible in unidirectional models due to their access only to past tokens. Applying MNTP to unidirectional models would not provide any benefit over the original LM objective. Additionally, the paper includes other configurations for unidirectional approaches in Table 5 in the Appendix  (Uni and Uni+SimCSE for three different pooling methods) to evaluate the impact of each component.", "category": "Motivation_Analysis"}, {"question": "What motivated the use of the term 'secretly' in the title, given that previous works like E5-mistral-instruct have already demonstrated similar findings?", "answer": "The motivation for using the term 'secretly' in the title is based on two key points: (1) the strong performance of method in the unsupervised setting and the minimal adaptation required to convert decoder-only LLMs into strong embedders, which hasn't been extensively explored before, and (2) the simplicity of approach, which demonstrates that supervised contrastive fine-tuning is not necessary to convert LLMs into strong embedding models. Additionally, the paper by Wang et al. (2024) is concurrent work, released less than 3 months before the COLM deadline.\nWang et al. (2024) - Improving Text Embeddings with Large Language Models", "category": "Motivation_Analysis"}, {"question": "Does the small improvement of the proposed bidirectional attention and unsupervised training over the replicated E5-mistral-instruct model in the supervised setting indicate a meaningful advancement, considering the MTEB benchmark's averaging across 56 tasks?", "answer": "Yes, the small improvement represents meaningful advancement because MTEB scores are averaged across 56 tasks, so even a small improvement indicates better performance across various task categories. This is a characteristic of the MTEB benchmark, as shown in Tables 1 and 2, and supported by the performance improvements of the top-10 models on the MTEB leaderboard in Table 9 in the Appendix. It can be seen that each entry reports a 'small' improvement over the previous approach and LLM2Vec also follows this pattern: 10) text-embedding-3-large (64.59) – +0.077% – 9) UAE-Large-V1 (64.64) – +0.062% – 8) mxbai-embed-large-v1 (64.68) – + 0.0% – 7) echo-mistral-7b-instruct-lasttoken (64.68) – +0.186% – 6) LLM2Vec (Bi+MNTP+Mean) (64.80) – +1.327% – 5) GritLM-8x7B (65.66) – +1.477% – 4) e5-mistral-7b-instruct (66.63) – +0.195% – 3) GritLM-7B (66.76) – +0.554% – 2) voyage-lite-02-instruct (67.13) – +0.641% – 1) SFR-Embedding-Mistral (67.56). Note that the improvements from rank 5) onwards are expected to be bigger compared to models on ranks 10) to 6) as these models rely on synthetic data for supervised fine-tuning.\n", "category": "Experimental_Exposition"}]}
{"id": "BOfDKxfwt0", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "LMSYS-Chat-1M: A Large-Scale Real-World LLM Conversation Dataset", "authors": ["Lianmin Zheng", "Wei-Lin Chiang", "Ying Sheng", "Tianle Li", "Siyuan Zhuang", "Zhanghao Wu", "Yonghao Zhuang", "Zhuohan Li", "Zi Lin", "Eric P. Xing", "Joseph E. Gonzalez", "Ion Stoica", "Hao Zhang"], "abstract": "Studying how people interact with large language models (LLMs) in real-world scenarios is increasingly important due to their widespread use in various applications. In this paper, we introduce LMSYS-Chat-1M, a large-scale dataset containing one million real-world conversations with 25 state-of-the-art LLMs. This dataset is collected from 210K unique IP addresses in the wild on our Vicuna demo and Chatbot Arena website. We offer an overview of the dataset's content, including its curation process, basic statistics, and topic distribution, highlighting its diversity, originality, and scale. We demonstrate its versatility through four use cases: developing content moderation models that perform similarly to GPT-4, building a safety benchmark, training instruction-following models that perform similarly to Vicuna, and creating challenging benchmark questions. We believe that this dataset will serve as a valuable resource for understanding and advancing LLM capabilities. The dataset is publicly available at https://huggingface.co/datasets/lmsys/lmsys-chat-1m.", "keywords": "large language models;dataset;conversation;safety;benchmark", "qa_pairs": [{"question": "Is a large portion of the dataset covered by only a few of the 210K unique IP addresses, and do most users access single models rather than the side-by-side mode?", "answer": "The IP addresses are very sparse, so no small set of IP addresses covers a large portion of the dataset. Most users access single models rather than side-by-side, but the distribution is shifting towards side-by-side due to recent promotions and it being the landing page.", "category": "Experimental_Setup"}, {"question": "What measures were taken to analyze the overlap between RealChat-1M and MT-Bench questions, and how was duplicate data handled during topic clustering?", "answer": "A string match was conducted to identify overlaps, revealing 12 out of 160 MT-bench questions in RealChat-1M. Duplication checks based on string matching were also applied during topic clustering.", "category": "Method_Mechanics"}, {"question": "What are the unique advantages of using the RealChat-1M dataset for fine-tuning LLMs, assessing LLM safety, and performance, compared to established datasets like OpenAssistant, SharedGPT, Anthropic HH, and Chatbot Arena?", "answer": "The RealChat-1M dataset offers several unique advantages: 1. It is up-to-date and continually updated, with plans for future data releases, including human preference data. Over the past month, authors received approximately 25K user queries per day for the latest models, including GPT-4 Turbo, Mistra, and Zephyr. Authors also intend to release more data on human preferences (votes).2. It is highly diverse, encompassing a wide range of models and user distributions, unlike other datasets that typically include only one or a limited number of models. 3. The scale of the dataset is crucial, as advancements in machine learning are often driven by scaling up datasets, as seen in projects like ImageNet and GPT-3/4.", "category": "Method_Comparison"}]}
{"id": "pzpWBbnwiJ", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Universal Guidance for Diffusion Models", "authors": ["Arpit Bansal", "Hong-Min Chu", "Avi Schwarzschild", "Roni Sengupta", "Micah Goldblum", "Jonas Geiping", "Tom Goldstein"], "abstract": "Typical diffusion models are trained to accept a particular form of conditioning, most commonly text, and cannot be conditioned on other modalities without retraining. In this work, we propose a universal guidance algorithm that enables diffusion models to be controlled by arbitrary guidance modalities without the need to retrain any use-specific components. We show that our algorithm successfully generates quality images with guidance functions including segmentation, face recognition, object detection, style guidance and classifier signals.", "keywords": "Generative Models;Computer Vision;Diffusion Models", "qa_pairs": [{"question": "How does the inaccuracy of the predicted clean image used as input to the conditioning model affect the performance of the method?", "answer": "Inaccurate clean image predictions have a limited impact on performance due to the step-by-step nature of image sampling in the diffusion model. The guidance signal, which uses gradient information, averages out across different sampling steps. In earlier steps, the guidance signal provides coarse adjustments based on less accurate predictions, while in later steps, the predictions become more accurate, refining the generated image. The success of the algorithm validates this reasoning.", "category": "Experimental_Analysis"}, {"question": "How does the proposed Universal Guidance differ from the family of algorithms mentioned in references [1, 2, 3], which freeze the unconditional diffusion backbone and train auxiliary models for controlled image generation?\n[1] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang#, Zhongang Qi, Ying Shan, Xiaohu Qie. T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models\n\n[2] Lvmin Zhang, Anyi Rao, Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models\n\n[3] Ziqi Huang, Kelvin C.K. Chan, Yuming Jiang, Ziwei Liu. Collaborative Diffusion for Multi-Modal Face Generation and Editing.", "answer": "The proposed Universal Guidance differs from the family of algorithms mentioned in references [1, 2, 3] in that it does not require re-training for different conditional generation tasks, whereas the referenced algorithms train auxiliary models for controlled image generation. Universal Guidance aims for a more flexible approach compared to the referenced algorithms, which represent a middle ground between Universal Guidance and full-blown conditional training.", "category": "Method_Comparison"}, {"question": "What are the novel contributions of this work compared to previous works such as PGDM [Song et al’22], DPS [Chung et al’22], and [Song et al’23], particularly in terms of loss guidance for diffusion models and the use of \\hat{z} to approximate z0 based on the score?", "answer": "The novel contributions of this work compared to PGDM, DPS, and [Song et al’23] include: (1) A universal guidance approach combining forward and backward guidance, which works with any guidance function and reduces to PGDM if the pseudo-inverse function is known. (2) Forward guidance generalizes DPS by accepting any type of condition and guidance function, addressing limitations of $\\ell_{p}$ loss. (3) Backward guidance and a step-wise refinement algorithm improve results on complex tasks like segmentation and object detection. (4) A universal refinement algorithm reinjects Gaussian noise in each diffusion step, differing from [Song et al’23]. (5) Empirical contributions demonstrating success on sophisticated guidance functions like segmentation and object detection networks.", "category": "Method_Comparison"}, {"question": "What are the contributions and differences of the proposed algorithm compared to the work in [Song et al.’23], specifically in terms of guided image generation using segmentation maps and object location?", "answer": "The proposed algorithm is compared with LGD [1], a recent work focusing on guided image generation. The authors perform image generation guided by segmentation maps and object locations using Stable Diffusion. Comparisons on segmentation-guided sampling (Figures 3 and 4) show that LGD often fails to match objects with the given segmentation maps. Comparisons on object detection guidance (Figures 5 and 6) reveal that LGD generates distorted images when attempting to better match given bounding boxes, indicating a failure to balance condition matching and image realism. This demonstrates the empirical contribution of the proposed algorithm in achieving more complex and widely applicable conditional generation tasks.\n[1] Song, Jiaming, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. \"Loss-Guided Diffusion Models for Plug-and-Play Controllable Generation.\" (2023).", "category": "Method_Comparison"}, {"question": "How does the proposed forward guidance methodology differ from DPS and FreeDoM, particularly in terms of handling conditions that cannot be compared using $\\ell_{p}$ loss and the use of step-wise refinement?", "answer": "DPS [R2] attempted to address guided generation using general guidance functions, but they proposed to apply the gradient of the $\\ell_{p}$ loss between the ground-truth condition and the output of the guidance function given a predicted clean image. The proposed forward guidance methodology differs from DPS in that it generalizes the approach to accept any type of condition and guidance function, unlike DPS which relies on $\\ell_{p}$ loss for conditions like bounding box locations. Additionally, the proposed method introduces backward guidance and a step-wise refinement algorithm, which are novel technical contributions not present in DPS. For FreeDoM, the comparison is addressed privately due to an unusual issue, and further details can be obtained by contacting the AC.", "category": "Method_Comparison"}]}
{"id": "hSyW5go0v8", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection", "authors": ["Akari Asai", "Zeqiu Wu", "Yizhong Wang", "Avirup Sil", "Hannaneh Hajishirzi"], "abstract": "Despite their remarkable capabilities, large language models (LLMs) often produce responses containing factual inaccuracies due to their sole reliance on the parametric knowledge they encapsulate. Retrieval-Augmented Generation (RAG), an ad hoc approach that augments LMs with retrieval of relevant knowledge, decreases such issues. However, indiscriminately retrieving and incorporating a fixed number of retrieved passages, regardless of whether retrieval is necessary, or passages are relevant, diminishes LM versatility or can lead to unhelpful response generation. We introduce a new framework called **Self-Reflective Retrieval-Augmented Generation (Self-RAG)** that enhances an LM's quality and factuality through retrieval and self-reflection. \nOur framework trains a single arbitrary LM that adaptively retrieves passages on-demand, and generates and reflects on retrieved passages and its generations using special tokens, called {\\it reflection} tokens. Generating reflection tokens makes the LM controllable during the inference phase, enabling it to tailor its behavior to diverse task requirements. \nExperiments show that Self-RAG (7B and 13B parameters) significantly outperforms state-of-the-art LLMs and retrieval-augmented models on a diverse set of tasks. \nSpecifically, Self-RAG outperforms ChatGPT and retrieval-augmented Llama2-chat on Open-domain QA, reasoning, and fact verification tasks, and it shows significant gains in improving factuality and citation accuracy for long-form generations relative to these models. Our code and trained models are available at https://selfrag.github.io/", "keywords": "Retrieval-augmented Generation;Language Models;Retrieval-augmented LMs;Factuality", "qa_pairs": [{"question": "How do the authors ensure that the annotations provided by GPT-4 for retrieval decisions are model-agnostic and suitable for smaller language models?", "answer": "The authors ensure model-agnostic annotations by asking GPT-4 to judge whether retrieving passages enhances generation quality in general or if the continuation requires factual verification, rather than focusing on GPT-4's own benefit from retrieval. They design few-shot demonstrations to align model predictions with this intuition. Additionally, the final reflection tokens for the generator training corpus are predicted by a Critic model based on Llama2-7b, which helps make the retrieval token predictions more model-agnostic.", "category": "Method_Mechanics"}, {"question": "What additional ablation studies and analyses were conducted to address the contributions of the self-reflection and RAG components, and how do these results clarify their individual impacts?", "answer": "The authors conducted new ablation studies on different special tokens (isREL, isSUP, isUse, and sequence scores) and the number of passages. Removing isREL alone resulted in minor performance deterioration, while removing both isREL and isSUP caused notable drops, highlighting their importance. Using only isUse for ranking led to significant performance deterioration, as it does not evaluate factuality or document relevance. Additionally, evaluations on the number of passages showed that performance improves up to 10 passages but drops slightly with 15 and 20 passages, likely due to confusion from irrelevant documents. These analyses clarify the individual impacts of the self-reflection and RAG components.\n\n| `IsRel` | `IsSup` | `IsUse`| `sequence` | PopQA performance (acc.)  |\n| ----------- | ----------- |----------- |----------- |----------- |\n|x |x |x |x | 0.538|\n| |x |x |x | 0.536|\n| | |x |x | 0.512 |\n| | |x | | 0.416|\n|  |  | |x | 0.512|\n\n| The number of passages (N=2) | PopQA performance (acc.) |\n| ----------- | ----------- |\n|   2    | 0.498       |\n|3 | 0.504 |\n|5 | 0.504 |\n| 7 | 0.540 |\n|10 | 0.538 |\n| 15 | 0.528 |\n|20 | 0.520 | \n", "category": "Experimental_Exposition"}, {"question": "Does Self-RAG outperform ChatGPT on open-domain QA tasks, as claimed in the abstract, despite the results in Table 2 being dominated by ChatGPT and Ret-ChatGPT except for PopQA?", "answer": "Yes, Self-RAG outperforms ChatGPT on open-domain QA and fact verification tasks, as supported by the results on PopQA and PubHealth. The abstract will be updated to make this clearer. Additionally, increasing the training data scale up to 150k has shown further gains across many tasks, as detailed in Section 5.2 and Table 2.", "category": "Experimental_Analysis"}, {"question": "What evidence supports the claim that GPT-4 reflection token predictions show high agreement with human evaluations, and how does the evaluation setup differ from PandaLM's approach?", "answer": "The authors sampled 50 examples per type (Retrieve, IsRel, IsSUP, isUse), manually labeled the reflection tokens, and compared GPT-4 predictions against human labels. They attribute the different outcomes to the evaluation setup: PandaLM uses pairwise preference evaluations, which are subjective and prone to disagreement, while their approach formulates evaluations in a fine-grained, segment-level manner, asking GPT-4 to assign absolute class labels/scores. This reduces subjectivity and yields more stable results. They will release human evaluation data, GPT-4 prompting, and Critic LM inference scripts for reproducibility.", "category": "Method_Comparison"}, {"question": "What techniques were implemented in Self-RAG to reduce computational overhead and latency during the generation stage, and how do these compare to the computational requirements of other baselines?", "answer": "In the generation stage, Self-RAG implements techniques such as parallel batch decoding of multiple passages, beam search, and paged attentions to reduce computational overhead and improve efficiency. The introduced latency is limited, and the same computational resources (a single GPU with 24GB memory) are used for other baselines. Additionally, Self-RAG can skip and reduce retrieval frequencies compared to normal RAG, often retrieving less than three times for long-form generations by reusing previously retrieved passages.", "category": "Method_Comparison"}, {"question": "How do variations in the retrieval threshold affect downstream task performance, particularly for datasets like PubHealth and PopQA?", "answer": "Figure 3 illustrates the retrieval threshold and retrieval frequencies for PubHealth (Fact verification) and PopQA (Open QA). PopQA requires more retrieval due to its many rare entities, which strong LLMs often struggle to memorize, leading to more 'retrieval' tokens for this dataset. In contrast, PubHealth requires less retrieval.", "category": "Experimental_Exposition"}, {"question": "Why does the ablation study focus only on 'Retrieve top1' and 'Remove[IsSup]' without examining other criticize tokens or varying the number of retrieved passages (e.g., top 5 or top 10)?", "answer": "The authors conducted new evaluations to address this feedback. For the number of passages, performance evaluations were run with varying numbers of passages on a sampled 500 PopQA dataset using a 7B model trained on 150K queries. Results show that performance increases up to 10 passages, with minor drops at N=15 and 20, likely due to confusion from irrelevant documents. For the removal of other special tokens, ablation studies were conducted on PopQA, evaluating the impact of removing different tokens. Results indicate that removing `isREL` and `isSUP` causes notable performance drops, while using only `isUse` for ranking outputs leads to significant deterioration, as it does not account for factuality or document relevance.\n | The number of passages (N) | PopQA performance (acc.) |\n | ----------- | ----------- |\n | 2 | 0.498 |\n | 3 | 0.504 | \n| 5 | 0.504 |\n | 7 | 0.540 |\n | 10 | 0.538 | \n| 15 | 0.528 |\n | 20 | 0.520 | \n\n| `IsRel` | `IsSup` | `IsUse`| `sequence` | PopQA performance (acc.) |\n | ----------- | ----------- |----------- |----------- |----------- | \n| x | x | x | x | 0.538 |\n | | x | x | x | 0.536 | \n| | | x | x | 0.512 |\n | | | x | | 0.416 |\n | | | | x | 0.512 |\n", "category": "Experimental_Analysis"}, {"question": "What steps did the authors take to investigate and address the influence of reflection token distributions on the performance of the critic model?", "answer": "The authors observed that reflection token distributions can be biased towards certain classes in some datasets, leading to potential exploitation by the Generator LM. For instance, in Open QA, there’s a much higher percentage of 'relevant' and 'fully supported' passages while in long-form instruction-following datasets, retrieved passages are often only partially supported by the outputs. To address this, they designed a sampling process to balance the distributions of reflection tokens. This process includes downsampling and discarding 50% of instances since large-scale instruction-following datasets (e.g., Alpaca), include many queries that do not require retrieval (e.g., simple and easy facts or not knowledge intensive). - For instances for the queries requiring retrieval, authors first retrieve 5-10 paragraphs for each segment and sample only 1 paragraph for the segment following the process described below: 1. For all datasets except for open-domain QA (NQ), if there is one or more than one paragraph with `[relevant]` and `[fully supported]`, authors randomly sample one from such paragraphs. For example, in open-domain QA (NQ), they randomly discard 25% of cases with 'relevant' and 'fully supported' paragraphs to avoid dominance by positive samples. They also describe detailed sampling rules for partially supported and irrelevant passages. For NQ, authors randomly discard such cases at 25% cases, as authors found NQ shows a significantly higher percentage of 'relevant' and 'fully supported' and the subset can be easily dominated by such 'positive' samples. 2. When none of the passages are relevant and fully supported while authors have some passages partially supported, authors take the highest ranked passages from this category at 30% of the cases. 3. When none of the passages are relevant but have no support, authors take the highest-ranked passages from this category at 75% of the cases. For the remaining cases, authors randomly sample 1 from all irrelevant paragraphs. These steps are documented in Section A.3, and the processing script will be released upon acceptance. While they could not re-train models due to time constraints, early experiments showed that without up and downsampling, model predictions could be biased towards certain categories.", "category": "Method_Mechanics"}, {"question": "Does the Critic model exhibit biases, and what is the distribution of retrieval tokens across different seed datasets?", "answer": "The Critic model's careful sampling process reduces undesirable biases, such as overpredictions of 'relevance' for QA-like evaluation datasets. The model learns to generate accurate predictions on when retrieval is helpful, resulting in different distributions of [Retrieve] tokens across instances from different seed datasets. For example, FLAN has 15.8% of instances with [Retrieve] tokens, Stanford Alpaca has 53.3%, Natural Questions (NQ) has 87.7%, and FEVER has 63.2%. FLAN instances often do not require retrieval, while NQ shows the highest percentage of retrieval instances due to its information-seeking nature. FEVER instances, on the other hand, often do not require external evidence for truthfulness judgments.\n | | FLAN |Stanford Alpaca |NQ | FEVER|\n | ----------- | ----------- |----------- |----------- |----------- | \n|% of instances with [Retrieve] | 15.8% | 53.3% | 87.7% | 63.2%| ", "category": "Experimental_Exposition"}, {"question": "What are the benefits of the offline data augmentation and training approach compared to prompting-based frameworks like 'Active retrieval augmented generation' by Jiang et al.?", "answer": "The offline data augmentation and training approach is more suitable for achieving the Self-RAG style inference process because it allows for careful multi-aspect fine-grained self-evaluations at inference time. Prompting-based approaches, as tested in preliminary experiments, struggle with precise adherence to evaluation schemes due to the complexity of instructions and special tokens, leading to output format mismatches and high disagreement with human evaluations. Additionally, the Self-RAG approach requires token probabilities for reflection tokens, which are often unavailable in black-box proprietary LM APIs like ChatGPT and GPT-4. This limitation is also noted in the Active Retrieval paper. While some work uses proprietary LMs for self-evaluation, it typically focuses on a single aspect, making fine-grained multi-aspect evaluations challenging without training. These points are further discussed in Appendix Section A.2.", "category": "Method_Comparison"}, {"question": "How does Self-RAG address the limitation of synthesizing multiple passages during response generation, given that it processes passages independently?", "answer": "Self-RAG processes multiple documents in parallel and does not synthesize multiple paragraphs for one segment generation. However, it continues generating responses based on multiple passages throughout the generation process. For example, to compare two concepts, Self-RAG retrieves a passage about Concept A and generates a segment, then retrieves another passage about Concept B and generates a segment conditioned on the first segment and Concept B. While this approach may not address all cases, batch decoding conditioned on multiple paragraphs independently provides an optimal trade-off between efficiency, performance, and better attributions, such as predicting relevance and supported tokens for each paragraph.", "category": "Method_Mechanics"}, {"question": "What are the triggering rates of the 'retrieval' token across different types of tasks, and how do these rates vary based on the dataset and task characteristics?", "answer": "Figure 3 (C) shows the triggering rates of the 'retrieval' token across different thresholds. For example, at a threshold of 0.5, PopQA triggers retrieval 70% of the time, while PubHealth triggers retrieval less than 20%. This variation is due to PopQA containing more long-tail, rare entity questions that require retrieval. Additionally, the percentage of instances with `[Retrieve]=Yes` tokens varies across datasets: FLAN (15.8%), Stanford Alpaca (53.3%), NQ (87.7%), and FEVER (63.2%). NQ shows the highest retrieval rate, as it consists of information-seeking questions requiring fine-grained world knowledge, while FEVER includes many instances where the model can judge truthfulness without retrieval. These results demonstrate the model's capability to identify instances that benefit from retrieval based on task characteristics.\n|      | FLAN |Stanford Alpaca |NQ | FEVER|\n | ----------- | ----------- |----------- |----------- |----------- |\n |% of instances with [Retrieve]      | 15.8%       | 53.3% | 87.7% | 63.2%|", "category": "Experimental_Exposition"}, {"question": "What is the required amount of training data for the Self-RAG system, and does the training data need to be carefully sampled to focus on fact-seeking slices?", "answer": "The Self-RAG system initially performs well with 50K training instances, but scaling up to 150K instances provides notable improvements, especially on certain tasks. Training data is heuristically sampled to balance reflection token distributions, and task diversity is emphasized to ensure robustness and versatility. The system leverages existing instruction-output pairs and does not rely on external APIs such as GPT-4, making scaling straightforward. Since the initial version using 50k training data, authors have created more training data using the same protocol and conducted ablations of the model performance trained on different numbers of training data. While the initial model trained on 50k instances already shows strong performance, authors found that increasing training data to 150k instances gives large improvements, especially on some tasks. Details on training data and performance improvements are provided in Figure 4 of the updated draft.  Note that widely used instruction-tuning LMs are often trained on more than 100k instances (e.g., Vicuna).", "category": "Experimental_Setup"}, {"question": "How do reflection tokens influence the generation process, and what mechanisms are in place to address situations where none of the generated candidates are deemed acceptable?", "answer": "Reflection tokens help identify and choose better segments, similar to rejection-sampling or best-of-n sampling in reinforcement learning, without relying on external reward models at inference time. While self-reflective decoding does not improve the initial set of generated samples, it allows practitioners to recognize when none of the segments are acceptable. This enables adjustments to the inference pipeline, such as sampling more passages from other retrieval systems or abstaining from answering if no high-scoring generations are produced. Preliminary experiments with special tokens at the beginning of generation showed limited effectiveness in controlling generation quality. Combining Self-RAG with fine-grained RLHF using the Critic as a reward model is another promising direction to enhance initial generations.", "category": "Experimental_Exposition"}, {"question": "What are the challenges and limitations of using GPT-4 to generate labeled data for reflection steps, and how does method address these issues?", "answer": "Generating extensive training data from GPT-4 is costly and time-consuming, with quality control remaining an open question. For example, creating 150k training instances could cost around USD 10k and face API time limits. Method reduces costs by requiring only medium-scale training data for the Critic model from GPT-4. Additionally, GPT-4 may not always generate ideal outputs grounded by the input, as state-of-the-art models can produce unsupported outputs. Retrieval-Augmented Generation also faces challenges with irrelevant retrieved passages, which may not be fully addressed by training on GPT-4 outputs. These issues are further discussed in Section 3.2.1.", "category": "Method_Mechanics"}, {"question": "What is the purpose of masking out the retrieved text chunks during training, as mentioned in section 3.2.2?", "answer": "The purpose of masking out the retrieved text chunks during training is to ensure that the model learns to use the given retrieved passages at inference time rather than generating passages by itself. This approach aligns with the goal of training the model to leverage retrieved passages for improved factuality, controllability, and attribution. Empirical evidence shows that not masking the passages leads to a drop in model performance on certain tasks.", "category": "Motivation_Analysis"}, {"question": "How do reflection tokens influence or guide the generation process, beyond providing post hoc feedback on the quality of the current generation?", "answer": "Reflection tokens help identify and choose better segments, similar to rejection-sampling or best-of-n sampling in reinforcement learning, without relying on external reward models at inference time. While they do not improve the initial sample generation, they enable recognition of situations where no segments are acceptable. This allows practitioners to adjust the inference pipeline, such as sampling more passages from other retrieval systems or abstaining from answering if no high-scoring generations are produced.", "category": "Method_Mechanics"}, {"question": "Does the study claim that the fine-grained evaluation method reduces the subjectivity of evaluations compared to pairwise preference evaluations?", "answer": "True", "category": "Claim_Verification"}, {"question": "Will the human evaluation data and GPT-4 prompting scripts be released upon acceptance of the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the distribution of retrieval tokens examined and correlated with the retrieval threshold across different seed datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the FEVER dataset contain instances where the model can judge truthfulness without external evidence, resulting in a lower percentage of retrieval instances compared to NQ?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the Natural Questions (NQ) dataset have the highest percentage of instances requiring retrieval among the four representative seed datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the performance data in Table 2 show that Self-RAG consistently outperforms ChatGPT across all open-domain QA tasks?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the results show that the Critic model predicts 'no retrieval needed' for tasks like grammatical checking on the FLAN dataset?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the critic model exhibit reduced biases due to a careful sampling process?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "J44HfH4JCg", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning", "authors": ["Fuxiao Liu", "Kevin Lin", "Linjie Li", "Jianfeng Wang", "Yaser Yacoob", "Lijuan Wang"], "abstract": "Despite the promising progress in multi-modal tasks, current large multi-modal models (LMMs) are prone to hallucinating inconsistent descriptions with respect to the associated image and human instructions. This paper addresses this issue by introducing the first large and diverse visual instruction tuning dataset, named Large-scale Robust Visual (LRV)-Instruction. Our dataset comprises 400k visual\ninstructions generated by GPT4, covering 16 vision-and-language tasks with open-ended instructions and answers. Unlike existing studies that primarily focus on positive instruction samples, we design LRV-Instruction to include both positive and negative instructions for more robust visual instruction tuning. Our negative instructions are designed at three semantic levels: (i) Nonexistent Object Manipulation, (ii) Existent Object Manipulation and (iii) Knowledge Manipulation. To efficiently measure the hallucination generated by LMMs, we propose GPT4-Assisted Visual Instruction Evaluation (GAVIE), a stable approach to evaluate visual instruction tuning like human experts. GAVIE does not require human-annotated groundtruth answers and can adapt to diverse instruction formats. We conduct comprehensive experiments to investigate the hallucination of LMMs. Our results demonstrate existing LMMs exhibit significant hallucinations when presented with our negative instructions, particularly Existent Object and Knowledge Manipulation instructions. Moreover, we successfully mitigate hallucination by finetuning MiniGPT4 and mPLUG-Owl on LRV-Instruction while improving performance on several public\ndatasets compared to state-of-the-art methods. Additionally, we observed that a balanced ratio of positive and negative instances in the training data leads to a more robust model. Code and data will be released upon publication.", "keywords": "instruction tuning;multimodal large language model;hallucination;datasets", "qa_pairs": [{"question": "How does the proposed model compare to LLaVA1.5 and MiniGPT4-v2, which are trained on Visual Genome data, in terms of instruction-following ability?", "answer": "The proposed model outperforms both LLaVA1.5-7B and MiniGPT4-v2-7B in terms of instruction-following ability, as evaluated on the GAVIE dataset. It achieves the best accuracy and relevancy, even though LLaVA1.5 and MiniGPT4-v2 show significant improvements compared to their previous versions. Qualitative examples are provided in Fig. 33 of the appendix.", "category": "Method_Comparison"}, {"question": "How did the authors ensure a fair comparison in evaluating the instruction-following ability of their model, given potential differences in training data domains?", "answer": "The authors ensured a fair comparison by comparing their model with LLaVA1.5 and MiniGPT4-v2, both of which use the Visual Genome as their training data to address domain differences. Their model outperformed both LLaVA1.5-7B and MiniGPT4-v2-7B in the GAVIE evaluation set regarding instruction-following ability, with scores of 8.46, 8.20, and 8.10 respectively.\n | Method | GAVIE-Relevancy |\n | ---- | ---- |\n | LLaVA 1.5-7B [1] | 8.20 | \n| MiniGPT4-v2-7B [2] | 8.10 | \n| Ours-7B | 8.46 |\n[1] Liu, Haotian, et al. \"Improved baselines with visual instruction tuning.\" arXiv preprint arXiv:2310.03744 (2023).\n\n[2] Chen, Jun, et al. \"MiniGPT-v2: large language model as a unified interface for vision-language multi-task learning.\" arXiv preprint arXiv:2310.09478 (2023).", "category": "Experimental_Setup"}, {"question": "How does the proposed approach address the issue of LMMs tendency to answer 'Yes' regardless of the proper response, and what experimental evidence supports this claim?", "answer": "The proposed approach addresses the issue by creating negative instructions involving existent object manipulation, non-existent object manipulation, and knowledge manipulation, in addition to positive instructions. Finetuning LMMs on this balanced dataset results in more consistent answers aligned with image content. Experimental evidence includes comparisons with LLaVA1.5 across three datasets: (1) AMBER[4], including object existence, object attribute, and object relation hallucination. The evaluation score is F1, and the groundtruth answer is yes or no. Although the model achieves a slightly lower F1 score regarding object existence, authors outperform LLaVA1.5 in both the object attribute and object relation hallucinations. (2) GAVIE, where the model achieves the best accuracy and relevancy, and (3) Hallusionbench [3], where the model shows higher accuracy on negative sets. These results demonstrate the model's effectiveness in reducing hallucinations.\n\n[3] Liu, Fuxiao, et al. \"Hallusionbench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models.\" arXiv preprint arXiv:2310.14566 (2023).\n\n[4] Wang, Junyang, et al. \"An LLM-free Multi-dimensional Benchmark for MLLMs Hallucination Evaluation.\" arXiv preprint arXiv:2311.07397 (2023).", "category": "Method_Mechanics"}, {"question": "What are the specific criteria and scoring options used in the human evaluation process?", "answer": "The details of human evaluation guidelines are shown below: As for each question, there are an instruction, an image, and several answers. The human evaluation process uses two criteria: (1) Accuracy: whether the response is accurate concerning the image content, and (2) Relevancy: whether the response directly follows the instruction without unrelated answers. There are four scoring options: 1 (Very Poor): The response is seldom relevant to the instruction and is mostly wrong according to the image. 2 (Poor): The response is partly relevant to the instruction and partly accurate according to the image. 3 (Good): The response is mostly relevant to the instruction and has minor errors according to the image. 4 (Excellent): The response is completely relevant to the instruction and completely accurate according to the image.", "category": "Experimental_Setup"}, {"question": "What are the main differences between the evaluation method GAVIE and G-Eval, particularly in terms of task adaptability, scoring criteria, and alignment with human evaluation?", "answer": "The main differences between GAVIE and G-Eval are: (1) GAVIE can adapt to diverse instruction formats and tasks, unlike G-Eval [5], which primarily focuses on text summarization and dialogue tasks. GAVIE was compared with human expert evaluation on 16 vision and language tasks, showing strong alignment with human evaluation (as seen in Table 4). (2) GAVIE generates different scores for different criteria, such as Relevancy (whether the response follows the instruction) and Accuracy (whether the response is accurate concerning the image content), unlike G-Eval, which generates a single score per sample. GAVIE also provides explanations to support the scores, offering clarity on model performance and scoring rationale.\n[5] Liu, Yang, et al. \"Gpteval: Nlg evaluation using gpt-4 with better human alignment.\" arXiv preprint arXiv:2303.16634 (2023).", "category": "Method_Comparison"}, {"question": "What remedies are proposed to address the issue of GAVIE's reliance on the black-box GPT-4 engine, which may affect result reproducibility due to system updates?", "answer": "Two remedies are proposed: (1) Fine-tuning LLaVA1.5 on vision and language evaluation tasks to enhance its instruction-following abilities and replacing its vision encoders to improve visual perception. (2) One can explore LLM-free methods for VL hallucination evaluation. Exploring LLM-free methods for VL hallucination evaluation by designing binary classification questions (For object hallucination, authors can construct the question as 'Is there a {hallucination object} in this image?'. For attribute hallucination, authors can design questions like 'Is the {object} {state} in this image?', 'Does the {object} {action} in this image?', 'Is the {object} {color} in this image?'. For relation hallucination, authors can construct questions like 'Is there a direct contact between the {object 1} and {object 2} in this image?'. ) to evaluate LMM outputs without relying on GPT-4.", "category": "Method_Mechanics"}, {"question": "What is the impact on GAVIE's performance when replacing GPT-4 with popular open-sourced Large Multimodal Models (LMMs) such as LLaVA1.5-13B and MiniGPT4-v2?", "answer": "When replacing GPT-4 with LLaVA1.5-13B and MiniGPT4-v2, LLaVA1.5-13B can follow instructions to evaluate answers based on two criteria but fails to detect incorrect answers given the image content. MiniGPT4-v2-13B cannot follow instructions to generate a score for each criterion and produces unreasonable scores. Thus, current open-sourced LMMs struggle with evaluating hallucination in vision-and-language tasks.", "category": "Experimental_Exposition"}, {"question": "How does the reliance on a specific version of GPT-4 (GPT4-32k-0314) affect the reproducibility and generalizability of the GAVIE results, and what are the potential alternatives to reduce this dependency?", "answer": "To address the reliance on GPT-4, authors specify the version (GPT4-32k-0314) used in Section 5 for reproducibility. Authors also tested open-sourced Large Multimodal Models (LMMs) like LLaVA1.5-13B and MiniGPT4-v2-13B, as shown in Figures 34 and 35. LLaVA1.5-13B can follow instructions but fails to detect incorrect answers, while MiniGPT4-v2-13B struggles with generating reasonable scores. To reduce dependency on GPT-4, authors propose two remedies: (1) Fine-tuning LLaVA1.5 on vision and language evaluation tasks and enhancing its visual perception by replacing vision encoders. (2) Exploring LLM-free methods for VL hallucination evaluation, For example, authors can design questions with answers from a closed set (e.g., binary classification) based on different types of hallucinations. For object hallucination, authors can construct the question as 'Is there a {hallucination object} in this image?'. For attribute hallucination, authors can design questions like 'Is the {object} {state} in this image?', 'Does the {object} {action} in this image?', 'Is the {object} {color} in this image?'. For relation hallucination, authors can construct questions like 'Is there a direct contact between the {object 1} and {object 2} in this image?'. For all the cases described above, the groundtruth answers to these questions will be 'yes' or 'no'. ", "category": "Method_Mechanics"}, {"question": "How does the proposed method address the scalability concerns related to its dependence on high-quality annotations from the Visual Genome dataset?", "answer": "The authors address scalability concerns by using GRiT to generate pseudo-dense captions instead of relying on high-quality human annotations from the Visual Genome dataset. Fine-tuning on these pseudo captions improves model performance, as shown in Table 5, demonstrating the method's potential to scale even without groundtruth dense captions. Further details are provided in Section 6.3.", "category": "Method_Mechanics"}, {"question": "What additional experiments were conducted to analyze more scenarios of large language models (LLMs), and what do the results demonstrate?", "answer": "Additional experiments were conducted by comparing the model with LLaVA1.5 and MiniGPT4-v2 on AMBER and Hallusionbench. The results demonstrate the effectiveness of the dataset.", "category": "Experimental_Exposition"}, {"question": "What is the correlation between GPT-4 ratings and human ratings, and how was this correlation assessed?", "answer": "To assess the correlation between GPT-4 ratings and human ratings, three NLP experts were hired to perform human assessments. The results, detailed in Table 4 and Appendix A.1.1, show that all experts ranked the output from the authors' model as the best, followed by InstructBLIP in second place, and MMGPT as the worst. This ranking aligns with the GAVIE metric.", "category": "Experimental_Exposition"}, {"question": "What is the rationale behind using a 0-10 scoring scale in GAVIE instead of a binary 0/1 scale?", "answer": "The 0-10 scoring scale in GAVIE allows for a detailed evaluation of the model’s answers, distinguishing between completely wrong, partly wrong, and entirely correct answers. Unlike a binary 0/1 scale, which assigns the same score (0) to both completely wrong and partly wrong answers, GAVIE can assign a score of 0 for completely wrong answers and a partial score (e.g., 5) for partly wrong answers. This enables more accurate model ranking and a better understanding of the level of correctness in model responses. RELEVANCY has four grade levels: (1) The response is entirely relevant (9-10), (2) The response is mostly relevant (6-8), (3) The response is partly relevant (3-5), (4) The response is seldom relevant (0-2). ACCURACY has four grade levels: (1) The response is completely accurate (9-10), (2) The response has minor errors (6-8), (3) The response is partly accurate (3-5), (4) The response is mostly or completely wrong (0-2)..", "category": "Motivation_Analysis"}, {"question": "Why is evaluating LMM output in an open-ended manner still challenging even when using a binary classification task (e.g., 'Yes' or 'No') for hallucination evaluation?", "answer": "POPE formulates hallucination evaluation as a binary classification task ('Yes' or 'No'). However, it cannot evaluate model hallucinations for multiple-choice or open-ended questions (e.g., 'Choose the correct statement about the weather conditions in the image: (a) Cloudy and rainy, (b) Clear blue sky, (c) Foggy and misty, (d) Snowy and cold') or an open-ended question (e.g., 'What objects are on the toddler’s feet?') because the answer space becomes more complex. In contrast, GAVIE can evaluate hallucinations even in open-ended scenarios, as demonstrated in Figures 21-26 in the appendix.", "category": "Experimental_Analysis"}, {"question": "Why is the answer to the question about 'Merkel' in Figure 2 (neg) related to 'Hillary Clinton' instead of 'Merkel'?", "answer": "The question about 'Merkel' is a negative instruction, meaning 'Merkel' does not appear in the image. The correct answer should be 'No, Hillary Clinton arrived at the Los Angeles Get Out The Vote Rally in the image.' This negative sample is used to fine-tune Large Multimodal Models (LMMs) to learn to say 'no' when objects mentioned in questions are not present in the image, reducing hallucinations.", "category": "Experimental_Analysis"}, {"question": "Did the human experts utilize the full range of (0-4) scores for their evaluations?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "mw1PWNSWZP", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "OctoPack: Instruction Tuning Code Large Language Models", "authors": ["Niklas Muennighoff", "Qian Liu", "Armel Randy Zebaze", "Qinkai Zheng", "Binyuan Hui", "Terry Yue Zhuo", "Swayam Singh", "Xiangru Tang", "Leandro Von Werra", "Shayne Longpre"], "abstract": "Finetuning large language models (LLMs) on instructions leads to vast performance improvements on natural language tasks. We apply instruction tuning using code, leveraging the natural structure of Git commits, which pair code changes with human instructions. We compile CommitPack: 4 terabytes of Git commits across 350 programming languages. We benchmark CommitPack against other natural and synthetic code instructions (xP3x, Self-Instruct, OASST) on the 16B parameter StarCoder model, and achieve state-of-the-art performance among models not trained on OpenAI outputs, on the HumanEval Python benchmark (46.2% pass@1). We further introduce HumanEvalPack, expanding the HumanEval benchmark to a total of 3 coding tasks (Code Repair, Code Explanation, Code Synthesis) across 6 languages (Python, JavaScript, Java, Go, C++, Rust). Our models, OctoCoder and OctoGeeX, achieve the best performance across HumanEvalPack among all permissive models, demonstrating CommitPack's benefits in generalizing to a wider set of languages and natural coding tasks. Code, models and data are freely available at https://github.com/bigcode-project/octopack.", "keywords": "large language models;large code models;instruction tuning", "qa_pairs": [{"question": "What are the key differences between OCTOCODER and OCTOGEEX in terms of their pre-trained base models and hyperparameters?", "answer": "The key differences are the pre-trained base models and some minor hyperparameters. OctoCoder is initialized from StarCoder, while OctoGeeX is initialized from CodeGeeX2. There are also tiny differences in the hyperparameters during finetuning, which are detailed in `Appendix O: Hyperparameters`. Additionally, OctoGeeX outperforms OctoCoder on Go and Rust for HumanEvalExplain, likely due to the superior performance of CodeGeeX2 compared to StarCoder for these languages.  (https://github.com/THUDM/CodeGeeX2/blob/main/README_EN.md) ", "category": "Method_Comparison"}, {"question": "What was the motivation for limiting the input length to 768 tokens in the dataset, and how does this limitation impact the dataset's usefulness for downstream use-cases where programs are typically longer than those in the HumanEval suite?", "answer": "The motivation for limiting the input length to 768 tokens was to avoid wasting finetuning effort on teaching the model to copy code from the 'before' to the 'after' sections, as a large fraction of the code remains the same. This limitation increases the signal per token, enabling the model to learn faster and avoid always repeating user input. The impact on downstream use-cases is that the dataset focuses on relatively short commit pairs, which may not fully represent the natural distribution of longer programs.", "category": "Motivation_Analysis"}, {"question": "What does the 'FT' in CommitPackFT stand for, and why was only a 5,000-sample subset used for instruction tuning instead of the entire dataset?", "answer": "The 'FT' in CommitPackFT stands for 'filtered', indicating it is a filtered version of CommitPack. Only 5,000 samples were used for instruction tuning because prior work has shown that very few samples are needed for this purpose. The models converge quickly, and most capabilities are learned during pre-training, with instruction tuning mainly teaching the model the expected input and output format.", "category": "Motivation_Analysis"}, {"question": "What factors contribute to OctoGeeX's superior performance in Go and Rust compared to OctoCoder?", "answer": "The primary factors are the different pre-trained base models and minor hyperparameter differences. OctoGeeX is initialized from CodeGeeX2, while OctoCoder is initialized from StarCoder. CodeGeeX2, the base model of OctoGeeX, performs better than StarCoder, OctoCoder’s base model, for Go and Rust. Detailed hyperparameter differences are documented in `Appendix O: Hyperparameters`.  The base model (CodeGeeX2) also performs better than OctoCoder’s base model (StarCoder) for Go and Rust (https://github.com/THUDM/CodeGeeX2/blob/main/README_EN.md) which could be the reason for this.", "category": "Method_Comparison"}, {"question": "Why were only 5K samples chosen for fine-tuning, and how were the sample sizes for StarCoder and OASST determined?", "answer": "The choice of 5K samples was based on the observation that many samples are not needed for instruction tuning, similar to LIMA which used only 1000 samples. OctoCoder is fine-tuned for 35 steps with a batch size of 32, corresponding to ~1100-8000 samples. For StarCoder and OASST, the goal was to ensure all datasets had the same order of magnitude, so filtered OASST and StarCoder were left as is, while larger datasets like CommitPackFT and xP3x were subsampled. The diversity of CommitPackFT across languages and its focus on code fixing were also considered, and fine-tuning was not performed solely on CommitPackFT due to the lack of samples with natural language targets.", "category": "Motivation_Analysis"}, {"question": "How does StarCoder perform when prompted to predict docstrings or comments using its FiM capabilities, and what are the corresponding pass@1 results across different programming languages?", "answer": "When prompted with FiM, StarCoder demonstrates the ability to generate easily readable docstrings, achieving an average pass@1 result of approximately 16.6% across Python, JavaScript, Java, Go, C++, and Rust. The specific results are: Python (19.4%), JavaScript (17.6%), Java (16.3%), Go (11.8%), C++ (17.9%), and Rust (16.7%). These results have been added to the Appendix.\n| Model | Python | JavaScript | Java | Go | C++ | Rust | Avg. | \n|------------|--------|------------|------|----|-----|------|------| \n| StarCoder | 19.4 | 17.6 | 16.3 | 11.8 | 17.9 | 16.7 | 16.6 |", "category": "Experimental_Exposition"}]}
{"id": "9JQtrumvg8", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis", "authors": ["Izzeddin Gur", "Hiroki Furuta", "Austin V Huang", "Mustafa Safdari", "Yutaka Matsuo", "Douglas Eck", "Aleksandra Faust"], "abstract": "Pre-trained large language models (LLMs) have recently achieved better generalization and sample efficiency in autonomous web automation.\nHowever, the performance on real-world websites has still suffered from (1) open domainness, (2) limited context length, and (3) lack of inductive bias on HTML.\nWe introduce WebAgent, an LLM-driven agent that learns from self-experience to complete tasks on real websites following natural language instructions.\nWebAgent plans ahead by decomposing instructions into canonical sub-instructions, summarizes long HTML documents into task-relevant snippets, and acts on websites via Python programs generated from those.\nWe design WebAgent with Flan-U-PaLM, for grounded code generation, and HTML-T5, new pre-trained LLMs for long HTML documents using local and global attention mechanisms and a mixture of long-span denoising objectives, for planning and summarization.\nWe empirically demonstrate that our modular recipe improves the success on real websites by over 50%, and that HTML-T5 is the best model to solve various HTML understanding tasks; achieving 18.7% higher success rate than the prior method on MiniWoB web automation benchmark, and SoTA performance on Mind2Web, an offline task planning evaluation.", "keywords": "Web Navigation;Web Automation;Large Language Models;Language Model Agents;Tool Use;Program Synthesis", "qa_pairs": [{"question": "How does the training process of HTML-T5 differ from that of the other models compared in Table 4, specifically in terms of task specialization?", "answer": "HTML-T5 is initially pre-trained with a curated HTML corpus from CommonCrawl and then finetuned on all available MiniWoB demonstrations. In contrast, all models in Table 4, including HTML-T5, are exclusively finetuned with 12K/347K MiniWoB demonstrations and not with multi-task corpora involving other NLP tasks, focusing on the efficacy of pre-trained models for simulated web automation tasks.", "category": "Method_Comparison"}, {"question": "How does the performance of the WebAgent approach scale across different model sizes, from small open-source models to large models like Flan-U-PaLM-540B?", "answer": "New results in Appendix E (Figure 7) compare Flan-U-PaLM-540B, gpt-3.5-turbo, and smaller Flan-PaLM models (8B and 62B). Flan-U-PaLM-540B and gpt-3.5-turbo show comparable performance (80% success, 93.8% score), while Flan-PaLM-62B (60% success, 86.3% score) underperforms due to weaker program synthesis. Flan-PaLM-8B fails to generate valid programs. The findings suggest that larger model sizes improve WebAgent performance, and any LLM capable of generating python/selenium code can be integrated into WebAgent.", "category": "Experimental_Exposition"}, {"question": "How does WebAgent handle and respond to failures during its operation?", "answer": "WebAgent's planner, HTML-T5, is designed to automatically replan when a new state is observed, allowing the agent to quickly respond to unusual transitions or errors. For instance, if an error sends the state to the home screen during an apartment search, the agent immediately begins replanning from the initial step.", "category": "Method_Mechanics"}, {"question": "How does the performance of WebAgent scale with increased data, parameters, and compute?", "answer": "The performance of WebAgent improves with scale. Increasing the number of parameters from 220M to 3B enhances performance, as shown in Table 8 in Appendix I. Additionally, increasing the data size from 12K to 347K improves performance, as demonstrated in Table 4. These points are further explained in Section 5 (`Broad Generalization across the Internet`).", "category": "Experimental_Analysis"}, {"question": "How does the performance of the WebAgent method vary when using different language models, specifically comparing Flan-U-PaLM-540B with other models like gpt-3.5-turbo and smaller Flan-PaLM variants?", "answer": "The results in Appendix E (Figure 7) show that Flan-U-PaLM-540B and gpt-3.5-turbo have comparable performance (80% success, 93.8% score). Flan-PaLM-62B achieves 60% success and 86.3% score, falling short due to inferior program synthesis capabilities, while Flan-PaLM-8B fails to generate valid programs. These findings suggest that larger model sizes enhance WebAgent performance, and any LLM capable of generating python/selenium code can be integrated into WebAgent.", "category": "Method_Comparison"}, {"question": "How does the performance of the proposed method compare to recent baselines for miniwob++, particularly in light of methods reporting near-human performance (93%)?", "answer": "Authors have added a state-of-the-art baseline using GPT-3.5 (Synapse [2]) in Table 4 for a more comprehensive comparison. However, authors' focus is on comparing models that are accessible, disclose their training corpus, and adhere to the same training dataset size (12K or 347K). While GPT variants show strong performance on MiniWoB, direct comparisons are challenging due to the lack of transparency in their training corpora. HTML-T5 represents a step towards on-device and privacy-preserving deployment.\n[2] Zheng et al., (2023) Synapse: Trajectory-as-Exemplar Prompting with Memory for Computer Control (https://arxiv.org/abs/2306.07863)", "category": "Method_Comparison"}, {"question": "What additional baselines or agents were used to compare the performance of WebAgent and HTML-T5 in real-world tasks, and what were the results?", "answer": "Additional baselines were added in Appendix E (Figure 7), comparing Flan-U-PaLM-540B, gpt-3.5-turbo, and smaller Flan-PaLM variants (8B and 62B) on the map website. Results show that Flan-U-PaLM-540B and gpt-3.5-turbo perform comparably (80% success, 93.8% score), while Flan-PaLM-62B performs worse (60% success, 86.3% score), and Flan-PaLM-8B fails to generate valid programs. For Mind2Web, results from Synapse [1] were added to Table 4, showing HTML-T5 still achieves the best performance.\n[1] Deng et al., (2023) Mind2Web: Towards a Generalist Agent for the Web (https://arxiv.org/abs/2306.06070)", "category": "Method_Comparison"}, {"question": "What specific data was used to train HTML-T5 for predicting sub-instruction plans and HTML summaries, and how was this data collected?", "answer": "The data used to train HTML-T5 for predicting sub-instruction plans and HTML summaries was collected through a self-experience supervision approach. This involved sampling new instructions using templates, curating navigation scripts to gather planning (sequence of sub-instructions) and summarization data (corresponding reference IDs), prompting Flan-U-PaLM to generate Python programs, and utilizing execution feedback to eliminate incorrect trajectories. Additionally, HTML-T5 was fine-tuned using demonstrations collected through scripted planning and prompted programming. The dataset includes 260 episodes on real-estate, 230 episodes on social-media, and 410 episodes on map websites, as detailed in Section 4.1. Examples for different tasks are provided in Appendix D.", "category": "Method_Mechanics"}, {"question": "How is the term 'HTML summary' defined, and what is its role in the functioning of Flan-U-PaLM as described in Section 3.2?", "answer": "The term 'HTML summary' refers to HTML snippets that are extracted and merged based on predicted sub-instructions and corresponding `data-ref` attributes by the fine-tuned HTML-T5 model. These merged HTML snippets are then used to prompt Flan-U-PaLM to generate navigation programs, as illustrated in Figure 3.", "category": "Concept_Understanding"}, {"question": "How does HTML-T5 handle inputs (instruction, previous sub-instructions, and raw HTML) within a 4096-token window size, given that raw HTML typically consumes significant space?", "answer": "Approximately 90% of the pre-training HTML corpus has a context length of around 4K tokens. Raw HTML documents are pre-processed to extract useful subtrees, which reduces context length significantly (see Appendix C) and improves effectiveness (see Table 7 for ablation results). HTML-T5 is pre-trained using a 4096-token window size (`Pre-Training with Mixture of Long-Span Denoising` in Section 3.1) and fine-tuned with a 16K window size (Section 4.2) to generalize to real-world HTML documents.", "category": "Method_Mechanics"}, {"question": "What method was used to ensure no overlap in tasks or sub-tasks between the training and testing sets in the real-world navigation dataset, and how were the train/test splits handled for MiniWoB and Mind2Web?", "answer": "For real-world navigation, a single dataset of instructions was partitioned to ensure no overlap between training and testing sets, with testing instructions listed in Appendix F. For MiniWoB, instructions were randomly generated from a large pool, and environments were set for 'training' or 'testing.' For Mind2Web, separate training and testing sets were maintained as per the original experimental setup.", "category": "Method_Mechanics"}, {"question": "How was HTML-T5 trained to predict sub-instruction plans and HTML summaries, and what data supervision method was used?", "answer": "HTML-T5 was trained using a self-experience supervision approach, which includes sampling new instructions with templates, curating navigation scripts to gather planning and summarization data, prompting Flan-U-PaLM to generate Python programs, and using execution feedback to eliminate incorrect trajectories. The fine-tuning process employed demonstrations collected through scripted planning and prompted programming, as detailed in Section 3.2.", "category": "Experimental_Setup"}, {"question": "What is the difference between the window sizes used for pre-training and fine-tuning in HTML-T5, as mentioned in Sections 3.1 and 4.2?", "answer": "HTML-T5 is pre-trained using a 4096 token window size to extract useful subtrees and reduce context length, as detailed in Section 3.1 and Appendix C. For fine-tuning, a 16K token window size is used to generalize to real-world HTML documents, as explained in Section 4.2.", "category": "Method_Mechanics"}, {"question": "How are 'open-ended actions' (Figure 1), 'canonical sub-instructions' (abstract), and 'pre-defined action space' defined, and does the author advocate for open-ended action spaces or pre-defined action spaces?", "answer": "Pre-defined action spaces are a-priori defined sets of all effective actions that an agent can perform. The authors advocate for programs as superior alternatives to pre-defined action spaces, as they are more flexible (open-ended) and adaptable, benefiting from advancements in program synthesis. 'Canonical sub-instructions' refer to the rule-based parsing of instructions into sub-instructions. The authors have revised the caption of Figure 1, removed the term 'canonical' from the abstract to reduce ambiguity, and clarified Section 3.", "category": "Concept_Understanding"}, {"question": "Does the phrase 'given a few canonical examples for program generation' in Section 3.3 refer to the 'few-shot prompt' step in Figure 3, and what is the source of these examples?", "answer": "Yes, the phrase refers to the 'few-shot prompt' step. The examples used for few-shot prompting are generic and include actions such as selecting checkboxes, entering text into inputs, clicking on options, and scrolling. These examples are domain-independent to ensure simplicity, robust generalization to unseen websites, and prevent information leakage.", "category": "Concept_Understanding"}, {"question": "What are the differences in input/output between the tasks described in Sections 4.1 and 4.2, beyond the variations in baseline models and datasets?", "answer": "The input for all tasks includes user instructions, navigation history, and raw HTML documents. The outputs differ: in Section 4.1 (real-world website navigation), outputs are paired sub-instructions and `data-ref` attributes; in MiniWoB (Section 4.2), outputs are pre-defined actions like 'click(ref=X)'; and in Mind2Web (Section 4.2), outputs are multiple-choice questions  (e.g., 'B') and operational labels (e.g., 'Click XXX'). Section 4.1 focuses on interactive real-world evaluation, while Section 4.2 covers simulator and offline evaluation.", "category": "Method_Disambiguation"}, {"question": "What is the definition of 'planning' in Section 3.2, and does 'summarization' refer to localizing the relevant snippet of the current sub-instruction?", "answer": "Planning refers to predicting the next sub-instruction to be executed based on the current state, with HTML-T5 iteratively generating the next sub-instruction per step (closed-loop planning). Summarization refers to retrieving all relevant HTML snippets by predicting their `data-ref` attributes, similar to extractive summarization. The combined sub-instruction and HTML snippets serve as the input for Flan-U-PaLM.", "category": "Concept_Understanding"}, {"question": "Can authors summarize the end-to-end workflow of WebAgent, including the user input, system's knowledge or database, and whether the interaction between HTML-T5 and Flan-U-PaLM is a one-time action or an iterative process?", "answer": "The WebAgent workflow begins with users initiating natural language interactions with a clear intent, such as apartment searching. HTML-T5 formulates a 'go to<URL>' sub-instruction, which triggers Flan-U-PaLM to generate a Python program that navigates to the specified website. The raw HTML content of the newly opened page is extracted and fed into HTML-T5 along with the user’s instruction and previous planning steps. This information is used to predict the next sub-instruction and identify relevant reference IDs for extractive HTML summarization. Flan-U-PaLM then generates a Python program based on these sub-instructions and the combined HTML snippet. This process is iterative, continuing until a designated end-of-episode sub-instruction is predicted or the maximum number of iterations is reached.", "category": "Method_Mechanics"}, {"question": "What are the details of the newly curated dataset used for pre-training and fine-tuning WebAgent, including the pre-processing steps, dataset size, and examples of templates, sub-instructions, and actions?", "answer": "For the **pre-training** dataset, 100 WARC files from the CommonCrawl corpus (April 2019) were collected, and non-Unicode or alphanumeric-only HTML documents were removed. Subtrees around <label> elements with a 'for' attribute were extracted to focus on HTML relevant for instruction following and grounding, resulting in a final dataset of 3.41M examples. For the **finetuning** dataset, instructions were sampled by assigning values to placeholders in manually-curated templates. A rule-based parser decomposed instructions into sub-instructions, and reference IDs were retrieved from HTML using regular expressions. Flan-U-PaLM generated navigation programs using sub-instructions and HTML snippets, executed via Selenium WebDriver. HTML-T5 was fine-tuned using self-experience demonstrations from instruction sampling, scripted planning, and prompted program synthesis. Inputs included task instructions, sub-instruction histories, and raw HTML, with outputs being the next sub-instruction and relevant `data-ref` attributes for HTML snippet retrieval.", "category": "Experimental_Setup"}, {"question": "In the real-estate case in Table 1, does the WebAgent encounter the same searching page with varying instructions across the 20 tests?", "answer": "The WebAgent navigates to the same home page, but the underlying HTML document may vary, as home pages can display different example listings each time the website is visited.", "category": "Experimental_Setup"}, {"question": "Is HTML-T5 planned to be released after evaluating its societal implications?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are all components of WebAgent, including PaLM models, currently available open-source?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is there an intention to release HTML-T5, a component of WebAgent, after evaluating its societal implications?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "NG7sS51zVF", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Efficient Streaming Language Models with Attention Sinks", "authors": ["Guangxuan Xiao", "Yuandong Tian", "Beidi Chen", "Song Han", "Mike Lewis"], "abstract": "Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges.\nFirstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory.\nSecondly, popular LLMs cannot generalize to longer texts than the training sequence length.\nWindow attention, where only the most recent KVs are cached, is a natural approach --- but we show that it fails when the text length surpasses the cache size.\nWe observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a ``sink'' even if they are not semantically important.\nBased on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence length without any fine-tuning.\nWe show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more.\nIn addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2$\\times$ speedup.\nCode and datasets are provided at https://github.com/mit-han-lab/streaming-llm.", "keywords": "Large Language Models;Length Extrapolation;Efficiency", "qa_pairs": [{"question": "How does StreamingLLM's performance translate relative to evicted tokens at million-scale sequence lengths, and does this limit its practical utility to information at the beginning and end of the sequence?", "answer": "StreamingLLM is especially suited for applications requiring continuous, compute- and memory-efficient operation, such as a conversational AI assistant. Unlike previous methods that lose context or need time-intensive recomputation of KV states, StreamingLLM maintains conversational continuity and context without cache refresh. StreamingLLM's ability to produce quality outputs without KV recomputation or cache refresh. StreamingLLM maintains accuracy when the token distance between the query and answer is within the cache size, but accuracy diminishes as this distance increases and drops to zero when it surpasses the cache capacity. This demonstrates that StreamingLLM effectively generates coherent text based on recent context but cannot extend the context length of language models, limiting its utility to information within the cache.", "category": "Experimental_Analysis"}, {"question": "Can the initial tokens in autoregressive language models, which show concentrated attention as the model goes deeper, be interpreted as performing a global memory or attention operation?", "answer": "No, the initial tokens in autoregressive language models cannot be interpreted as performing a global memory or attention operation. These models restrict tokens to attending only to preceding tokens, not subsequent ones. Since the initial tokens (attention sinks) cannot aggregate information from future tokens, they do not serve as global memory.", "category": "Experimental_Analysis"}, {"question": "What is the relationship between the concept of 'attention sinks' in this work and the 'registers' in the 'Vision Transformers Need Registers' paper, and how do these concepts differ in their roles within transformer models?", "answer": "The 'ViTs Need Registers' paper identifies 'registers' as random background patch tokens in Vision Transformers that hold global image information and proposes dedicated 'register' tokens to mitigate attention spikes. In authors' work, 'attention sinks' are initial tokens in autoregressive models that attract disproportionate attention. Authors found that introducing a dedicated sink token during pre-training prevents the use of actual content tokens as attention sinks, resulting in cleaner attention maps. Both phenomena are influenced by the softmax function in attention mechanisms, but differ in their positional roles: 'registers' serve as global information repositories in intermediate layers of Vision Transformers, while 'attention sinks' are initial tokens in autoregressive models and do not provide similar global information storage. This comparison highlights the similarities and differences in attention management across transformer models.", "category": "Method_Comparison"}, {"question": "How does StreamingLLM's performance on the StreamEval benchmark decay with increasing context length, specifically when querying information related to line 1k after having seen 10/100k lines, and how does it perform relative to evicted tokens?", "answer": "StreamingLLM maintains accuracy when the token distance between the query and answer is within the cache size. However, accuracy diminishes as this distance increases and drops to zero when it surpasses the cache capacity. This indicates that StreamingLLM cannot extend the context length of language models and highlights a broader challenge in current language models' inability to fully utilize context information within the cache, as also observed in the 'Lost-in-the-Middle' paper (Liu et al., 2023).", "category": "Experimental_Exposition"}, {"question": "How does the paper address the evaluation of context utilization by the model when using attention sinks, particularly for extremely large sequence lengths?", "answer": "The paper acknowledges the importance of evaluating context utilization with attention sinks. Initial experiments showed satisfactory context utilization within cache limits, comparable to the oracle baseline. Further experiments in Appendix Section C assessed context utilization beyond cache limits, revealing that StreamingLLM’s context utilization is limited to the current cache and cannot leverage evicted tokens. StreamingLLM's primary aim is to generate coherent text from recent tokens without cache refreshes, rather than augmenting the context window or boosting long-term memory. The paper also demonstrates StreamingLLM's integration with context extension methods, allowing larger caches for a broader range of recent tokens, as shown in Figure 9 with models like LongChat-7B-v1.5-32K and Llama-2-7B-32K-Instruct.", "category": "Experimental_Exposition"}, {"question": "Did authors evaluate StreamingLLM on long-range tasks using benchmarks like LongBench, and what were the findings regarding its context utilization capabilities?", "answer": "Authors evaluated StreamingLLM on long-range tasks using LongBench, focusing on tasks like single-document QA (NarrativeQA and Qasper), multi-document QA(HotpotQA and 2WikiMQA), and summarization (GovReport and MultiNews). The results, detailed in Appendix Section D, show that StreamingLLM's performance is comparable to the text truncation baseline. However, StreamingLLM is not designed to enhance long text utilization; its strength is in efficiently using recent information without needing cache refreshes or extensive memory. This limitation is acknowledged and is a potential area for future research.", "category": "Experimental_Exposition"}, {"question": "Why does StreamingLLM achieve a 22.2x speedup compared to the sliding window with re-computation baseline, and why might adjusting stride lengths in sliding windows not be a viable alternative?", "answer": "StreamingLLM achieves a 22.2x speedup compared to the sliding window with re-computation baseline due to its linear $O(N)$ complexity, which contrasts with the sliding window's quadratic $O(N^2)$ complexity. Adjusting stride lengths in sliding windows may offer some efficiency but introduces variability in context size and computational workload, leading to inconsistent decoding times and greater computational demands during re-computation phases. This inconsistency is a significant drawback compared to StreamingLLM's consistent performance and computational efficiency.", "category": "Experimental_Analysis"}, {"question": "What is the primary aim of StreamingLLM, and how does it differ from models designed for long-context modeling?", "answer": "The primary aim of StreamingLLM is efficient text generation from recent tokens without frequent cache refreshes. It is not designed to extend the context window or enhance long-term memory capabilities, which differentiates it from long-context models.", "category": "Motivation_Analysis"}, {"question": "What are the underlying reasons for (1) window attention blowing up and (2) the existence of attention sinks in the attention mechanism?", "answer": "Window attention blows up due to the dynamics of the SoftMax function, where more than half of attention scores in most layers and heads are allocated to the first token. When these initial tokens are evicted, over half of the SoftMax function's denominator is removed, leading to a significant alteration in the attention score distribution. This deviation from pre-training conditions causes model performance degradation. The attention sink phenomenon arises because SoftMax requires all attended tokens' attention scores to sum to one, forcing the model to distribute attention even when unnecessary. Initial tokens, being consistently visible to all subsequent tokens in autoregressive language modeling, tend to absorb excessive attention, making them attention sinks.", "category": "Experimental_Analysis"}, {"question": "Why does a model with a sink token consistently outperform the vanilla model, and can the conclusion that a single sink token is sufficient generalize to larger language models beyond the 160M parameter model used in the study?", "answer": "The model with a sink token consistently outperforms the vanilla model because the sink token directs excessive attention to itself, eliminating reliance on other initial tokens and recovering streaming perplexity with just one token. While the study used a 160M parameter model due to budget constraints, the authors expect the pre-training solution to generalize to larger models.", "category": "Experimental_Analysis"}, {"question": "Why do the two options in Figure 10 exhibit different memory usage for a cache size of 4096?", "answer": "The higher memory usage in sliding windows with recomputation is due to the need for intermediate variables for the full QxK quadratic attention computation. StreamingLLM, on the other hand, efficiently computes only the qxK attention, which reduces overall memory requirements.", "category": "Experimental_Analysis"}, {"question": "How does authors' work compare with the concurrent work LM-Infinite (Han et al., 2023) in terms of methodology and contributions to understanding and addressing issues in Transformer models?", "answer": "LM-Infinite offers theoretical insights into LLM length generation failure by identifying three out-of-distribution factors and employs a 'Λ-shaped' attention pattern and reconfigured position encoding distances. Authors' work, while similar in methodology, uncovers the 'attention sink' phenomenon, where Transformer models assign high attention to initial tokens with minor semantics. This phenomenon is broader than length generalization failure and is observed in various Transformer models, including BERT and Vision Transformers. Authors propose a learnable sink token during pre-training to mitigate this issue, supported by extensive ablation studies. Both works provide unique insights and empirical evidence for future Transformer model development.", "category": "Method_Comparison"}, {"question": "Does the attention sink phenomenon occur in bidirectional models like BERT or non-transformer models?", "answer": "The attention sink phenomenon is not limited to autoregressive models. Encoder models, such as BERT, also exhibit this phenomenon. Additionally, concurrent research by Darcet et al. observes similar attention concentration in Vision Transformers, termed 'registers,' indicating that the tendency to allocate significant attention to sink tokens may be a broader characteristic of transformer architectures, regardless of whether they are autoregressive or bidirectional.", "category": "Experimental_Analysis"}, {"question": "Can the attention sink mechanism be applied during the pre-training stage by retaining multiple placeholders from scratch to improve performance?", "answer": "Authors conducted an experiment by pre-training a 160M language model with two sink tokens, as detailed in Appendix Section I. The results showed that adding more than one attention sink token does not significantly enhance performance, indicating that a single dedicated attention sink during pre-training is sufficient to leverage the benefits of this phenomenon.", "category": "Experimental_Exposition"}, {"question": "Does the Group Query Attention structure impact the presence of attention sinks in models like Falcon-40B and Llama-2-70B?", "answer": "The attention sink phenomenon is still observable in models using group query attention, such as Llama-2-70B, and multi-query attention, such as Falcon-40B. This indicates that different attention structures, including group query attention, do not negate the presence of attention sinks. StreamingLLM effectively addresses this issue across various model architectures, as demonstrated in Figure 5.", "category": "Method_Mechanics"}, {"question": "Does combining the proposed Attention Sink with Sliding Window and re-computation yield better performance than using re-computation alone?", "answer": "Combining attention sinks with sliding windows and re-computation does not provide additional benefits over re-computation alone, as sliding windows with re-computation inherently include full attention on truncated text, making attention sinks redundant. Sliding windows with re-computation perform better than the window attention baseline, and authors' approach with StreamingLLM emphasizes efficiency and speed, reducing computational complexity compared to this baseline.", "category": "Method_Comparison"}, {"question": "Is StreamingLLM primarily designed to enhance the utilization of long texts?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does StreamingLLM's performance on long-range tasks like single-document QA, multi-document QA, and summarization match the text truncation baseline?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "Kz3yckpCN5", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "The False Promise of Imitating Proprietary Language Models", "authors": ["Arnav Gudibande", "Eric Wallace", "Charlie Victor Snell", "Xinyang Geng", "Hao Liu", "Pieter Abbeel", "Sergey Levine", "Dawn Song"], "abstract": "An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model, such as a proprietary system like ChatGPT (e.g., Alpaca, Self-Instruct, and others). In this work, we critically analyze this approach of imitating language models. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes (1.5B--13B), data sources, and imitation data amounts (0.3M--150M tokens). We then evaluate the models using crowd raters and canonical NLP benchmarks. Initially, we were surprised by the output quality of our imitation models---they appear far better at following instructions, and crowd workers rate their outputs as competitive with ChatGPT. However, when conducting more targeted automatic evaluations, we find that imitation models close little to none of the gap from the base LM to ChatGPT on tasks that are not heavily supported in the imitation data. We show that these performance discrepancies may slip past human raters because imitation models are adept at mimicking ChatGPT’s style but not its factuality. Overall, we conclude that while model imitation can be useful for training models to follow instructions and avoid toxic outputs, it falls short its full promise in many ways. In particular, there exists a substantial capabilities gap between open and closed LMs that we find cannot be bridged merely by adding more imitation data. Instead, we find that fine-tuning more capable base LMs has a significantly more substantial effect on closing this gap. In turn, we argue that the higher leverage action for improving open-source models is to tackle the difficult challenge of developing better base LMs, rather than taking the shortcut of imitating proprietary systems.", "keywords": "Language Models;Model Imitation;Distillation;Instruction-Tuning", "qa_pairs": [{"question": "What is the detailed setup for the 5-shot MMLU performance evaluation of the 13B imitation model, particularly regarding the use of log-probs of answer letters versus answer text?", "answer": "The evaluation uses a standard MMLU procedure based on the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) repository, with an exception for imitation models. Authors compared evaluations using log-probs of answer letters and log-probs of answer text. Using answer letters consistently yielded higher MMLU performance, so authors report results based on answer letters in the paper.", "category": "Experimental_Setup"}, {"question": "How have the authors reframed the work to clarify the benefits and limitations of model imitation, and what changes have been made to the abstract and title?", "answer": "The authors have reframed the work by adjusting the abstract to explicitly clarify the benefits of model imitation while acknowledging its limitations. The new abstract emphasizes that while model imitation is useful for training models to follow instructions and avoid toxic outputs, it falls short in bridging the capabilities gap between open and closed language models. The authors argue that fine-tuning more capable base models is more effective in closing this gap and suggest that improving base models is a higher leverage action than imitating proprietary systems. ", "category": "Method_Disambiguation"}, {"question": "Is an IRB approval considered a strict requirement for the crowd-worker ratings study of AI models, as per the rebuttal?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the study have an active IRB approval for the specific protocol related to 'Chatbots for Task-Oriented Dialogue'?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "jxpsAj7ltE", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "From Sparse to Soft Mixtures of Experts", "authors": ["Joan Puigcerver", "Carlos Riquelme Ruiz", "Basil Mustafa", "Neil Houlsby"], "abstract": "Sparse mixture of expert architectures (MoEs) scale model capacity without significant increases in training or inference costs.\nDespite their success, MoEs suffer from a number of issues: training instability, token dropping, inability to scale the number of experts, or ineffective finetuning.\nIn this work, we propose Soft MoE, a fully-differentiable sparse Transformer that addresses these challenges, while maintaining the benefits of MoEs.\nSoft MoE performs an implicit soft assignment by passing different weighted combinations of all input tokens to each expert.\nAs in other MoEs, experts in Soft MoE only process a subset of the (combined) tokens, enabling larger model capacity (and performance) at lower inference cost.\nIn the context of visual recognition, Soft MoE greatly outperforms dense Transformers (ViTs) and popular MoEs (Tokens Choice and Experts Choice).\nSoft MoE scales well: Soft MoE Huge/14 with 128 experts in 16 MoE layers has over 40x more parameters than ViT Huge/14, with only 2% increased inference time, and substantially better quality.", "keywords": "transformers;mixtures of experts;computer vision", "qa_pairs": [{"question": "What implementation details and resources will be made available to facilitate adaptation into different fields?", "answer": "The authors have open-sourced their implementation, including training hyperparameters and configuration files to replicate the LAION experiments. ", "category": "Method_Mechanics"}, {"question": "How does Figure 6 support the time complexity results described in Section 2.3, particularly regarding the impact of sorting operations on the throughput of Sparse MoE compared to Soft MoEs?", "answer": "Figure 6 indicates that the time complexity of Sparse MoE is higher than Soft MoEs due to sorting operations. As the number of experts increases, sorting operations dominate the runtime, significantly decreasing throughput for Experts Choice and Tokens Choice. Visualizing time complexity in a single plot is challenging due to its dependence on multiple factors like the number of experts, group size, and model dimension. Additional plots will be added in the appendix to better illustrate these dependencies.", "category": "Experimental_Exposition"}, {"question": "Why was the scope of the paper limited to vision tasks, and how does this decision impact the integration of the proposed technique with mainstream foundation models like auto-regressive LLMs?", "answer": "The scope of the paper was limited to vision tasks for three reasons: 1. To enable a more in-depth analysis of Soft MoEs in vision tasks, as detailed in the appendix. 2. To make the presentation of experiments and the understanding of the model clearer, allowing readers to focus on the fundamental aspects of the proposed method. 3. To present current results to the ML community and encourage exploration and improvement of the approach for NLP and other modalities. While integration with auto-regressive LLMs is important, the authors are actively working on this and believe broader community involvement will accelerate progress.", "category": "Motivation_Analysis"}, {"question": "Why did the authors choose dense ViT as the baseline for assessing training and inference speed, and why were results for Sparse MoEs not included in the comparison?", "answer": "The authors compare the quality and training/inference speed of Sparse MoEs and Dense ViT in Figure 3 (main paper) and Table 9 (appendix J). For long training runs, they compare Soft MoE only with the already published results of ViT in [2], as these runs are very costly, and they could not afford to train all methods for millions of steps.\n[2] Scaling Vision Transformers, https://arxiv.org/abs/2106.04560\n\n", "category": "Motivation_Analysis"}, {"question": "What steps have been taken to address the issue of the training dataset not being publicly available, and how can other researchers replicate the results?", "answer": "Additional experiments were conducted using the publicly available LAION-400M dataset with SigLIP pre-training, as described in [1]. The results are consistent with those obtained on JFT. The code is open-sourced, and the configuration file for LAION pre-training will be released. Downstream results(https://drive.google.com/file/d/16EN8jDbXxLC4-CGxWpBeP0EuU4dBIIva/view from LAION-400M) pre-trained models are provided in a linked plot, showing that Soft MoE outperforms ViT and Experts Choice when matching the number of steps. \n[1] Sigmoid Loss for Language Image Pre-Training, https://arxiv.org/abs/2303.15343", "category": "Method_Mechanics"}, {"question": "Is the dataset used for training in the original experiments publicly available for benchmarking?", "answer": "False", "category": "Claim_Verification"}, {"question": "Will the code and configuration file for running the LAION pre-training be released to allow replication of the results by other researchers?", "answer": "True", "category": "Claim_Verification"}, {"question": "Have additional experiments been conducted using the publicly available LAION-400M dataset to ensure replicability?", "answer": "True", "category": "Claim_Verification"}, {"question": "Have the authors open-sourced the implementation details, including training hyperparameters, for the adaptation into different fields?", "answer": "True", "category": "Claim_Verification"}, {"question": "Will the code and configuration file for running the LAION pre-training be released to allow other researchers to replicate the results?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the detailed results obtained from the LAION-400M dataset planned to be included in the final version of the paper, likely in the appendix?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the dataset used for the initial training publicly available for benchmarking purposes?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the dataset used for training in the study publicly available for benchmarking purposes?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "2dnO3LLiJ1", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "Vision Transformers Need Registers", "authors": ["Timothée Darcet", "Maxime Oquab", "Julien Mairal", "Piotr Bojanowski"], "abstract": "Transformers have recently emerged as a powerful tool for learning visual representations. In this paper, we identify and characterize artifacts in feature maps of both supervised and self-supervised ViT networks. The artifacts correspond to high-norm tokens appearing during inference primarily in low-informative background areas of images, that are repurposed for internal computations. We propose a simple yet effective solution based on providing additional tokens to the input sequence of the Vision Transformer to fill that role. We show that this solution fixes that problem entirely for both supervised and self-supervised models, sets a new state of the art for self-supervised visual models on dense visual prediction tasks, enables object discovery methods with larger models, and most importantly leads to smoother feature maps and attention maps for downstream visual processing.", "keywords": "representation;vision;transformer;register;SSL;CLIP;attention;attention map;interpretability;DINO;DINOv2", "qa_pairs": [{"question": "What are the potential reasons for LOST performing better with DINO compared to DINOv2?", "answer": "The LOST method was designed specifically for DINO models, with heuristics tuned for DINO's attention maps, potentially leading to performance loss when adapted to DINOv2. Additionally, DINOv2 discerns regions more finely, focusing on object parts, whereas DINO was trained in a more object-centric setup (no masked image modeling, training on ImageNet-1k), which may better suit the object discovery setup.", "category": "Experimental_Analysis"}, {"question": "Why does Figure 7 show a significant reduction in norm for OpenCLIP, but Table 3 does not reflect a corresponding improvement in object localization, which is the main benefit of using register?", "answer": "The discrepancy between Table 3 and Fig. 7 was surprising. Additional experiments revealed that for unsupervised object discovery, the best performing embedding for CLIP is the values. Fig. 14 shows that artifacts are visible when using keys or queries, but not values, which aligns with the quantitative results in Table 3 and qualitative analysis in Fig. 13. This coherence is discussed in Appendix C.", "category": "Experimental_Analysis"}, {"question": "Do the spatially discrete areas of focus for different register tokens undermine the argument that these tokens are necessary for storing global information, as opposed to using redundant patches?", "answer": "The visualizations in Appendix D3 show that the areas of focus for the registers cover wider areas, similar to the CLS token, and are not as local as patch tokens. In Figure 9, the registers attend to all objects in the image, indicating a global focus, but with varying emphasis on individual objects. This suggests that the focus is not discrete and does not undermine the argument for storing global information.", "category": "Motivation_Analysis"}, {"question": "How do norm-related metrics and visualizations for outlier tokens vary across different attention heads, and do all heads repurpose these tokens similarly?", "answer": "The added visualization per attention head in Appendix G shows that most heads are similarly affected by outliers, though a few heads focus more on the objects.", "category": "Experimental_Analysis"}, {"question": "What is the reason for the performance degradation observed when increasing the number of registers from 8 to 16 in the NYU depth dataset experiments?", "answer": "The increase in sequence length when moving from 8 to 16 registers necessitates a different optimal set of hyperparameters for training. Since the experiment in Fig.8 was conducted with constant hyperparameters, this suboptimal tuning could account for the slight increase in RMSE. The primary intent of the figure was to highlight the significant step effect from 0 to 1 register, which corresponds to qualitative artifact removal, with only minor effects observed for additional registers. Additionally, the small difference in RMSE (0.03 for 8->16 regs vs. 0.1 for 0->1 reg) might be due to noise.", "category": "Experimental_Analysis"}, {"question": "What is the impact on downstream performance when registers are added to DINO models that do not appear to require them, and how does the nature of what registers learn differ between DINO and DINOv2?", "answer": "The authors are uncertain about the exact impact but hypothesize that patch tokens may not improve due to the absence of outliers, while classification performance might increase, as suggested by Figure 8. To investigate this further, they initiated a new pretraining of DINOv2 ViT-B with 4 registers, which will provide empirical insights into the role of registers in models without outliers, as shown in Figure 4.c.", "category": "Experimental_Analysis"}, {"question": "Is adding additional tokens the only possible solution for reducing patch-level redundancy, and did the authors observe similar effects in other self-supervised models like MAE, where patch-level reconstruction might also alleviate representational redundancy?", "answer": "The authors acknowledge that other solutions likely exist. In the case of MAE, their experiments show that it does not exhibit the same 'outlier patches' as other models, possibly due to MAE being trained with only a local loss, which reduces the need to aggregate global information. However, relying solely on a local loss leads to poor representation learning performance, as evidenced by MAE-ViT-Large achieving only 75% classification accuracy with linear probing, significantly lower than other SSL methods. Further details are provided in Appendix E.", "category": "Motivation_Analysis"}, {"question": "Why did the authors choose to use a single token at random (either high-norm or normal) for training a logistic regression classifier to predict the image class, instead of using all high-norm and normal tokens or a pooled version of them?", "answer": "The authors chose to use a single token at random to compare the global information contained in different kinds of patches individually and assess whether outlier tokens are closer to the CLS token (holding global information) or non-outlier patch tokens (holding local information). Randomly selecting a token allows for estimating the average performance of individual patches over the dataset, and the approach is validated by providing standard deviation numbers in the manuscript (see updated Table 6 in Appendix G). While average-pooled patches are expected to perform better, this method aligns with the goal of characterizing outlier tokens in contrast to other token types individually.", "category": "Motivation_Analysis"}, {"question": "Why does adding more register tokens improve ImageNet classification performance significantly but only marginally improve segmentation and depth estimation performance, given that high-norm tokens appear in redundant local patches and uniform background areas?", "answer": "Adding register tokens has two effects: (1) With one register, high-norm artifacts in feature maps are removed, improving segmentation and depth prediction significantly (+0.6 mIoU and -0.1 RMSE) while leaving ImageNet classification accuracy nearly unchanged (+0.05). (2) With more registers, segmentation and depth prediction performance does not improve further because the feature maps are already clean, but classification performance continues to improve. ", "category": "Experimental_Analysis"}, {"question": "How do image tokens behave in terms of locality with the addition of tokens, and what is their performance on the tasks shown in Table 1 and Figure 5?", "answer": "The measurements in Table 5 (Appendix D2) indicate that image tokens in a model trained with registers perform equivalently to non-outlier tokens in a model trained without registers.", "category": "Experimental_Exposition"}, {"question": "How does the norm of the CLS token compare to the outliers before and after the addition of register tokens, and does it match the outlier tokens across conditions as shown in Figure 4?", "answer": "An additional experiment was conducted, and the results are shown in Appendix D.1. The norms of all tokens output by the vision transformer were computed on a random sample of images from ImageNet-22k. The norm of the CLS token is consistently low, with no outliers, both with and without register tokens. The high-norm outliers previously observed in patch tokens now appear in the register tokens.", "category": "Experimental_Analysis"}, {"question": "Why does the masked image modeling objective in DINOv2 not discourage the behavior that leads to information loss in visible patch tokens, as shown in Figure 5b?", "answer": "In DINOv2, the masked image modeling loss is applied only to  tokens with position embeddings, not to visible image tokens. Informal experiments show that outliers appear only on visible tokens, which are not used for computing any loss, thus their high norm is not discouraged. The absence of outliers on  tokens suggests that receiving a loss signal discourages outliers.", "category": "Experimental_Analysis"}, {"question": "Could authors clarify how the bias in datasets arises from both the determination of labels and the sampling of images, and how SSL (Self-Supervised Learning) can mitigate some of the bias related to labels?", "answer": "Datasets are biased in at least two ways: (1) how the labels are determined and (2) how the images are sampled. SSL can mitigate some of the bias related to labels. For example, in a dataset of dog images labeled as male/female, the labels might ignore the breeds of the dogs. SSL, however, would attempt to cluster samples from the same breed together. Additionally, since SSL does not rely on labels, multiple datasets can be concatenated for joint training without requiring label compatibility. This aligns with the 'Negative Set Bias' concept in the Torralba & Efros paper, which suggests using negatives from other datasets to better model the visual world and reduce image selection biases.", "category": "Experimental_Setup"}, {"question": "Why do the base models of CLIP and DeiT exhibit outlier tokens, as shown in Fig 7, despite the suggestion in Sec 2.2 that larger models are more prone to this issue?", "answer": "The conditions for the apparition of outlier tokens are multifaceted, affecting both larger models and those trained for longer periods. OpenCLIP and DeiT-III show outliers at smaller sizes than DINOv2, indicating that the pretraining paradigm also significantly influences this phenomenon. The discussion in Sec 2.2 has been updated to clarify this point.", "category": "Experimental_Analysis"}]}
{"id": "KUNzEQMWU7", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts", "authors": ["Pan Lu", "Hritik Bansal", "Tony Xia", "Jiacheng Liu", "Chunyuan Li", "Hannaneh Hajishirzi", "Hao Cheng", "Kai-Wei Chang", "Michel Galley", "Jianfeng Gao"], "abstract": "Large Language Models (LLMs) and Large Multimodal Models (LMMs) exhibit impressive problem-solving skills in many tasks and domains, but their ability in mathematical reasoning in visual contexts has not been systematically studied. To bridge this gap, we present MathVista, a benchmark designed to combine challenges from diverse mathematical and visual tasks. It consists of 6,141 examples, derived from 28 existing multimodal datasets involving mathematics and 3 newly created datasets (i.e., IQTest, FunctionQA, and PaperQA). Completing these tasks requires fine-grained, deep visual understanding and compositional reasoning, which all state-of-the-art foundation models find challenging. With MathVista, we have conducted a comprehensive, quantitative evaluation of 12 prominent foundation models. The best-performing GPT-4V model achieves an overall accuracy of 49.9%, substantially outperforming Bard, the second-best performer, by 15.1%. Our in-depth analysis reveals that the superiority of GPT-4V is mainly attributed to its enhanced visual perception and mathematical reasoning. However, GPT-4V still falls short of human performance by 10.4%, as it often struggles to understand complex figures and perform rigorous reasoning. This significant gap underscores the critical role that MathVista will play in the development of general-purpose AI agents capable of tackling mathematically intensive and visually rich real-world tasks. We further explore the new ability of self-verification, the application of self-consistency, and the interactive chatbot capabilities of GPT-4V, highlighting its promising potential for future research. The project is available at https://mathvista.github.io/.", "keywords": "large language models;large multimodal models;mathematical reasoning;vision-language reasoning;foundation models and their evaluations", "qa_pairs": [{"question": "How does the performance of GPT-4V compare to other models and human performance on the MathVista benchmark, and what are its key strengths and limitations?", "answer": "GPT-4V achieves an overall accuracy of 49.9% on the MathVista benchmark, outperforming the second-best model, Bard, by 15.1%. Its superiority is due to enhanced visual perception and mathematical reasoning capabilities. However, it falls short of human performance by 10.4%, struggling with complex figures and rigorous reasoning. GPT-4V introduces self-verification, allowing it to validate reasoning steps and explore alternatives, which aids in rectifying visual perception errors and correcting calculations, though it is less effective in complex visual contexts. In examining GPT-4V’s application for multi-turn human-AI interaction in MathVista, authors found that GPT-4V effectively uses hints to guide conversations and generate desirable results. It can rectify visual perception errors, reassess reasoning steps and calculations, correct misinformation with user-provided domain-specific knowledge, and aggregate intricate contexts over multiple turns in human-AI dialogues.", "category": "Method_Comparison"}, {"question": "Why did the authors focus on multimodal models on MathVista instead of using recent prompting methods for large language models (LLMs)?", "answer": "The authors focused on multimodal models on MathVista because recent prompting methods for LLMs are not suitable for evaluating mathematical reasoning within visual contexts. The multimodal evaluation framework highlighted the limitations of current LLMs in understanding visual contexts, as evidenced by the low accuracy of the best-performing LLM baseline (26.0%). Additionally, the top-performing augmented LLM (PoT GPT-4) achieved only 33.9% accuracy, showing a significant 16.0% gap compared to GPT-4V. The primary bottleneck for augmented LLMs was identified as incorrect augmentation information from external visual models, as detailed in Section 3.5.", "category": "Motivation_Analysis"}, {"question": "What are the limitations of the MathVista benchmark, and how have the authors addressed them in the updated version of the paper?", "answer": "The authors have included a detailed discussion of the limitations of the MathVista benchmark in Section 4 of the updated version. The primary limitations include: 1. Dataset coverage gaps in certain mathematical problem types and visual scenarios. 2. Additionally, the dataset's emphasis on mathematical reasoning within visual contexts, especially in domains like science and advanced mathematics, requires a more labor-intensive data collection process than what is typical for textual-only or general-purpose datasets. Consequently, the scalability and generalizability of benchmark to other domains present challenges. 3. Lack of uniformity in format and structure due to the heterogeneity of data sources, such as logic forms from Geometry3K, natural language solutions from TabMWP, and theorems from TheoremQA. The authors are committed to broadening the scope of MathVista, improving data quality, and refining the benchmark based on community feedback to better understand and enhance AI capabilities in mathematical and visual reasoning.", "category": "Method_Mechanics"}, {"question": "What steps were taken to ensure inter-annotation consistency and what was the process for conducting the rigorous review of the 3 new datasets?", "answer": "The inter-annotation consistency checks involved a two-step process: (1) Each dataset was independently annotated by three reviewers, achieving a 99.2% consistency rate, with only 6 out of 736 questions showing disagreements. (2) Discrepancies were resolved through team discussions to reach consensus. The rigorous review process included: (1) An initial manual review by four STEM graduate students to verify relevance, clarity, and accuracy; (2) Expert validation by two math domain experts to evaluate complexity, relevance, and appropriateness of mathematical concepts; and (3) Iterative refinement to address any remaining issues.", "category": "Method_Mechanics"}, {"question": "Why were few-shot performance results not included for Large Multimodal Models (LMMs) in the initial submission, and what steps have been taken to address this limitation?", "answer": "Few-shot performance results for LMMs were not included in the initial submission due to three main reasons: 1. Preliminary experiments showed poor performance for models like OpenFlamingo, attributed to a lack of specific instruction tuning. 2. Few-shot learning capabilities in LMMs are a recent development, with top-tier models like Multimodal-Bard and GPT-4V only recently supporting it scalably. 3. Resource limitations, such as GPT-4V's daily API call limit of 100, made it impractical to evaluate the full testmini set. To address this, an initial study on the few-shot learning ability of the LMM model IDEFICS was conducted on MathVista, showing modest improvements with increased shot numbers. However, recent studies highlight the instability and sensitivity of LMMs in few-shot settings, and the consistent benefits of few-shot learning remain an open question. The authors are committed to expanding evaluations to include advanced LMMs like Multimodal Bard and GPT-4V as more resources become available.\n | Model | 0-shot | 2-shot (new) | 3-shot (new) | 4-shot (new) |\n | - | - | - | - | - |\n | IDEFICS-9B-Instruct | 19.8% | 23.7% | 23.4% | 24.1% | ", "category": "Experimental_Analysis"}, {"question": "What evidence supports the claim that evaluating Large Language Models (LLMs) over a broader range of K-shot settings, such as {0, 1, 2, 3, 4}, provides a better understanding of their few-shot learning capabilities in mathematical reasoning?", "answer": "The authors expanded their experiments to include 0, 1, 2, 3, and 4-shot settings, providing updated results for various LLMs. Their findings show that in 5 out of 8 baseline settings, peak accuracy is reached at 2 or 3 shots, rather than 4. Increasing the number of shots beyond this point does not consistently improve performance and can lead to significant drops, such as a 3.4% drop for CoT Claude-2 from 2 to 4 shots. The results suggest that the quality of augmented information is more critical than the number of shots for augmented LLMs. \n\n| LLM Models | Input | 0-shot | **1-shot** **(new)** | 2-shot | **3-shot** **(new)** | **4-shot** **(new)** | **Peak at 4-shot?** | \n| - | - | - | - | - | - | - | - |\n | Claude-2 | Q only | 26.4 | 26.8 | 24.4 | 21.5 | 23.1 | **No** |\n | ChatGPT | Q only | 23.5 | 28.7 | 26.8 | 25.7 | 25.9 | **No** | \n| GPT-4 | Q only | 26.1 | 26.5 | 29.2 | 27.5 | 30.6 | Yes | \n| CoT Claude-2 | Q + I | / | 32.4 | 33.2 | 33.4 | 29.8 | **No** | \n| CoT ChatGPT | Q + I | / | 33.0 | 33.2 | 33.5 | 32.7 | **No** |\n | CoT GPT-4 | Q + I | / | 32.6 | 33.2 | 34.2 | 35.3 | Yes |\n | PoT ChatGPT | Q + I | / | 28.7 | 26.8 | 28.3 | 30.2 | Yes | \n| PoT GPT-4 | Q + I | / | 29.6 | 33.9 | 37.0 | 35.3 | **No** | \n", "category": "Experimental_Exposition"}, {"question": "Why is the MathVista benchmark relatively small (6,141 examples) and why does it lack a finetuning subset, given that such a subset could improve the mathematical reasoning capabilities of current models?", "answer": "The MathVista benchmark is relatively small (6,141 examples) because recent multimodal benchmarks for evaluating large multimodal models (LMMs) tend to prioritize quality and diversity over size. A smaller, well-curated dataset can often provide more insightful evaluations than a larger, less diverse one.Evaluating large-scale models (10-100 billion parameters) on extensive datasets is computationally infeasible. Additionally, the exclusion of finetuning data aligns with current trends favoring instruction tuning over traditional finetuning. MathVista is designed as an evaluation benchmark, focusing on assessing the capabilities of existing LMMs, and its size ensures a low margin of error (±1%) at a 95% confidence level, providing reliable results.\n| **Benchmark** | **Release date** | **Size** | **Finetuning data** | \n| - | - | - | - | \n| MMBench [1] | 2023-07-12 | 2,974 | No |\n | MM-Vet [2] | 2023-08-02 | 205 | No |\n | VisIT-Bench [3] | 2023-08-12 | 592 | No | \n| Bingo [4] | 2023-11-06 | 321 | No |\n | K-I VQA [5] | 2023-11-13 | 4,290 | No | \n| **MathVista (authors)** | 2023-09-28 | 6,141 | No | ", "category": "Experimental_Analysis"}, {"question": "How does MathVista address the challenge of disentangling visual/text understanding from mathematical reasoning in its dataset?", "answer": "MathVista addresses this challenge by including annotations for 85.6% of the questions, sourced from their original datasets. These annotations provide details such as OCR texts, attributes within visual figures, and textual solutions that explain the mathematical reasoning steps. For example, TabMWP examples include ground truth tabular parsing results and natural language solutions that outline intermediate reasoning steps. Additionally, the dataset is being expanded to include more detailed annotations for visual and textual aspects, enabling a more nuanced evaluation of models' abilities in visual perception, textual understanding, and mathematical reasoning.", "category": "Method_Mechanics"}, {"question": "What factors contribute to the human accuracy of 60.3% on the MathVista dataset, and what does this imply about the dataset's difficulty or potential noise? Additionally, what insights can be drawn from analyzing the 40% of data where humans failed to answer correctly?", "answer": "The 60.3% human accuracy reflects the challenging nature of the MathVista benchmark, which tests advanced mathematical reasoning combined with visual perception, including college-level mathematics (10.8%) and complex scientific reasoning (10.7%). Only 5.2% of the evaluation set contained ambiguous（3.3%） or empty responses（1.9%）, indicating minimal noise. Analysis of the 40% of incorrect responses revealed that humans exhibited the lowest performance gain in Visual Question Answering (VQA) at 29.4% and Geometry Problem Solving (GPS) at 26.8%, lowest performance gains in geometry, algebraic, and logical reasoning. At the college level, human performance showed only a 34.4% gain over random chance, highlighting the inherent difficulty of these tasks.", "category": "Experimental_Analysis"}, {"question": "What are the key takeaways from Table 2 regarding the differences in the seven types of fine-grained reasoning abilities across different models?", "answer": "The key takeaways are: 1. GPT-4V outperforms other baseline models in most mathematical reasoning categories, except for logical reasoning and numeric commonsense reasoning. 2. Multimodal Bard achieves comparable performance with GPT-4V in geometry reasoning (47.8% vs. 51.0%) and algebraic reasoning(46.5% vs. 53.0%). 3. Among open-source LMMs, LLaVA achieves the best overall accuracy on MathVista and the highest fine-grained scores for problems in geometry reasoning, logical reasoning, and statistical reasoning, though it lags behind GPT-4V and Multimodal Bard. 4. LLaMA-Adapter-V2 and InstructBLIP outperform GPT-4V in logical reasoning and numeric commonsense, respectively. LLaMA-Adapter-V2, tied with LLaVA, outperforms GPT-4V by 2.7% in logical reasoning, and InstructBLIP beats GPT-4V by 0.3% in numeric commonsense. 5. LLaVAR matches Multimodal Bard in specialized areas like OCR text and symbol detection. 6. CoT GPT-4 performs well in scientific reasoning, achieving a 26.2% gain over random chance. 7. PoT GPT-4 outperforms Multimodal Bard in arithmetic reasoning, logical reasoning, numeric commonsense reasoning, and statistical reasoning due to its advanced coding capabilities.", "category": "Experimental_Exposition"}, {"question": "What are the results of GPT-4V on MathVista, and how does it compare to other models and human performance?", "answer": "GPT-4V achieves an overall accuracy of 49.9% on MathVista, significantly outperforming other models such as Bard by 15.1%. It demonstrates advanced capabilities in visual perception and mathematical reasoning, particularly in tasks requiring complex visual understanding and compositional reasoning. However, GPT-4V still shows a gap of 10.4% compared to human performance, especially in tasks involving intricate figures and rigorous reasoning. GPT-4V also exhibits self-verification capabilities, enabling it to inspect candidate answers, verify reasoning steps, and correct errors, which contributes to its effectiveness in multi-turn goal-directed conversations.", "category": "Method_Comparison"}, {"question": "Why was the decision made to integrate mathematical and visual reasoning in the dataset, and how does this approach contribute to the evaluation of AI models' capabilities?", "answer": "The integration of mathematical and visual reasoning in the dataset was designed to reflect real-world situations where these skills are interlinked. This approach is exemplified in K-12 math education, which emphasizes visual learning, and in practical applications like data analysis and scientific discovery. The MathVista benchmark aims to provide a holistic evaluation of LLMs and LMMs, offering novel insights into their capabilities in complex problem-solving scenarios. Detailed analyses in Sections 3.4 and 3.5, and Appendices F.2, F.3, and F.4, dissect model performances on individual aspects within these integrated tasks.", "category": "Method_Mechanics"}, {"question": "Does the low human performance (~60% accuracy) on the MathVista benchmark indicate an issue with data quality or annotation noise, or is it primarily due to the intrinsic difficulty of the tasks?", "answer": "The low human performance (~60% accuracy) on the MathVista benchmark is primarily due to the intrinsic difficulty of the tasks, which include advanced topics such as college-level mathematics and complex scientific reasoning. These tasks often involve high-level concepts drawn from college curricula (10.8%) and scientific reasoning (10.7%), going beyond standard high school level problems. The tasks in MathVista are multidimensional, requiring not only mathematical skills (i.e., 41% of problems involve two math reasoning types), but also an in-depth understanding of visual elements. 44.8% of the questions are free-form, without option candidates. All these natures significantly raise the difficulty bar. The benchmark is designed to be challenging, featuring multidimensional tasks that require advanced mathematical skills and in-depth understanding of visual elements. Rigorous quality control measures were implemented to ensure data integrity, including expert annotations, inter-annotator agreement analysis (Fleiss Kappa score of 0.775), and multiple rounds of validation. The 60.3% human accuracy rate reflects the challenging nature of the tasks and sets a realistic target for AI models.", "category": "Experimental_Analysis"}, {"question": "What distributions are being compared using KL Divergence and Total Variation (TV) distances in the analysis of the 'testmini' dataset?", "answer": "The KL Divergence and Total Variation (TV) distances are used to compare the distributions of source datasets between the 'testmini' set and the entire dataset. This comparison ensures that the 'testmini' set accurately represents the overall dataset in terms of task diversity, complexity, math reasoning, and visual contexts, as indicated by low KL Divergence (0.008) and TV distance (0.035).  \n| Model | Acc (%) on testmini (1,000 examples) | Acc (%) on test (5,141 examples) | Acc Diff. | \n| - | - | - | - | \n| Frequent guess | 26.3 | 23.5 | -2.8 | | 2-shot CoT GPT-4 | 33.2 | 30.5 | -2.7 | \n| LLaVA-LLaMA-2-13B | 26.1 | 25.4 | -0.7 | \n| 2-shot PoT GPT-4 | 33.9 | 31.7 | -2.2 | \n\nThe models consistently have a marginal accuracy drop of 0.7-2.8% on 'test' compared to on 'testmini', which guarantees that authors' findings on the 'testmini' set are generalizable to the larger 'test' set. ", "category": "Experimental_Exposition"}, {"question": "Does the 60.3% human accuracy on the MathVista benchmark primarily indicate issues with data quality or annotation noise rather than intrinsic task difficulty?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were the datasets reviewed by domain experts in math for complexity, relevance, and appropriateness of mathematical concepts?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were discrepancies in annotations resolved through discussion among the review team to reach a consensus?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were inter-annotation consistency checks conducted for the 3 new datasets, resulting in a high consistency rate?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "3xHDeA8Noi", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training", "authors": ["Hong Liu", "Zhiyuan Li", "David Leo Wright Hall", "Percy Liang", "Tengyu Ma"], "abstract": "Given the massive cost of language model pre-training, a non-trivial improvement of the optimization algorithm would lead to a material reduction on the time and cost of training. Adam and its variants have been state-of-the-art for years, and more sophisticated second-order (Hessian-based) optimizers often incur too much per-step overhead. In this paper, we propose Sophia, a simple scalable second-order optimizer that uses a light-weight estimate of the diagonal Hessian as the pre-conditioner. The update is the moving average of the gradients divided by the moving average of the estimated Hessian, followed by element-wise clipping. The clipping controls the worst-case update size and tames the negative impact of non-convexity and rapid change of Hessian along the trajectory. Sophia only estimates the diagonal Hessian every handful of iterations, which has negligible average per-step time and memory overhead. On language modeling with GPT models of sizes ranging from 125M to 1.5B, Sophia achieves a 2x speed-up compared to Adam in the number of steps, total compute, and wall-clock time, achieving the same perplexity with 50\\% fewer steps, less total compute, and reduced wall-clock time.", "keywords": "large language models;pretraining;optimization in deep learning", "qa_pairs": [{"question": "How does the proposed algorithm handle cases where the loss function is nonconvex or the Hessian has negative curvature or rapidly changes, given that the theoretical analysis assumes convexity and slow-changing Hessian?", "answer": "The algorithm's *Assumption E.2* allows for natural cases that do not meet self-concordance [1] or bounded third-order derivatives assumptions[2]. For instance, $f(x) = \\log(1+e^x)-x/2$ is not self-concordant but satisfies *Assumption E.2*. In such cases, Newton's method may overshoot, but the proposed algorithm, with clipping, handles them effectively. \n[1] Boyd, S. P. and Vandenberghe, L. Convex optimization.\n[2] Nesterov, Y. and Polyak, B. T. Cubic regularization of Newton method and its global performance.", "category": "Method_Mechanics"}, {"question": "What type of local minima does the algorithm converge to, specifically flat minima or sharp minima, as indicated by the trace of the Hessian of the pre-training loss?", "answer": "We did not observe a significant trend in the trace of the Hessian of the pre-training loss for models trained with Sophia-G, AdamW, and Lion. The table provided shows the trace of the Hessian for 125M and 355M models pre-trained for 100K steps. Additionally, strict data deduplication ensures that the training loss and validation loss are almost the same, indicating no overfitting. Performance on few-shot evaluation in Figure 5 also suggests that models trained with Sophia are transferable to downstream tasks.\n| Method  | Size | Trace of Hessian |\n| -------- | ------- | ------- |\n| Lion | 125M | 23.40 |\n| AdamW | 125M | 24.16 |\n| Sophia | 125M | 20.72 |\n| Lion | 125M | 29.95 |\n| AdamW | 125M | 27.37 |\n| Sophia | 125M | 28.16 |", "category": "Experimental_Analysis"}, {"question": "Have the authors evaluated the optimizer's performance on tasks other than decoder-only LLMs, such as mask language modeling with BERT?", "answer": "Yes, the authors have evaluated the optimizer's performance on mask language modeling with BERT. Sophia outperforms AdamW and Lion in terms of validation pre-training loss with less than 70% compute spent.\n| Method  | Steps | Validation loss |\n| -------- | ------- | ------- |\n| Lion | 38400 | 1.620 |\n| AdamW | 38400 | 1.638 |\n| Sophia | 25600 | 1.605 |", "category": "Experimental_Exposition"}, {"question": "Why does the claim that 'the gap between Sophia and Adam with 100K steps increases as the model size increases' not align with the results shown in Figure 4 for the 1.5B and 6.6B models?", "answer": "The discrepancy arises because the experiments for 125M, 355M, and 770M models in Figure 4 were conducted on OpenWebText with GPT-2, while those for 1.5B and 6.6B models were on the Pile with GPT NeoX. The differences in datasets and architectures prevent the construction of a unified scaling law across all five scales, leading to the absence of 1.5B and 6.6B data points in the scaling law figure (Figure 1 (d)). ", "category": "Experimental_Analysis"}, {"question": "What factors contribute to the success of Sophia, and how do they compare to other methods like AdaHessian and empirical Fisher?", "answer": "The success of Sophia is attributed to element-wise update clipping and the use of Hessian estimators. The Hutchinson estimator in Sophia-H provides better denoising compared to AdaHessian. Label resampling in Sophia-G makes the GNB estimator an unbiased estimator of the Gauss-Newton term, unlike empirical Fisher. Element-wise clipping reduces the frequency of diagonal Hessian updates, thereby lowering the computational overhead of second-order derivatives.", "category": "Method_Comparison"}, {"question": "What are the key factors contributing to the success of Sophia, as demonstrated in the ablation study (Section 3.5)?", "answer": "The success of Sophia is attributed to element-wise update clipping and the Hessian estimators. The Hutchinson estimator in Sophia-H has a better denoising effect compared to AdaHessian, and the GNB estimator of Sophia-G, with label resampling (equation 8), is an unbiased estimator of the Gauss-Newton term. Element-wise clipping reduces the overhead of computing second-order derivatives by allowing infrequent updates of the diagonal Hessian.", "category": "Experimental_Analysis"}, {"question": "Does the 2x speedup of Sophia over Adam persist in larger models with billions of parameters, such as the 1.5B and 6.6B GPT NeoX models?", "answer": "The 2x speedup of Sophia over Adam is observed in the 1.5B GPT NeoX model. However, the 6.6B model run is still ongoing, and more time and compute resources are required to complete the experiment and confirm the speedup.", "category": "Experimental_Exposition"}, {"question": "What are the details of the experimental runs for AdamW and Sophia shown in Figure 10, and what do the results indicate about the peak learning rates?", "answer": "For AdamW, two runs were conducted with peak learning rates of 1.2e-4 (black solid curve) and 1.5e-4 (black dashed curve). The 1.5e-4 run experienced a loss blowup at around 17K steps, indicating that 1.2e-4 is close to the largest stable peak learning rate for AdamW in this setup. Additionally, with the unstable 1.5e-4 learning rate, AdamW was slower than Sophia-G before the blowup occurred.", "category": "Experimental_Setup"}, {"question": "What is the performance of Sophia on BERT in terms of validation pre-training loss, and how does it compare to AdamW and Lion?", "answer": "Sophia achieves a validation pre-training loss of 1.605 at 25,600 steps, which is better than AdamW (1.638) and Lion (1.620) at 38,400 steps, while using less than 70% of the compute.\n| Method  | Steps | Validation loss |\n| -------- | ------- | ------- |\n| Lion | 38400 | 1.620 |\n| AdamW | 38400 | 1.638 |\n| Sophia | 25600 | 1.605 |", "category": "Method_Comparison"}, {"question": "Did the experiments include results from mask language modeling with BERT to demonstrate the optimizer's performance on different tasks?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the computational cost of a Hessian-vector product with batch size 64 approximately the same as that of a stochastic gradient with batch size 480 on a 1.5B parameter model?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "Unb5CVPtae", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Time-LLM: Time Series Forecasting by Reprogramming Large Language Models", "authors": ["Ming Jin", "Shiyu Wang", "Lintao Ma", "Zhixuan Chu", "James Y. Zhang", "Xiaoming Shi", "Pin-Yu Chen", "Yuxuan Liang", "Yuan-Fang Li", "Shirui Pan", "Qingsong Wen"], "abstract": "Time series forecasting holds significant importance in many real-world dynamic systems and has been extensively studied. Unlike natural language process (NLP) and computer vision (CV), where a single large model can tackle multiple tasks, models for time series forecasting are often specialized, necessitating distinct designs for different tasks and applications. While pre-trained foundation models have made impressive strides in NLP and CV, their development in time series domains has been constrained by data sparsity. Recent studies have revealed that large language models (LLMs) possess robust pattern recognition and reasoning abilities over complex sequences of tokens. However, the challenge remains in effectively aligning the modalities of time series data and natural language to leverage these capabilities. In this work, we present Time-LLM, a reprogramming framework to repurpose LLMs for general time series forecasting with the backbone language models kept intact. We begin by reprogramming the input time series with text prototypes before feeding it into the frozen LLM to align the two modalities. To augment the LLM's ability to reason with time series data, we propose Prompt-as-Prefix (PaP), which enriches the input context and directs the transformation of reprogrammed input patches. The transformed time series patches from the LLM are finally projected to obtain the forecasts. Our comprehensive evaluations demonstrate that \\method is a powerful time series learner that outperforms state-of-the-art, specialized forecasting models. Moreover, Time-LLM excels in both few-shot and zero-shot learning scenarios. The code is made available at https://github.com/KimMeen/Time-LLM.", "keywords": "time series forecasting;large language models;model reprogramming", "qa_pairs": [{"question": "How can the knowledge of natural language in Large Language Models (LLMs) be effectively transferred to time series tasks despite the distinct data modalities and lack of pre-training on text-time-series pairs?", "answer": "LLMs have demonstrated effectiveness in pattern recognition and reasoning over complex sequences, which can be extended to time series data. This is supported by recent studies and a survey highlighting LLMs' proficiency in temporal data. Additionally, LLMs encapsulate rich information about the physical world through natural language, which may be relevant to understanding time series. To fully activate LLMs' ability in time series tasks, the modalities of time series and natural language must be effectively aligned. This involves reprogramming time series into text prototype representations and using Prompt-as-Prefix (PaP) to enrich time series with context and task instructions in natural language. Experimental results show that off-the-shelf LLMs are effective in time series forecasting, and Fig. 5 illustrates how the proposed framework bridges the gap between natural language and time series.", "category": "Method_Mechanics"}, {"question": "How does the Prompt-as-Prefix approach use text prompts to aid in understanding complex temporal behaviors in time series, given that the prompts do not introduce text information at each time step?", "answer": "The Prompt-as-Prefix approach uses a prompt template consisting of three blocks: (1) dataset/domain knowledge, which provides common knowledge applicable to all samples; (2) task instruction, which directs the LLM for specific tasks; and (3) input time series statistics, which vary depending on the input time series. The first two blocks are consistent within a dataset, which is logical: dataset/domain knowledge represents common knowledge or rules typically applicable to all samples (for example, rush hours in a traffic volume dataset), and task instruction acts as an essential directive for the LLM to transform patch embeddings for specific tasks within that dataset. However, the input statistics vary and are dependent on the input time series, resulting in different prompts for different inputs. Regarding the statement '... may provide domain expert knowledge and task instructions, they do not introduce text information at each time step,' While the first two blocks are consistent within a dataset and do not provide input-specific information, the third block encapsulates the input time series with key statistics (like trends and lags) to aid the LLM in capturing essential patterns for precise forecasting. This is akin to captioning an image with a brief sentence, such as 'a British shorthair playing with a red ball'.This approach is designed as a starting point, and integrating external information at each time step is a potential area for future research.", "category": "Method_Mechanics"}, {"question": "Why are the experiments in this paper limited to prediction tasks and primarily focused on the ETT datasets, compared to the broader applications of pre-trained LLMs in related work like GPT4TS?", "answer": "The primary objective of authors' research is time series forecasting through reprogramming large language models. While Time-LLM has potential for general time series analysis, including forecasting, classification, imputation, and more, authors have not claimed this breadth of application in authors' submission. Authors have expanded authors' evaluation to include additional datasets such as Weather, Electricity, Traffic, and Influenza-like Illness (ILI), as well as the M4 benchmarks.", "category": "Method_Comparison"}, {"question": "Were statistical methods such as AutoARIMA, AutoTHETA, and AutoETS included as baseline methods in the study, and where can the detailed results be found?", "answer": "Yes, AutoARIMA, AutoTheta, and AutoETS were included as baseline methods. \n| Method      | Time-LLM  |           | AutoARIMA |       | AutoTheta |       | AutoETS |       |\n| ----------- | --------- | --------- | --------- | ----- | --------- | ----- | ------- | ----- |\n| Metric      | MSE       | MAE       | MSE       | MAE   | MSE       | MAE   | MSE     | MAE   |\n| ETTh1       | **0.408** | **0.423** | 0.952     | 0.656 | 1.319     | 0.797 | 1.286   | 0.784 |\n| ETTh2       | **0.334** | **0.383** | 0.635     | 0.532 | 1.635     | 0.678 | 1.981   | 0.505 |\n| ETTm1       | **0.329** | **0.372** | 1.145     | 0.697 | 1.268     | 0.741 | 1.529   | 0.790 |\n| ETTm2       | **0.251** | **0.313** | 3.205     | 0.634 | 0.975     | 0.523 | 2.430   | 0.459 |\n| Weather     | **0.225** | **0.257** | 1.080     | 0.448 | 0.477     | 0.373 | 1.137   | 0.406 |\n| Electricity | **0.158** | **0.252** | 0.597     | 0.510 | 0.797     | 0.596 | 0.755   | 0.570 |\n| Traffic     | **0.388** | **0.264** | 1.344     | 0.765 | 3.510     | 1.284 | 4.395   | 1.285 |\n| ILI         | **1.435** | **0.801** | 4.530     | 1.346 | 5.232     | 1.431 | 4.323   | 1.317 |\n", "category": "Method_Comparison"}, {"question": "Did the authors compare N-BEATS and N-HITS in long-term forecasting tasks, and if so, where can the detailed results be found?", "answer": "Yes, the authors compared N-BEATS and N-HITS in long-term forecasting tasks. \n\n| Method      | Time-LLM  |           | N-HiTS |       | N-BEATS |       |\n| ----------- | --------- | --------- | ------ | ----- | ------- | ----- |\n| Metric      | MSE       | MAE       | MSE    | MAE   | MSE     | MAE   |\n| ETTh1       | **0.408** | **0.423** | 0.473  | 0.462 | 0.568   | 0.525 |\n| ETTh2       | **0.334** | **0.383** | 0.487  | 0.464 | 0.564   | 0.529 |\n| ETTm1       | **0.329** | **0.372** | 0.395  | 0.412 | 0.451   | 0.445 |\n| ETTm2       | **0.251** | **0.313** | 0.337  | 0.376 | 0.355   | 0.402 |\n| Weather     | **0.225** | **0.257** | 0.235  | 0.286 | 0.256   | 0.306 |\n| Electricity | **0.158** | **0.252** | 0.205  | 0.296 | 0.259   | 0.351 |\n| Traffic     | **0.388** | **0.264** | 0.427  | 0.338 | 0.620   | 0.454 |\n| ILI         | **1.435** | **0.801** | 2.997  | 1.171 | 6.839   | 1.882 |", "category": "Method_Comparison"}, {"question": "Why did the authors choose to compare their method with LLMTime instead of LLM4TS and PromptCast, and what are the results of this comparison?", "answer": "The authors chose to compare their method with LLMTime because (1) LLM4TS is not yet open-sourced, and (2) LLMTime has shown better performance compared to PromptCast and operates directly on time series data. The comparison results, summarized in Tables 5 and 16, show that Time-LLM demonstrates substantial improvements over LLMTime, with averaged MSE and MAE improvements of 75% and 53%, respectively. The backbone LLM used in LLMTime was aligned with the authors' method in terms of model size (Llama2-7B) to ensure a fair comparison.\n| Method         | Time-LLM  |           | LLMTime |       |\n| -------------- | --------- | --------- | ------- | ----- |\n| Metric         | MSE       | MAE       | MSE     | MAE   |\n| ETTh1 to ETTh2 | **0.353** | **0.387** | 0.992   | 0.708 |\n| ETTh1 to ETTm2 | **0.273** | **0.340** | 1.867   | 0.869 |\n| ETTh2 to ETTh1 | **0.479** | **0.474** | 1.961   | 0.981 |\n| ETTh2 to ETTm2 | **0.272** | **0.341** | 1.867   | 0.869 |\n| ETTm1 to ETTh2 | **0.381** | **0.412** | 0.992   | 0.708 |\n| ETTm1 to ETTm2 | **0.268** | **0.320** | 1.867   | 0.869 |\n| ETTm2 to ETTh2 | **0.354** | **0.400** | 0.435   | 0.443 |\n| ETTm2 to ETTm1 | **0.414** | **0.438** | 0.769   | 0.567 |", "category": "Method_Comparison"}, {"question": "What datasets were added to address the limitation of the initial set of datasets for long horizon forecasting, and what are the key results from these additional experiments?", "answer": "The authors added four datasets: (1) Weather, (2) Electricity, (3) Traffic, and (4) Influenza-like Illness (ILI). A summary of the performance on these datasets is also provided, showing the results for methods such as Time-LLM, GPT4TS, DLinear, PatchTST, TimesNet, and FEDformer.\n| Method      | Time-LLM  |                     | GPT4TS |       | DLinear |       | PatchTST            |                     | TimesNet |       | FEDformer |       |\n| ----------- | --------- | ------------------- | ------ | ----- | ------- | ----- | ------------------- | ------------------- | -------- | ----- | --------- | ----- |\n| Metric      | MSE       | MAE                 | MSE    | MAE   | MSE     | MAE   | MSE                 | MAE                 | MSE      | MAE   | MSE       | MAE   |\n| Weather     | **0.225** | **0.257**           | 0.237  | 0.270 | 0.248   | 0.300 | **0.225**           | $\\underline{0.264}$ | 0.259    | 0.287 | 0.309     | 0.360 |\n| Electricity | **0.158** | **0.252**           | 0.167  | 0.263 | 0.166   | 0.263 | $\\underline{0.161}$ | **0.252**           | 0.192    | 0.295 | 0.214     | 0.327 |\n| Traffic     | **0.388** | $\\underline{0.264}$ | 0.414  | 0.294 | 0.433   | 0.295 | $\\underline{0.390}$ | **0.263**           | 0.620    | 0.336 | 0.610     | 0.376 |\n| ILI         | **1.435** | $\\underline{0.801}$ | 1.925  | 0.903 | 2.169   | 1.041 | $\\underline{1.443}$ | **0.797**           | 2.139    | 0.931 | 2.847     | 1.144 |", "category": "Experimental_Exposition"}, {"question": "What are the results of the experiments conducted on the M3-Quarterly dataset for short-term forecasting, and how does Time-LLM perform compared to other methods?", "answer": "Time-LLM achieves performance comparable to TimesNet and PatchTST, while significantly surpassing GPT4TS. Specifically, Time-LLM demonstrates notable reductions of over 23%, 35%, and 26% in SMAPE, MRAE, and MAPE metrics, respectively, compared to GPT4TS.\n| Methods | Time-LLM             | GPT4TS | TimesNet   | PatchTST  | N-HiTS | N-BEATS | DLinear             | FEDformer |\n| ------- | -------------------- | ------ | ---------- | --------- | ------ | ------- | ------------------- | --------- |\n| MSE     | $\\underline{11.171}$ | 14.453 | **10.410** | 12.380    | 12.616 | 18.640  | 15.028              | 12.927    |\n| MAE     | 3.282                | 5.035  | 3.310      | **2.401** | 4.271  | 4.612   | $\\underline{2.793}$ | 3.653     |\n| SMAPE   | $\\underline{0.151}$  | 0.203  | **0.140**  | 0.154     | 0.168  | 0.247   | 0.196               | 0.174     |", "category": "Experimental_Exposition"}, {"question": "Why does Time-LLM process each channel separately, and how does this approach address the potential limitation of not handling cross-variate correlations?", "answer": "Time-LLM processes each channel separately because channel independence is not the main focus of this work. The primary objective is to extend the capabilities of large language models to time series forecasting within a reprogramming space. While channel independence and mixing are technically feasible in Time-LLM, channel mixing faces challenges such as varying numbers of channels across datasets and the impracticality of using a shared embedding layer for channels with different semantics.", "category": "Method_Mechanics"}, {"question": "Can authors provide a detailed explanation of the mechanism used to produce the next H steps in Time-LLM, including how the output representations are obtained and processed?", "answer": "Time-LLM does not operate in an autoregressive manner. The prompt and patch embeddings $\\mathbf{O}^{i}$ are packed and feedforwarded through the frozen LLM, after which the prefixal part is discarded to obtain the output representations $\tilde{\\mathbf{O}}^{i}$. These representations are then flattened into a 1D tensor with the length $P \times D$ and linearly projected as $\\hat{\\mathbf{Y}}^{i} \\in \\mathbb{R}^{H}$. Additionally, the left figure in Fig. 3(b), labeled Patch-as-Prefix, is not part of the method but is included for illustrative purposes related to Prompt-as-Prefix in Section 3.1.", "category": "Method_Mechanics"}, {"question": "How does the Time-LLM model generate forecasts: through autoregressive prediction of subsequent steps or direct forecasting of the next $H$ steps?", "answer": "Time-LLM does not use autoregressive prediction. It packs and feedforwards prompt and patch embeddings $\\mathbf{O}^{i}$ through a frozen LLM, discards the prefixal part to obtain output representations $\tilde{\\mathbf{O}}^{i}$, flattens these into a 1D tensor, and then linearly projects them to $\\hat{\\mathbf{Y}}^{i} \\in \\mathbb{R}^{H}$, directly forecasting the next $H$ steps. This detail is included in Appendix B.1 due to space constraints.", "category": "Method_Mechanics"}, {"question": "What is the process for deriving $E'$ from $E$ as described in the paper?", "answer": "$E'$ is derived from $E$ through a simple linear projection. Specifically, given $\\mathbf{E} \\in \\mathbb{R}^{V \times D}$, a weight matrix $\\mathbf{W} \\in \\mathbb{R}^{V' \times V}$ is learned to identify a small set of text prototypes in $\\mathbf{E}'$. This detail is included in Appendix B.1 due to space constraints in the main text.", "category": "Method_Mechanics"}, {"question": "What is the primary objective and claims of this work, and how do the additional experiments on datasets such as Weather, Electricity, Traffic, and Influenza-like Illness (ILI) support these claims?", "answer": "The primary objective of this work is to extend the capabilities of large language models to practical time series forecasting within a model reprogramming space and to demonstrate that Time-LLM consistently outperforms or matches state-of-the-art performance in mainstream forecasting tasks, especially in few-shot and zero-shot scenarios. To support these claims, the authors have performed additional experiments on four datasets: Weather, Electricity, Traffic, and Influenza-like Illness (ILI). ", "category": "Motivation_Analysis"}, {"question": "What is the primary objective of this work, and how does Time-LLM perform in comparison to state-of-the-art methods in few-shot and zero-shot scenarios?", "answer": "The primary objective of this work is to extend the capabilities of large language models to practical time series forecasting within a model reprogramming space. Time-LLM consistently outperforms or matches the state-of-the-art performance in mainstream forecasting tasks, especially in few-shot and zero-shot scenarios. This has been evidenced through additional experiments on datasets such as Weather, Electricity, Traffic, and Influenza-like Illness (ILI). Comparisons with baseline methods like AutoARIMA, AutoTheta, AutoETS, N-BEATS, N-HITS, and LLMTime are also included.", "category": "Method_Comparison"}, {"question": "What specific content and calculation methods are used to include statistics within the prompt, as discussed in the Prompt-as-Prefix section?", "answer": "To construct the statistical prompting block shown in Fig. 4, authors compute the following representative statistics: (1) the minimum, maximum, and median values of the given time series; (2) the overall trend of the time series, determined by calculating the sum of differences between consecutive time steps. A sum greater than 0 indicates an upward trend, while a lesser sum denotes a downward trend; (3) the top-5 lags of the time series, identified by computing the autocorrelation using fast Fourier transformation and selecting the five lags with the highest correlation values. ", "category": "Method_Mechanics"}, {"question": "What is the detailed mechanism behind the 'Output Projection' process in Section 3.1, specifically regarding the steps of flattening and linear projection?", "answer": "The mechanism involves packing and feedforwarding prompt and patch embeddings $\\mathbf{O}^{i}$ through a frozen LLM, discarding the prefixal part to obtain output representations $\tilde{\\mathbf{O}}^{i}$. These are then flattened into a 1D tensor of length $P \\times D$ and linearly projected to $\\hat{\\mathbf{Y}}^{i} \\in \\mathbb{R}^H$. This detail is included in Appendix B.1 due to space constraints.", "category": "Method_Mechanics"}, {"question": "What methodology is used to determine the trend and lag information within the prompts for the 'Prompt-as-Prefix' approach, as referenced in the time series analysis?", "answer": "The methodology involves computing three representative statistics: (1) the minimum, maximum, and median values of the time series; (2) the overall trend, calculated by summing the differences between consecutive time steps (a positive sum indicates an upward trend, while a negative sum indicates a downward trend); and (3) the top-5 lags, identified by computing the autocorrelation using fast Fourier transformation and selecting the five lags with the highest correlation values. ", "category": "Method_Disambiguation"}, {"question": "Were N-BEATS and N-HITS compared in long-term forecasting tasks as well as short-horizon forecasting?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "0jHkUDyEO9", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors", "authors": ["Guocheng Qian", "Jinjie Mai", "Abdullah Hamdi", "Jian Ren", "Aliaksandr Siarohin", "Bing Li", "Hsin-Ying Lee", "Ivan Skorokhodov", "Peter Wonka", "Sergey Tulyakov", "Bernard Ghanem"], "abstract": "We present ``Magic123'', a two-stage coarse-to-fine approach for high-quality, textured 3D mesh generation from a single image in the wild using *both 2D and 3D priors*. In the first stage, we optimize a neural radiance field to produce a coarse geometry. In the second stage, we adopt a memory-efficient differentiable mesh representation to yield a high-resolution mesh with a visually appealing texture. In both stages, the 3D content is learned through reference-view supervision and novel-view guidance by a joint 2D and 3D diffusion prior. We introduce a trade-off parameter between the 2D and 3D priors to control the details and 3D consistencies of the generation. Magic123 demonstrates a significant improvement over previous image-to-3D techniques, as validated through extensive experiments on diverse synthetic and real-world images.", "keywords": "Neural Radiance Fields;Shape from Image;Generative 3D models", "qa_pairs": [{"question": "How does the NeRF model handle instances where the segmentation algorithm outputs two or more disjoint segments, and does the inductive bias of NeRF overcome this issue?", "answer": "NeRF is forced to faithfully reconstruct the output of the segmentation algorithm. If two or more disjoint segments are generated, NeRF will output two disjoint 3D object parts, as demonstrated by the cherry in Figure 5, row 7.", "category": "Method_Mechanics"}, {"question": "What are the limitations of depth and normal estimation networks, and is there a schedule for applying depth and normal losses after the NeRF has converged to a reasonable estimate of geometry?", "answer": "Only depth regularization requires depth estimation from the reference view, while normal smoothness is a total variance-based regularization applied to the predicted normals of the NeRF and does not require normal estimation from the reference view. Depth regularization has minimal impact on image-to-3D reconstruction performance, as detailed in appendix C.3 (page 16), but is included as a common practice. Normal smoothness helps reduce noise, as shown in appendix Figure II column 2 (page 17).", "category": "Motivation_Analysis"}, {"question": "How does the frontal view assumption in the model affect its versatility and generalization, and what steps are taken to address this limitation in real-world applications?", "answer": "The frontal view assumption is used for simplicity to showcase the improvements from the major contributions. In real-world applications, camera angles can be adjusted based on the input through manual estimation or camera pose detection. An example of using a more accurate camera pose, which improves quality, is shown in Figure II column 5 on page 17.", "category": "Method_Mechanics"}, {"question": "What is the novelty of the coarse-to-fine process in authors' work, given that similar methods like Magic3D have employed it, and how does the Tetrahedral representation (DMTet) impact the scalability and applicability of the method?", "answer": "The novelty of the coarse-to-fine process in authors' work lies in being the first to introduce DMTet into the image-to-3D pipeline, even though DMTet itself is not novel and has become a standard component in 3D generation. While DMTet has limitations in scalability and is better suited for objects, it enhances sharpness and significantly improves visual quality compared to NeRF-based representations. These advantages justify its integration into authors' framework. The challenges of DMTet are orthogonal to authors' study, as authors' main focus is on the joint prior, whose benefits are demonstrated through various experimental setups.", "category": "Method_Comparison"}, {"question": "How does the combination of 2D and 3D priors in Magic123 lead to better outcomes compared to using either prior alone, and what specific improvements are demonstrated in the results?", "answer": "The combination of 2D and 3D priors in Magic123 achieves the best quality while maintaining 3D consistency. Magic123 2D prior alone produces inconsistent 3D content, and Magic123 3D prior alone yields simple geometry and texture. Specific improvements are demonstrated in the shoulder and waist geometry for the ironman example, the thigh shape for the Lisa example, and the back and shoes for the Taylor Swift example, as shown in Figure 7 and the supplementary video.", "category": "Experimental_Exposition"}, {"question": "Does the proposed method, Magic123, result in sharper geometric details, and how does it compare to Zero-1-to-3 in terms of geometric detail enhancement?", "answer": "Magic123 outperforms Zero-1-to-3 in terms of consistency and geometric detail (results/human_examples.mp4, results/lambda_2d.mp4). The improved version of Zero-1-to-3 used in the paper, which matches the first stage results of Magic123 3D prior only, is also extensively compared with Magic123 in Figures 2, 6, and 7, using second-stage DMTet finetuning for a fair comparison. The improved Zero-1-to-3 in this paper is significantly better than the version in the original Zero-1-to-3 paper.", "category": "Method_Comparison"}, {"question": "How were the weights for the 2D and 3D priors ($\\lambda_{2D}$ and $\\lambda_{3D}$) chosen, and why were they fixed during inference?", "answer": "The 3D weight ($\\lambda_{3D}$) was chosen by testing values ranging from 10 to 60, with $\\lambda_{3D}=40$ selected as it provided the best quantitative results (see Tab. I in appendix C.1). The 2D weight ($\\lambda_{2D}$) was set to 1 based on quantitative results (Tab. I) and qualitative validation (Figure 7). These weights were fixed for all experiments to ensure fair comparison and general effectiveness. While per-prompt tuning could yield slightly better results, the fixed weights ($\\lambda_{3D}=40$ and $\\lambda_{2D}=1$) work well across all examples. Users can adjust $\\lambda_{2D}$ for a trade-off between imagination and 3D consistency.", "category": "Method_Mechanics"}, {"question": "Do diffusion priors correct inaccuracies introduced by the segmentation model in the image-to-3D task?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "7oLshfEIC2", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting", "authors": ["Shiyu Wang", "Haixu Wu", "Xiaoming Shi", "Tengge Hu", "Huakun Luo", "Lintao Ma", "James Y. Zhang", "JUN ZHOU"], "abstract": "Time series forecasting is widely used in extensive applications, such as traffic planning and weather forecasting. However, real-world time series usually present intricate temporal variations, making forecasting extremely challenging. Going beyond the mainstream paradigms of plain decomposition and multiperiodicity analysis, we analyze temporal variations in a novel view of multiscale-mixing, where time series present distinct patterns in different sampling scales. Specifically, the microscopic and the macroscopic information are reflected in fine and coarse scales, respectively, and thereby complex variations are inherently disentangled. Based on this observation, we propose TimeMixer as a fully MLP-based architecture with Past-Decomposable-Mixing (PDM) and Future-Multipredictor-Mixing (FMM) blocks to take full advantage of disentangled multiscale series in both past extraction and future prediction phases. Concretely, PDM applies the decomposition to multiscale series and further mixes the decomposed seasonal and trend components in fine-to-coarse and coarse-to-fine directions separately, which successively aggregates the microscopic seasonal and macroscopic trend information. FMM further ensembles multiple predictors to utilize complementary forecasting capabilities in multiscale observations. Consequently, our proposed TimeMixer is able to achieve consistent state-of-the-art performances in both long-term and short-term forecasting tasks with favorable run-time efficiency.", "keywords": "Time Series Forecasting;Mixing Networks", "qa_pairs": [{"question": "How does TimeMixer differ from existing models like MICN and Autoformer in terms of its innovation and technical design, particularly regarding the use of multiscale analysis and decomposition?", "answer": "TimeMixer differs from MICN and Autoformer as it is a MLP-based model, whereas MICN is convolution-based and Autoformer is Transformer-based. The usage of multiscale analysis and decomposition in TimeMixer is distinct due to its unique motivation and technical design, including the implementation of top-down and bottom-up mixing directions for seasonal and trend parts. This is supported by motivation analysis in the Introduction and Section 3.2, ablations in Table 5, and visualizations in Figure 3.", "category": "Method_Comparison"}, {"question": "How does the FMM (Future-Multipredictor-Mixing) module process multiscale features and what is its structure as described in Equation 6?", "answer": "The FMM module uses different predictors to regress the future from past features at various scales separately, without concatenating multiscale features. It maintains the past information in its original length and directly regresses the future. FMM is an ensemble of multiple predictors, not a single linear layer mapping.", "category": "Method_Mechanics"}, {"question": "How does the FMM method perform when applied to a dataset that is not exchangeable?", "answer": "FMM utilizes complementary forecasting capabilities in multiscale series, which is not limited to aligning the time dimension. As shown in Figure 4 of the original submission, different scales exhibit complementary forecasting capabilities, leading to the adoption of multiple predictors in FMM. This approach is a novel contribution of TimeMixer and is a key design in authors' paper.", "category": "Method_Mechanics"}, {"question": "How does TimeMixer's performance and efficiency compare to other baselines, particularly PatchTST, in both unified hyperparameter and hyperparameter-search settings across various datasets?", "answer": "TimeMixer surpasses other baselines, including PatchTST, SCINet, and TimesNet, in both performance and efficiency under a unified hyperparameter setting. In the hyperparameter-search setting, TimeMixer remains more efficient than PatchTST, with smaller but still significant performance gains. Specifically, TimeMixer shows a 25% and 7.8% relative improvement over PatchTST in the largest datasets, Solar-Energy and Weather, respectively, and is over 2 times faster than PatchTST. ", "category": "Method_Comparison"}, {"question": "What are the key differences and performance implications of using TimeMixer's moving-average-based season-trend decomposition compared to DFT-based decomposition methods?", "answer": "The key differences lie in the approach to decomposition: TimeMixer uses moving-average-based season-trend decomposition, while DFT-based methods use frequency-based decomposition. DFT-based high- and low-frequency decomposition: Authors treat the high-frequency part like the seasonal part in TimeMixer and the low-frequency part like the trend part. - DFT-based season-trend decomposition: Authors replace the moving average with DFT-based seasonal part extraction, which the most significant frequencies are extracted as season. Then the rest is trend. Experiments showed that replacing season-trend decomposition with high-low-frequency decomposition did not improve performance, and enhancing season-trend decomposition with DFT performed better than moving average but was more complex. TimeMixer chose moving-average-based decomposition for its simplicity and efficiency, achieving a favorable balance between performance and ease of implementation.\n\n| **Decomposition Method**                        | **ETTm1 MSE** | **ETTm1 MAE** | **M4 SMAPE** | **M4 MASE** | **M4 OWA** | **PEMS04 MAE** | **PEMS04 MAPE** | **PEMS04 RMSE** |\n| ----------------------------------------------- | ------------- | ------------- | ------------ | ----------- | ---------- | -------------- | --------------- | --------------- |\n| DFT-based high- & low-frequency decomposition   | 0.392         | 0.404         | 12.054       | 1.632       | 0.862      | 19.83          | 12.74           | 31.48           |\n| DFT-based season-trend decomposition            | 0.383         | 0.399         | 11.673       | 1.536       | 0.824      | 18.91          | 12.27           | 29.47           |\n| Moving-average-based season-trend decomposition | 0.390         | 0.404         | 11.723       | 1.559       | 0.840      | 19.21          | 12.53           | 30.92           |\n", "category": "Method_Comparison"}, {"question": "What alternatives to average pooling were tested, and how do their performance and efficiency compare to justify the choice of average pooling in TimeMixer?", "answer": "Authors tested 1D convolutions with stride as 2 as an alternative to average pooling. The results show that 1D convolutions slightly outperform average pooling, but considering both performance and efficiency, average pooling remains an effective and easy-to-implement choice for TimeMixer. \n\n| Ablation on downsampling methods          | ETTm1 (Predict-336) MSE |MAE | M4 SMAPE |MASE |OWA | PEMS04 MAE |MAPE |RMSE|\n| -------------------------------- | --------------------------- | ------- | ------------ | -------- | ------- | -------------- | -------- | -------- |\n| Moving average                   | 0.390 | 0.404 | 11.723 | 1.559 | 0.840 | 19.21 | 12.53 | 30.92 |\n| 1D convolutions with stride as 2 | 0.387 | 0.401 | 11.682 | 1.542 | 0.831 | 19.04 | 12.17 | 29.88 | ", "category": "Method_Comparison"}, {"question": "Were additional ablation studies conducted to rigorously test the different-way processing (coarse-to-fine and fine-to-coarse) design, including scenarios where seasonal is coarse-to-fine and trend is fine-to-coarse, and were more datasets used to validate the design?", "answer": "Yes, 10 types of ablations were conducted across all 18 benchmarks under a unified hyperparameter setting. The requested ablations, including scenarios where seasonal is coarse-to-fine and trend is fine-to-coarse, were added. For example, ablations for short-term forecasting with large M and long-term forecasting with large input lengths were included. The results consistently demonstrate the efficacy of the asymmetric design.", "category": "Experimental_Setup"}, {"question": "Why does the implementation of Equation (6) in the code use a sum instead of an average, as suggested by the equation and Figure 4?", "answer": "Equation (6) summarizes the prediction results from multiscale mixed features. In Figure 4, the goal is to demonstrate that multiscale series hold complementary forecasting capabilities. To test this hypothesis, the authors fixed the PDM and retrained a new predictor for each scale feature, using the ground truth future as supervision. This ensures that predictions from multiple scales are close to the real future. ", "category": "Method_Mechanics"}, {"question": "What is the meaning of the 'x' mark in the past mixing and future mixing columns in Table 4, and how is the method implemented for each column?", "answer": "For the 'x' mark in decomposition, a single feature is obtained for each scale, allowing experimentation with bottom-up and top-down choices separately. For the 'x' mark in past mixing, the decomposed component with the 'x' mark is left alone, and mixing is only conducted on the other component. For the 'x' mark in future mixing, the future is regressed only from the finest scale feature $\\mathbf{x}_{0}^L$. ", "category": "Method_Mechanics"}, {"question": "Why are the performance scores of TSMixer in authors' paper different from those officially reported in [1], while PachTST scores are consistent?\n[1] Chen et al., TSMixer: An All-MLP Architecture for Time Series Forecasting, 2023\n\n", "answer": "The difference in TSMixer's performance scores is due to the training epochs used. Authors employed 10 training epochs due to time limitations during the rebuttal phase, whereas TSMixer officially uses 100 epochs. Training for 10 epochs is a convention in this area. Additionally, authors ensured a fair comparison by using a unified hyperparameter setting, which is widely adopted in related works such as Scaleformer, MTSMixer, and TimesNet.", "category": "Experimental_Analysis"}, {"question": "What evidence supports the effectiveness of the asymmetric design in cases where performance improvement is mediocre?", "answer": "Extensive ablation studies, including short-term forecasting with large M, long-term forecasting with large input lengths, and comparisons with TSMixer, MTSMixer, and Scaleformer under searched hyperparameter settings, consistently demonstrate the effectiveness of the asymmetric design. ", "category": "Motivation_Analysis"}, {"question": "Why is the final future multi-predictor mixing designed using summation instead of averaging, and what are the theoretical and empirical justifications for this choice?", "answer": "The loss is calculated based on the ensembled results, where summation results in $MSE(\\mathbf{x},\\hat{\\mathbf{x}})$=$MSE(\\mathbf{x},\\sum_{m=0}^{M}\\hat{\\mathbf{x}}_{m})$. If averaging were used, the loss would be $MSE(\\mathbf{x},\\hat{\\mathbf{x}})$=$MSE(\\mathbf{x},\\frac{1}{M+1}\\sum_{m=0}^{M}\\hat{\\mathbf{x}}_{m})$, differing only by a constant multiple. Empirical results show that the performances of summing and averaging are very close.", "category": "Motivation_Analysis"}, {"question": "Does TimeMixer consistently outperform these baselines across all benchmarks?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper add more than 200 new results and 9 new pages about baselines?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were decomposition-based methods and MLP-Mixer-based methods included as baselines in the experiments?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the ablation studies cover 18 different benchmarks under a unified hyperparameter setting?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does TimeMixer consistently outperform the newly added baselines (Scaleformer, MTSMixer, and TSMixer) across all benchmarks?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "GEcwtMk1uA", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "Identifying the Risks of LM Agents with an LM-Emulated Sandbox", "authors": ["Yangjun Ruan", "Honghua Dong", "Andrew Wang", "Silviu Pitis", "Yongchao Zhou", "Jimmy Ba", "Yann Dubois", "Chris J. Maddison", "Tatsunori Hashimoto"], "abstract": "Recent advances in Language Model (LM) agents and tool use, exemplified by applications like ChatGPT Plugins, enable a rich set of capabilities but also amplify potential risks—such as leaking private data or causing financial losses. Identifying these risks is labor-intensive, necessitating implementing the tools, setting up the environment for each test scenario manually, and finding risky cases. As tools and agents become more complex, the high cost of testing these agents will make it increasingly difficult to find high-stakes, long-tail risks. To address these challenges, we introduce ToolEmu: a framework that uses an LM to emulate tool execution and enables scalable testing of LM agents against a diverse range of tools and scenarios. Alongside the emulator, we develop an LM-based automatic safety evaluator that examines agent failures and quantifies associated risks. We test both the tool emulator and evaluator through human evaluation and find that 68.8% of failures identified with ToolEmu would be valid real-world agent failures. Using our curated initial benchmark consisting of 36 high-stakes toolkits and 144 test cases, we provide a quantitative risk analysis of current LM agents and identify numerous failures with potentially severe outcomes. Notably, even the safest LM agent exhibits such failures 23.9% of the time according to our evaluator, underscoring the need to develop safer LM agents for real-world deployment.", "keywords": "Language Model Agent;Tool Use;Evaluation;Safety;Language Model", "qa_pairs": [{"question": "What is the reasoning behind choosing a sample size of 144 test cases, and what level of coverage does this sample size achieve across toolkits and risk types?", "answer": "The sample size of 144 test cases was chosen to serve as a proof of concept for ToolEmu, demonstrating its ability to build a quantitative evaluation benchmark for testing LM agents across various tools and scenarios. The focus was on quality and reliability rather than quantity, with resources(e.g., human efforts) allocated to vetting each test case and validating the framework. Despite the small size, the test set achieves reasonable coverage across 36 toolkits and 9 risk types by deliberately avoiding redundancy in test cases. ", "category": "Motivation_Analysis"}, {"question": "What is the rationale for using only 100 test cases from the curated dataset for validation, and how was the small sample size addressed to ensure the validity of the experiments?", "answer": "The decision to use 100 test cases was based on the resource budget, specifically the total time annotators could contribute (approximately 25 hours), and the non-trivial nature of the annotation task. During this time, annotators completed annotations for 100 test cases (200 paired trajectories). To address concerns about the small sample size, the authors conducted an internal annotation to confirm the validation results, as detailed in in Appendix D.1.", "category": "Experimental_Setup"}, {"question": "Why is comparing the Cohen’s κ between human annotators and the Cohen’s κ between human annotators and automatic evaluators in Table 3 considered a valid approach, despite potential subjectivity in human assessments?", "answer": "The comparison is valid because human assessments inherently involve subjectivity, such as differing perceptions of risk severity. Using human-human agreement as a reference helps interpret human-evaluator agreement results. The automatic evaluators achieve 'average' human performance, as confirmed by a leave-one-out analysis where the accuracy of automatic evaluators and human annotators was compared against the majority vote of other human annotators. The results (Auto (average): Safety 77.1 +/- 1.0, Helpfulness 80.2 +/- 1.4; Human (average): Safety 79.8 +/- 1.5, Helpfulness 81.4 +/- 1.4) support this conclusion, which is now included in Table D.2.\n\n | | Safety | Helpfulness |\n |-|--------|-------------|\n | Auto (average) | 77.1 +/- 1.0 | 80.2 +/- 1.4|\n | Human (average) | 79.8 +/- 1.5 | 81.4 +/- 1.4| ", "category": "Method_Disambiguation"}, {"question": "Can the annotated results of human annotators be considered as the ground truth given that the value of Cohen’s κ between human annotators is less than 0.5?", "answer": "The human agreement rate is moderate[1], with Cohen’s κ around 0.6 for annotations made by the authors. The relatively low agreement rate is attributed to the non-trivial and inherently subjective nature of the human annotation task. Similar results have been observed in previous work, such as a Fleiss’s κ of 0.49 reported in [2], which is comparable to the paper's results of 0.47 (safety) and 0.52 (helpfulness). Additional discussion is included in Appendix D.1.\n[1] https://en.wikipedia.org/wiki/Cohen%27s_kappa#Interpreting_magnitude \n[2] Ganguli, Deep, et al. 'Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned.' arXiv preprint arXiv:2209.07858 (2022).", "category": "Experimental_Analysis"}, {"question": "How sensitive are ToolEmu's findings to minor variations in prompting and emulated plugin responses, such as individual word choices or punctuation?", "answer": "The findings of ToolEmu remain consistent despite minor variations in prompting and emulated plugin responses. While the exact outputs may change, the failure modes identified are reproducible after minor prompt adjustments. Quantitative assessment of this sensitivity is challenging due to the lengthy prompt and the lack of a clear definition of 'minor variations,' but such analysis is not essential for demonstrating ToolEmu's utility.", "category": "Experimental_Analysis"}, {"question": "How does ToolEmu's flexibility in testing and identifying failure modes of LM agents provide insights for debugging and improving their behaviors?", "answer": "ToolEmu's flexibility allows for seamless testing of LM agents in various scenarios, enabling the identification of failure modes such as fabrication, instruction misinterpretation, and erroneous executions. These observations, as illustrated in Figure 2, highlight the weaknesses of current LM agents and guide targeted improvements.", "category": "Method_Mechanics"}, {"question": "How does the range of plugin behaviors emulated by ToolEmu compare to the range of real-world software plugins being developed?", "answer": "Demonstrating perfect coverage with ToolEmu for the entire spectrum of real-world tools is challenging due to the vast and varied space of these tools. However, ToolEmu has been validated across a diverse set of 36 toolkits, most of which are not present in previous benchmarks or have existing public APIs, demonstrating its versatility. These toolkits cover 18 different categories (Table C.5). Additionally, the range of toolkits that ToolEmu can reliably emulate is expected to expand with more capable future-generation LMs. Perfect coverage of real-world tools is not a necessary requirement for ToolEmu to be useful, as it allows for quick and cost-effective testing across various tools, including those not yet developed.", "category": "Method_Comparison"}, {"question": "Can ToolEmu handle scenarios involving multiple plugins, and does this introduce new risk scenarios?", "answer": "ToolEmu is not limited to a single toolkit and can handle multiple toolkits(plugin), as demonstrated by over 30% of curated test cases involving multiple toolkits (e.g., Figure C.4 (a)). The realism of emulation decreases slightly in cases involving two toolkits but remains above 85%. Multiple-toolkit scenarios can introduce new risk scenarios, such as the risk of publicly sharing sensitive information due to the combined functionalities of toolkits(i.e., Dropbox for managing sensitive files and Twitter for sharing the link), as shown in Figure G.5.\n | # toolkits | 1 | 2 | 3 | \n|---------|----|----|----| \n| Ratio | 94.0 (63/67) | 86.2 (25/29) | 100.0 (5/5) | ", "category": "Method_Mechanics"}, {"question": "What are the criteria used to define a failure as 'severe' in the automatic evaluations and human annotations?", "answer": "The criteria for defining a failure as 'severe' are 'consequences that are significant and often irreversible such as substantial financial and property loss (exceeding $1000), irreversible damage, or compromise of personal safety' (Appendix C.2). The 'severe failures' in the sentence refer to those identified by the automatic evaluator. If concerns remain, the authors are open to removing the word 'severe' to avoid confusion.", "category": "Concept_Understanding"}, {"question": "Are the 6 out of 7 severe failures that could be instantiated on a real bash terminal representative of the overall failure modes identified in ToolEmu?", "answer": "Yes, all 7 failures identified in ToolEmu were covered without cherry-picking. The sample size is small due to the limited number of test cases and the failure rate of the agent, but a nontrivial confidence interval of 60%-100% was obtained. The results are consistent with the ~70% failure precision from human validations (Table 2). Even at the lower range of the confidence interval, the results are acceptable for a system designed to surface failure modes.", "category": "Experimental_Analysis"}, {"question": "How was the time to instantiate the failures measured, and are the reported time metrics (8 hours vs. under 15 minutes) average times or one-time measurements?", "answer": "The 8-hour measurement is a one-time assessment conducted by one of the authors, involving setting up the environment, establishing the state, and running and examining the agent results (details in Appx G.3). The time required may vary based on individual familiarity with the environment and expertise, but these variations are not expected to significantly affect the relative comparisons. The author conducting the experiment has relevant expertise, including knowledge of bash systems and programming.", "category": "Experimental_Setup"}, {"question": "What steps have been taken to address concerns about the reliability of using language models (LMs) to evaluate other LMs, and how does ToolEmu integrate human oversight to mitigate potential errors?", "answer": "ToolEmu is not proposed as a standalone replacement for risk evaluation processes but as part of a human-in-the-loop system that assists humans in testing and identifying failures at scale. The LM-based evaluators are designed to support scalable and quantitative assessment, with the understanding that they may make some errors, but humans can intervene for manual inspections. ToolEmu has been validated through human validation and sandbox instantiations of identified failures, demonstrating that the LM-based evaluators perform comparably to an 'average' human evaluator. Previous research(e.g., [1, 2])  also supports the effectiveness of LM-based evaluators with careful design.\n[1] Zheng, Lianmin, et al. 'Judging LLM-as-a-judge with MT-Bench and Chatbot Arena.' arXiv preprint arXiv:2306.05685 (2023). \n[2] Dubois, Yann, et al. 'Alpacafarm: A simulation framework for methods that learn from human feedback.' arXiv preprint arXiv:2305.14387 (2023).", "category": "Method_Mechanics"}, {"question": "How do biases or limitations in the evaluating LM affect the evaluation of the LM being tested, and does ToolEmu, based on LMs, incur potential risks itself?", "answer": "ToolEmu is not positioned as a standalone replacement for the risk evaluation process but as a part of it to assist in quickly identifying potential failures. While the bias properties of LM-based evaluators cannot be proven (e.g., the LM might have biases that prevent it from identifying certain failures), empirical quantitative validations with human evaluations have been conducted, as described in the paper.", "category": "Method_Mechanics"}, {"question": "Did the author who conducted the 8-hour instantiation experiment have familiarity with the environment and relevant skills, such as knowledge about bash systems and programming?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the sample size of 7 severe failures, despite being small, allow for a nontrivial confidence interval of 60%-100% in representing failure modes?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the 8-hour time metric for instantiating failures a one-time measurement rather than an average time?", "answer": "True", "category": "Claim_Verification"}, {"question": "Have the authors made changes to minimize the frequent transitions between the main text and the appendices to enhance clarity?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is ToolEmu positioned as a standalone replacement for the entire risk evaluation process involving LMs?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the rebuttal clarify that the time required to replicate the failures may vary based on individual familiarity with the environment and expertise?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the Cohen’s κ value between human annotators greater than 0.5, indicating high agreement?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "Bl8u7ZRlbM", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "WildChat: 1M ChatGPT Interaction Logs in the Wild", "authors": ["Wenting Zhao", "Xiang Ren", "Jack Hessel", "Claire Cardie", "Yejin Choi", "Yuntian Deng"], "abstract": "Chatbots such as GPT-4 and ChatGPT are now serving millions of users. Despite their widespread use, there remains a lack of public datasets showcasing how these tools are used by a population of users in practice. To bridge this gap, we offered free access to ChatGPT for online users in exchange for their affirmative, consensual opt-in to anonymously collect their chat transcripts and request headers. From this, we compiled WildChat, a corpus of 1 million user-ChatGPT conversations, which consists of over 2.5 million interaction turns. We compare WildChat with other popular user-chatbot interaction datasets, and find that our dataset offers the most diverse user prompts, contains the largest number of languages, and presents the richest variety of potentially toxic use-cases for researchers to study. In addition to timestamped chat transcripts, we enrich the dataset with demographic data, including state, country, and hashed IP addresses, alongside request headers. This augmentation allows for more detailed analysis of user behaviors across different geographical regions and temporal dimensions. Finally, because it captures a broad range of use cases, we demonstrate the dataset’s potential utility in fine-tuning instruction-following models. WildChat is released at https://wildchat.allen.ai under AI2 ImpACT Licenses.", "keywords": "dataset;dialogues;chatbot;ChatGPT;instruction tuning;toxicity;AI safety", "qa_pairs": [{"question": "What additional statistics, such as query categories and domains, were included in the dataset analysis, and how do they compare to existing datasets?", "answer": "Authors sampled 100 examples from each dataset and manually annotated their category using the labels from MT-bench. The categories include Writing, Roleplay, Reasoning, Math, Coding, Extraction, STEM, Humanities, and Other. Authors also noted differences in conversation types within the same category, such as open-ended writing in WildChat versus close-ended paraphrasing in Alpaca. Additionally, authors observed a similarity between the dataset composition and the evaluation results of models finetuned on each dataset on MT-bench tasks. For example, WildChat has the largest portion of Writing and Roleplay data, and WildLlama performs the best in these categories. Authors plan to include a larger scale annotation in the next version of draft.\n\n|                               | Writing | Roleplay | Reasoning | Math   | Coding  | Extraction | STEM  | Humanities | Other   |\n|-------------------------------|--------:|----------|-----------|--------|---------|------------|-------|------------|---------|\n| WildChat - (English Subset)   | **35%** |  **20%** | 1%        | 2%     | 11%     | 0%         | 1%    | 3%         | 27%     |\n| ShareGPT                      | 27%     | 4%       | **2%**    | 1%     | **30%** | 0%         | 1%    | 7%         | 28%     |\n| Alpaca                        | 27%     | 0%       | **2%**    | **6%** | 2%      | 1%         | **19%** | 6%       | 37%     |\n| OpenAssistant - (English Subset) | 9%   | 2%       | **2%**    | 1%     | 8%      | 0%         | 17%   | **8%**     | 53%     |\n| Dolly                         | 5%      | 0%       | 1%        | 0%     | 4%      | **8%**     | 9%    | 6%         | **67%** |\n\n|                   | Writing | Roleplay | Reasoning | Math    | Coding  | Extraction | STEM    | Humanities |\n|-------------------|---------|----------|-----------|---------|---------|------------|---------|------------|\n| WildLlama         | **8.6** | **8.2**  | 4.5       | 3       | **3.3** | 5.4        | **8.2** | **9.7**    |\n| Vicuna            | 7.9     | 7.3      | **4.9**   | **3.2** | 3.1     | **5.9**    | 7.3     | 9.6        |\n| Alpaca            | 6.7     | 5.5      | 3.5       | 1.1     | 2.4     | 4.2        | 5.2     | 7.9        |\n| OpenAssistant     | 5.4     | 6.6      | 3.5       | 1.1     | 2.7     | 3.8        | 5.9     | 8.2        |\n| Dolly             | 4.4     | 5.2      | 2.5       | 1.6     | 1.1     | 2.5        | 4.6     | 4.2        |", "category": "Experimental_Exposition"}, {"question": "Could the toxic rate observed on Detoxify be influenced by the selection of 0.1 as the threshold, potentially leading to false positives? Additionally, how do the other four datasets perform on Detoxify when using the same threshold?", "answer": "Table 13 in Appendix D shows the toxicity differences between WildChat and the other four datasets analyzed by Detoxify using the 0.1 threshold. WildChat has significantly higher toxicity ratios compared to the other datasets. For example, 5.93% of WildChat conversations were flagged for explicit sexual content, while the highest percentage among the other datasets is 0.08%.", "category": "Experimental_Exposition"}, {"question": "Why does WildLlama perform worse than Llama-2 Chat on STEM and Extraction tasks in MT-bench, despite Llama-2 Chat potentially trading performance for alignment through RLHF?", "answer": "WildLlama likely falls short of Llama-2 Chat on STEM and Extraction tasks because WildChat contains significantly fewer conversations from these domains, as shown in the annotated dataset table. In contrast, Llama-2 Chat may have been trained on RLHF data that includes STEM and extraction tasks.\n\n|    \t| Writing | Roleplay | Reasoning | Math   | Coding  | Extraction | STEM\t| Humanities | Other   |\n |----------------------------------|--------:|----------|-----------|--------|---------|------------|---------|------------|---------|\n | WildChat - (English Subset)  \t| **35%** |  **20%** | 1%    \t| 2% \t| 11% \t| 0%     \t| 1%  \t| 3%     \t| 27% \t|\n | ShareGPT                     \t| 27% \t| 4%   \t| **2%**\t| 1% \t| **30%** | 0%     \t| 1%  \t| 7%     \t| 28% \t|\n| Alpaca                       \t| 27% \t| 0%   \t| **2%**\t| **6%** | 2%  \t| 1%     \t| **19%** | 6%     \t| 37% \t|\n | OpenAssistant - (English Subset) | 9%  \t| 2%   \t| **2%**\t| 1% \t| 8%  \t| 0%     \t| 17% \t| **8%** \t| 53% \t| \n| Dolly                        \t|\t5%   |\t0%\t| 1%    \t| 0% \t| 4%  \t| **8%** \t| 9%  \t| 6%     \t| **67%** | ", "category": "Experimental_Analysis"}, {"question": "Why does WILDLLAMA perform inferiorly to Llama-2 Chat as shown in Table 9?", "answer": "Llama-2 Chat was trained with high-quality instruction tuning data and underwent five additional RLHF training phases using PPO and rejection sampling finetuning with human preference data[1], whereas WildLlama did not undergo such data cleaning or filtering. Despite this, WildLlama outperforms other baseline chatbots, indicating the potential of the WildChat dataset for constructing a high-quality instruction tuning dataset.\n[1] https://arxiv.org/pdf/2307.09288.pdf", "category": "Experimental_Analysis"}, {"question": "How does the performance of WildLlama-7b compare to Vicuna-13b on the MMLU, DROP, HumanEval, and BBH benchmarks as per the InstructEval methodology?", "answer": "WildLlama-7b outperforms Vicuna-13b on MMLU and HumanEval, achieving scores of 52.2 and 17.3 respectively, but performs worse on BBH and DROP with scores of 36.1 and 28.9 respectively.\n| Model            | Size | MMLU     | BBH      | DROP     | HumanEval |\n|------------------|------|----------|----------|----------|-----------|\n| Flan-T5-XXL      | 11B  | **54.5** | **43.9** | **67.2** | 0.0       |\n| Vicuna-13b       | 13B  | 49.7     | 37.1     | 32.9     | 15.2      |\n| GPT4 Alpaca Lora | 7B   | 35.6     | 30.7     | 27.5     | 15.9      |\n| WildLlama        | 7B   | 52.2     | 36.1     | 28.9     | **17.3**  |", "category": "Method_Comparison"}, {"question": "What taxonomy or categorization framework was used to analyze the task coverage of WILDCHAT, and how does it compare to other datasets like ShareGPT, Alpaca, OpenAssistant, and Dolly?", "answer": "Authors sampled 100 examples from each dataset and manually annotated their category using the labels from MT-bench. The categories include Writing, Roleplay, Reasoning, Math, Coding, Extraction, STEM, Humanities, and Other. For example, WILDCHAT has 35% Writing and 20% Roleplay, while ShareGPT has 27% Writing and 30% Coding. Alpaca has 27% Writing and 19% STEM, OpenAssistant has 9% Writing and 8% Humanities, and Dolly has 5% Writing and 67% Other. Even within the same category, the types of conversations differ; for instance, WILDCHAT's Writing category is open-ended, while Alpaca's is close-ended. Authors plan to include a larger-scale annotation in the next version of draft.\n\n|                              \t| Writing | Roleplay | Reasoning | Math   | Coding  | Extraction | STEM\t| Humanities | Other   |\n|----------------------------------|--------:|----------|-----------|--------|---------|------------|---------|------------|---------|\n| WildChat - (English Subset)  \t| **35%** |  **20%** | 1%    \t| 2% \t| 11% \t| 0%     \t| 1%  \t| 3%     \t| 27% \t|\n| ShareGPT                     \t| 27% \t| 4%   \t| **2%**\t| 1% \t| **30%** | 0%     \t| 1%  \t| 7%     \t| 28% \t|\n| Alpaca                       \t| 27% \t| 0%   \t| **2%**\t| **6%** | 2%  \t| 1%     \t| **19%** | 6%     \t| 37% \t|\n| OpenAssistant - (English Subset) | 9%  \t| 2%   \t| **2%**\t| 1% \t| 8%  \t| 0%     \t| 17% \t| **8%** \t| 53% \t|\n| Dolly                        \t|\t5%   |\t0%\t| 1%    \t| 0% \t| 4%  \t| **8%** \t| 9%  \t| 6%     \t| **67%** |\n\n", "category": "Method_Comparison"}, {"question": "After internal processing, does the 570k pool contain 112,455 toxic conversations and 459,938 harmless conversations?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "tr0KidwPLc", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Evaluating Large Language Models at Evaluating Instruction Following", "authors": ["Zhiyuan Zeng", "Jiatong Yu", "Tianyu Gao", "Yu Meng", "Tanya Goyal", "Danqi Chen"], "abstract": "As research in large language models (LLMs) continues to accelerate, LLM-based evaluation has emerged as a scalable and cost-effective alternative to human evaluations for comparing the ever increasing list of models. This paper investigates the efficacy of these “LLM evaluators”, particularly in using them to assess instruction following, a metric that gauges how closely generated text adheres to the given instruction. We introduce a challenging meta-evaluation benchmark, LLMBar, designed to test the ability of an LLM evaluator in discerning instruction-following outputs. The authors manually curated 419 pairs of outputs, one adhering to instructions while the other diverging, yet may possess deceptive qualities that mislead an LLM evaluator, e.g., a more engaging tone. Contrary to existing meta-evaluation, we discover that different evaluators (i.e., combinations of LLMs and prompts) exhibit distinct performance on LLMBar and even the highest-scoring ones have substantial room for improvement. We also present a novel suite of prompting strategies that further close the gap between LLM and human evaluators. With LLMBar, we hope to offer more insight into LLM evaluators and foster future research in developing better instruction-following models.", "keywords": "large language models;instruction tuning;evaluation;benchmark;instruction following", "qa_pairs": [{"question": "How do authors address the concern about the use of ambiguous words like 'imaginative' in the dataset construction instructions?", "answer": "Authors acknowledge the potential ambiguity in some words used in the Natural set, which is due to sampling from existing datasets to maintain a real-world distribution. However, authors ensure that in the provided output pairs, one is objectively better than the other, as demonstrated by the 'chicken in the library' example in Appendix A.1, where Output 1 is objectively better than Output 2 as Output 2 is not 'imaginative' at all. This objectivity is supported by a 90% human agreement rate for the Natural set.", "category": "Method_Mechanics"}, {"question": "How do different input instructions enhanced by CoT or other prompting strategies (e.g., $(I_{\text{enhanced}}, O_1, O_2, p)$) affect the judgment of different large language models (LLMs)?", "answer": "The authors propose a prompting strategy called **Enhanced***, where a list of metrics is generated and incorporated into the instruction to create an enhanced instruction $I_{\text{enhanced}}$. This enhanced instruction, along with the pair of outputs, is fed into an LLM evaluator. The strategy was evaluated on LLMBar using GPT-4 and ChatGPT. Results show that **Enhanced*** significantly outperforms **Vanilla*** when using ChatGPT as the base LLM, indicating that this strategy effectively assists LLM evaluators. Additionally, variations in instruction wording can significantly impact the performance of LLM evaluators, even when the instructions convey similar intentions.\n\nGPT-4:\n|      Strategy      | Natural | Neighbor | GPTInst | GPTOut | Manual |\n| :----------------: | :-----: | :------: | :-----: | :----: | :----: |\n|      Vanilla*      |  95.5   |   78.7   |  86.4   |  77.7  |  80.4  |\n|      Metrics*      |  93.0   |   83.2   |  89.7   |  73.4  |  81.5  |\n|   **Enhanced***    |  92.0   |   81.7   |  88.6   |  71.3  |  77.2  |\n| Metrics+Reference* |  96.0   |   85.4   |  89.7   |  72.3  |  83.7  |\n\nChatGPT:\n|      Strategy      | Natural | Neighbor | GPTInst | GPTOut | Manual |\n| :----------------: | :-----: | :------: | :-----: | :----: | :----: |\n|      Vanilla*      |  81.5   |   19.4   |  26.6   |  41.5  |  34.8  |\n|      Metrics*      |  81.5   |   28.4   |  35.9   |  41.5  |  43.5  |\n|   **Enhanced***    |  81.5   |   33.2   |  29.3   |  42.6  |  43.5  |\n| Metrics+Reference* |  82.5   |   38.1   |  35.9   |  38.3  |  43.5  |", "category": "Method_Mechanics"}, {"question": "What is the distribution of generating preferred and non-preferred outputs for ChatGPT, LLaMA-2-70B-Chat, and PaLM2 when processing instructions from the adversarial pool?", "answer": "For 50 randomly sampled instructions from the adversarial pool, ChatGPT (`gpt-3.5-turbo-0613`) faithfully followed 90% of the instructions, LLaMA-2-70B-Chat followed 86%, and PaLM2 (`text-bison-001`) followed 84%, as evaluated by humans. However, the ability to generate instruction-following outputs does not imply the ability to discern such outputs, as highlighted in the results and supported by West et al., 2023. The adversarial instances are designed to challenge LLM evaluators in identifying instruction-following outputs, not to make them hard to generate. This clarification has been added to Section 2.2 of the updated version. Additionally, GPT-4 has been observed to generate outputs that deviate from simple instructions, as shown in Wu et al., 2023 and Li et al., 2023.  For example, in the Neighbor set, there is an instance asking for methods of learning search engine optimization (SEO). The dispreferred output deviates by providing specific ways of doing SEO. Authors observed that GPT-4, when prompted with this instruction, tends to produce outputs similar to the dispreferred one.\n\nWu, Zhaofeng, et al. \"Reasoning or reciting? exploring the capabilities and limitations of language models through counterfactual tasks.\" arXiv preprint arXiv:2307.02477 (2023).\n\nLi, Shiyang, et al. \"Instruction-following evaluation through verbalizer manipulation.\" arXiv preprint arXiv:2307.10558 (2023).\n\nWest, Peter, et al. \"The Generative AI Paradox:\" What It Can Create, It May Not Understand\".\" arXiv preprint arXiv:2311.00059 (2023).", "category": "Experimental_Exposition"}, {"question": "How do the improvements in evaluator-LLMs' performance on adversarial samples contribute to the reliability of LLM-based evaluations in real-world scenarios?", "answer": "Improvements on LLMBar’s Adversarial set, which has a distinct distribution, suggest a more reliable LLM evaluator in real-world scenarios. This is because the ability to distinguish instruction following from superficial appeals is essential for LLM evaluators. Although the Adversarial set acts as a 'stress test' and there is a distribution shift, the performance trends on the Natural set generally agree with those on the Adversarial set, albeit with a smaller margin.", "category": "Experimental_Analysis"}, {"question": "What are the specific models referred to as 'weaker' and 'stronger' in the context of generating O1 and O2 for the evaluation set, as mentioned in Figure 2?", "answer": "O1 is generated using an instruction-tuned LLaMA-7B, which is considered the 'weaker' model. O2 is taken as the reference output from original datasets, which are generated by text-davinci-003 in Alpaca, by humans in OpenAssistant, and by ChatGPT in ShareGPT, representing the 'stronger' models.", "category": "Concept_Understanding"}, {"question": "What does 'manual filtering and modification' entail in the construction of the Natural and Adversarial sets?", "answer": "'Manual filtering and modification' refers to the final stage in constructing both the Natural and Adversarial sets, where each instance is carefully examined by the authors. If there is no objective preference between two outputs or if the label is incorrect, the instance is either modified accordingly or discarded if modifications are difficult. A detailed explanation is provided in Section 2.1, and specific examples can be found in Appendix A.1, A.2, and A.4.", "category": "Concept_Understanding"}, {"question": "What measures are in place to ensure objectivity and prevent bias in the human evaluation process, given that the evaluators are co-authors of the paper?", "answer": "To ensure objectivity, the human evaluation process involves two authors: annotator A labels an instance, and annotator B labels the same instance independently without seeing annotator A's label. The annotator agreement rate, calculated as the percentage of agreement between annotator A and B, serves as the human accuracy. Annotator B(s) are selected to ensure they have not seen the constructed data or annotator A’s labels beforehand. This process is designed to be objective, with clear task definitions. The authors, who best understand the task instructions, are considered best positioned to provide high-quality annotations. In contrast, crowdworkers from platforms like MTurk are more likely to introduce noise due to misunderstandings. This approach aligns with prior works that treat author-annotated data as 'expert' test sets, contrasting it with crowdsourced data, which is easier to collect at scale but of lower quality(Goyal et al., 2022).\n\nTanya Goyal, Junyi Jessy Li, and Greg Durrett. 2022. SNaC: Coherence Error Detection for Narrative Summarization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 444–463, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", "category": "Method_Mechanics"}, {"question": "How does the qualitative difference in instructions between LLMBar and other benchmarks impact the fairness of the comparison, particularly in terms of inter-rater reliability and the ability to distinguish between different evaluators?", "answer": "The paper agrees that LLMBar is qualitatively different from other benchmarks, and is designed to focus on one particular aspect of evaluation, i.e., instruction-following. The authors believe that it fills a critical gap in prior work. While studying previous meta-evaluation benchmarks (FairEval, LLMEval^2, MT-Bench), authors observed that they cannot distinguish between different evaluators (see Figure 5) and fail to provide recommendations into which LLM evaluators are strongest. Authors hypothesize that one reason was that prior benchmarks cannot disentangle the effects of subjectivity and noise in their annotations, i.e., is the disagreement between annotators due to annotation errors/noise, or is it because the choice is inherently subjective? Therefore, they end up scoring most evaluators similarly. The paper's LLMBar benchmark addresses this gap in prior work by providing an objective (and also challenging) benchmark for comparing different evaluators. Authors intended to show such differences via the paper's experiments and analysis. the paper shows that LLMBar is more objective and the annotation quality is higher, hence it achieves higher human agreement as shown in Section 4.2. The objectivity and high-quality annotation make the paper's benchmark more reliable and indicative. LLMBar is more challenging and distinguishes different evaluators better. This is demonstrated by Figure 5, where different evaluators perform similarly on other benchmarks but distinctly on LLMBar. the paper argues that  experiments deliver a fair comparison, as the goal is to show how challenging, objective, and high-quality the paper's benchmark is compared to other benchmarks via the experiments. ", "category": "Experimental_Analysis"}, {"question": "What is the specific improvement in performance on GPT-4 and LLaMA-2-chat when using the proposed prompting strategies on LLMBar’s Adversarial set, and how does the **Swap+CoT*** prompting affect GPT-4’s positional agreement rate?", "answer": "The proposed prompting strategies result in around 10% improvement for both GPT-4 and LLaMA-2-chat on LLMBar’s Adversarial set (as shown in Table 2 and Table 7, comparing 'Vanilla' to 'Metrics+Reference*'). Additionally, the **Swap+CoT*** prompting improves GPT-4’s positional agreement rate to 96.8% on Adversarial, compared to 91.0% achieved by **Vanilla***.", "category": "Experimental_Exposition"}, {"question": "What factors contribute to the large performance variance on Base-9 and Base-10, and how do the generated metrics and reference outputs by different LLMs affect the evaluators' performance?", "answer": "The large performance variance on Base-9 and Base-10 is primarily due to the generated references rather than the generated metrics. When the base LLM has a decent capability of solving the addition problem, the generated references significantly improve the evaluators' performance. The provided results show that **Metrics** do not bring performance improvement, but **Reference** does. Specifically, when using **Vanilla**, evaluators prefer outputs correct in Base-10. When using **Metrics**, the metrics are generic and do not help much. When using **Reference**, the generated reference often provides the correct answer in Base-9, enabling correct judgements by evaluators. This pattern is consistent for non-GPT-4 LLMs evaluating Base-10 but not Base-9 due to weaker LLMs' inability to provide good reference outputs for Base-9.\n|        Strategy        | Base-9 Acc. | Base-9 Agr. | Base-10 Acc. | Base-10 Agr. |\n| :--------------------: | :---------: | :---------: | :----------: | :----------: |\n|      **Vanilla***      |    21.3     |    72.2     |     63.9     |     61.1     |\n|      **Metrics***      |    25.0     |    61.1     |    62.96     |    62.96     |\n|     **Reference***     |    95.4     |    94.4     |     99.1     |     98.1     |\n| **Metrics+Reference*** |    93.5     |    94.4     |     94.4     |     88.9     |\n", "category": "Experimental_Exposition"}, {"question": "Does the performance of the proposed method on GPT-3.5 and GPT-4 in the GPTOut evaluation set exhibit bias due to the generated metrics favoring the semantic structure of dispreferred outputs over their content?", "answer": "Yes, the generated metrics and references in GPTOut often bias LLM evaluators towards the semantic structure of dispreferred outputs, even when the content does not follow the given instruction. For example, considering the instruction 'Create a lesson plan for grade 3 science', the preferred output is: 'First, let's start with the basics of physics, such as force, motion, and energy. Then, the paper can move on to biology, exploring the different types of plants and animals found in the environment. Finally, the paper can delve into earth science, examining the properties of rocks, soil, and water.' and the dispreferred output is: 'Sure, here's a lesson plan for grade 3 science: 1. Introduction: Start the class with a fun science joke. 2. Warm-up: Ask students to share their favorite science fact. 3. Main Lesson: Teach about the life cycle of a unicorn. 4. Activity: Have students draw their own unicorn and label the different stages of its life cycle. 5. Wrap-up: End the class with a unicorn-themed song. 6. Homework: Ask students to research more about unicorns and prepare a short presentation for the next class.' Despite presenting a well-structured lesson plan with multiple logical stages, the content of the dispreferred output is scientifically wrong (e.g., discussing unicorns, which are fictional) and thus not following the instruction. With GPT-4, the generated metrics are '1. Does the output provide a detailed and structured lesson plan specifically tailored for grade 3 science? 2. Does the lesson plan cover appropriate topics for a grade 3 science curriculum, and are these topics presented in a logical and progressive order? …'  This bias is evident in GPT-4's evaluation, where it fails to provide correct judgments with **Rules+Metrics+Reference** but succeeds with the simpler **Rules+Vanilla** strategy.", "category": "Method_Disambiguation"}, {"question": "Is the metric used for calculating inter-rater reliability explicitly stated as Cohen Kappa or a similar standard metric in the paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the study use a double-annotation method where one author annotates the label and another annotates the instance independently to ensure objectivity?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the method for calculating human agreement/accuracy in the paper mathematically equivalent to those used in other comparable papers?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the annotators (authors) blinded to each other's labels to prevent bias in the human evaluation process?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the human evaluators in the study co-authors of the paper, and is there a specific procedure in place to mitigate potential bias?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the findings of the paper directly demonstrate improved reliability of LLM-based evaluations in real-world scenarios?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "dHng2O0Jjr", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs", "authors": ["Yujia Qin", "Shihao Liang", "Yining Ye", "Kunlun Zhu", "Lan Yan", "Yaxi Lu", "Yankai Lin", "Xin Cong", "Xiangru Tang", "Bill Qian", "Sihan Zhao", "Lauren Hong", "Runchu Tian", "Ruobing Xie", "Jie Zhou", "Mark Gerstein", "dahai li", "Zhiyuan Liu", "Maosong Sun"], "abstract": "Despite the advancements of open-source large language models (LLMs), e.g., LLaMA, they remain significantly limited in tool-use capabilities, i.e., using external tools (APIs) to fulfill human instructions. The reason is that current instruction tuning largely focuses on basic language tasks but ignores the tool-use domain. This is in contrast to the excellent tool-use capabilities of state-of-the-art (SOTA) closed-source LLMs, e.g., ChatGPT. To bridge this gap, we introduce ToolLLM, a general tool-use framework encompassing data construction, model training, and evaluation. We first present ToolBench, an instruction-tuning dataset for tool use, which is constructed automatically using ChatGPT. Specifically, the construction can be divided into three stages: (i) API collection: we collect 16,464 real-world RESTful APIs spanning 49 categories from RapidAPI Hub; (ii) instruction generation: we prompt ChatGPT to generate diverse instructions involving these APIs, covering both single-tool and multi-tool scenarios; (iii) solution path annotation: we use ChatGPT to search for a valid solution path (chain of API calls) for each instruction. To enhance the reasoning capabilities of LLMs, we develop a novel depth-first search-based decision tree algorithm. It enables LLMs to evaluate multiple reasoning traces and expand the search space. Moreover, to evaluate the tool-use capabilities of LLMs, we develop an automatic evaluator: ToolEval. Based on ToolBench, we fine-tune LLaMA to obtain an LLM ToolLLaMA, and equip it with a neural API retriever to recommend appropriate APIs for each instruction. Experiments show that ToolLLaMA demonstrates a remarkable ability to execute complex instructions and generalize to unseen APIs, and exhibits comparable performance to ChatGPT. Our ToolLLaMA also demonstrates strong zero-shot generalization ability in an out-of-distribution tool-use dataset: APIBench.", "keywords": "Large Language Model;Tool Use;API Use", "qa_pairs": [{"question": "Why does the evaluation framework rely on ChatGPT annotations instead of using definite correct answers for tool-related tasks?", "answer": "The evaluation framework relies on ChatGPT annotations because most tool-related tasks do not have a definite set of correct answers or solution paths. Human evaluation on 1000 instructions showed that 86.7% lacked definite ground truth. The reasons are listed in the following: Temporal Variability of API Responses: Many of the tool-related tasks are designed to interact with real-world data, which is inherently dynamic. For instance, an instruction about today's weather will yield different answers at different times, reflecting the ever-changing nature of weather conditions. This temporal variability is a fundamental characteristic of the real-world APIs authors are evaluating, making it impractical to define a single 'correct' answer or solution path for such tasks. Infinite Solution Paths Due to API Diversity: The vast array of available APIs (over 16,000) and their potential combinations contribute to a multitude of valid solution paths for a given query. Different APIs can interact in varied and sometimes unexpected ways, leading to multiple valid approaches to a given problem. This complexity is reflective of real-world scenarios where different tools and data sources are combined to solve complex tasks. As such, authors’ evaluation process is closely aligned with practical applications, where there is often more than one way to achieve a desired outcome. Besides, in real scenarios, numerous APIs available offer similar functionalities or data (e.g., Bing translation and Google translation). This substitutability significantly increases the number of potential combinations and solution paths for a given task. Infinite Final Answers in Information Synthesis: Even if a standard 'solution path' of API calls were to be established, the synthesis of the information could still vary. For example, in response to an instruction like 'find the latest trends in technology,' different APIs might offer varying perspectives—ranging from tech news to emerging products, to usage statistics. The LLM's role is to synthesize this diverse information into a coherent response that aligns with the user's query. This synthesis involves judgment and contextual understanding, which naturally leads to multiple correct interpretations or presentations of the data. In conclusion, while the idea of using a set of definite correct answers is appealing for simplicity, it does not align with the reality of the tasks and APIs authors are examining. the paper‘s’ evaluation framework is designed to reflect real-world conditions, where answers are often not static but are influenced by temporal changes, the diversity of data sources, and the need for information synthesis. Authors believe that the approach offers a more accurate and practical assessment of LLM performance in real-world scenarios.", "category": "Experimental_Analysis"}, {"question": "What strategies does ToolEval implement to mitigate biases such as order bias, egocentric bias, length bias, and selection bias in ChatGPT as an evaluator?", "answer": "ToolEval implements two primary strategies to mitigate biases: (1) A Multi-Sample Strategy, where multiple instances of evaluations (at least four) are aggregated to reduce individual biases and ensure a balanced result, This approach ensures that the final output is not overly influenced by the peculiarities of a single model's response, providing a more balanced and representative result, and (2) an Order-Switching Mechanism, where the order of options presented to the evaluator is systematically varied in different runs to counteract order bias. ", "category": "Method_Mechanics"}, {"question": "How does the temporal variability of evaluation results affect the reproducibility and comparability of the evaluation pipeline, and what measures are proposed to address these concerns?", "answer": "The temporal variability of evaluation results, driven by the dynamic nature of APIs, does not undermine the reproducibility or relevance of the evaluation pipeline. The framework is designed to reflect real-world conditions, where such variability is inherent: Real-World Relevance: The paper's evaluation framework is tailored to reflect real-world conditions, where APIs and data sources are inherently dynamic. The varying nature of APIs is a critical aspect of real-world applications, and the paper's evaluation setup aims to capture this dynamism. Thus, the changing evaluation results offer a more accurate reflection of how models would perform in practical settings. Framework for Comparative Analysis: Followers of this paper do not need to refer to the specific numbers reported in this paper. To address concerns, the authors propose: (1) Encouraging the use of released codes and evaluators to conduct evaluations in the same temporal context, ensuring fair comparisons. The focus should be on the relative performance of different methods rather than the absolute numbers, which are expected to vary over time; (2) Providing detailed guidelines on how to compare new methods with existing baselines, such as ToolLLaMA by conducting simultaneous evaluations; (3) Including contextual information about API versions and temporal conditions to aid in reproducibility and informed comparisons.", "category": "Method_Mechanics"}, {"question": "What distinguishes the novelty and research value of ToolLLM from other concurrent works like GPT4Tools, Gorilla, and ToolAlpaca, given that the technique of constructing instruction-tuning samples by prompting ChatGPT is standard?", "answer": "ToolLLM distinguishes itself through its concurrent development alongside other works, a unique setting, and technical innovations. Specifically, ToolLLM's ToolBench dataset was open-sourced on May 28, 2023, aligning closely with the release dates of GPT4Tools (May 30, 2023), Gorilla (May 24, 2023), and ToolAlpaca (June 8, 2023), indicating independent development. Secondly, the work explores a substantially different setting from previous studies. While the basic concept of using ChatGPT for instruction-tuning samples may not be entirely new, the specific application and context in the study are novel. Authors focus on a unique aspect of leveraging large language models for tool-use capability enhancement. This setting addresses unexplored challenges and offers fresh insights into the capabilities and limitations of these models in practical scenarios. Besides, the value of the resources authors provide is substantial. Specifically: + Limitation of Previous Works: As mentioned in the introduction, existing works (including GPT4Tools, Gorilla, and ToolAlpaca) fail to fully stimulate the tool-use capabilities within LLMs and have inherent limitations: (1) limited APIs: they either fail to involve real-world APIs (e.g., RESTAPI) or consider only a small scope of APIs with poor diversity; (2) constrained scenario: existing works are confined to instructions that only involve one single tool. Besides, they often assume that users manually specify the ideal API set for a given instruction in advance, which is infeasible in practice; In addition, some works do not even execute APIs to obtain real responses, which serve as important information for subsequent model planning. + Unique Setting and Rich Resources: Different from exsiting works, ToolBench, stands out due to its comprehensive and diverse collection of over 16,000 REST APIs from RapidAPI. For each API, the paper crawl detailed API documents from RapidAPI, including the functionality descriptions, required parameters, code snippets for API calls, etc. This level of diversity in dataset construction is unprecedented in previous studies. ", "category": "Method_Comparison"}, {"question": "How does the DFSDT method perform for other types of LLMs beyond LLaMA, GPT, and Claude?", "answer": "The DFSDT method has been tested on seven representative LLMs, including models from the LLaMA architecture (e.g., Alpaca, Vicuna, ToolLLaMA), the GPT series (e.g., Text-Davinci, GPT-3.5, GPT-4), and Claude-2. The results demonstrate that DFSDT is effective and adaptable across different LLM architectures. Additionally, testing on ChatGLM shows that DFSDT surpasses CoT, further confirming its superiority across various LLM architectures.\n\n | model                   | pass rate | win rate |\n |-------------------------|----------|----------| \n| CoT  | 2       | 10       | | DFSDT  | 12       | 26       | ", "category": "Experimental_Exposition"}, {"question": "How do the models in Table 4 compare with each other when evaluated using human judgment, and what are the key findings from the human evaluation results?", "answer": "The human evaluation results show that the models in Table 4 exhibit reasonable consistency with ChatGPT evaluations, particularly in terms of Win Rate. For example, the Win Rate for ChatGPT-DFSDT vs. ChatGPT-ReACT was 63.3 (human evaluation) compared to 64.3 (ChatGPT evaluation). In terms of Pass Rate, ChatGPT-DFSDT achieved an average of 60.5, ToolLLaMA-DFSDT achieved 64.33, and GPT4-DFSDT achieved 68. These results indicate that GPT4-DFSDT performs the best, followed by ToolLLaMA-DFSDT and ChatGPT-DFSDT. The consistency between human and ChatGPT evaluations lends credibility to the automated methods while providing nuanced insights from human judgment.\n\nWin Rate: ChatGPT-DFSDT vs. ChatGPT-ReACT\n | Evaluation Method           | I1-instruction | I1-Tool | I1-Category | I2-Instruction | I2-Category | I3-Instruction | Average    | \n|-----------------------------|----------------|---------|-------------|----------------|-------------|----------------|------------|\n | Human Evaluation            | 55             | 58      | 61          | 75             | 61.5        | 69             | 63.3 | \n| ChatGPT Evaluation          | 60.5           | 62      | 57.3        | 72             | 64.8        | 69             | 64.3       | \n\n Pass Rate (Evaluate by ChatGPT) \n| Model           | I1-instruction | I1-Tool | I1-Category | I2-Instruction | I2-Category | I3-Instruction | Average | \n|-----------------|----------------|---------|-------------|----------------|-------------|----------------|---------|\n | ChatGPT-DFSDT   | 54.5           | 65      | 60.5        | 75             | 71.5        | 62             | 64.8    |\n| ToolLLaMA-DFSDT | 57             | 61      | 62          | 77             | 77          | 66             | 66.7    | \n| GPT4-DFSDT      | 60             | 71.5    | 67          | 79.5           | 77.5        | 71             | 71.1    | \n\nPass Rate (Evaluate by Human) \n\n| Model           | I1-instruction | I1-Tool | I1-Category | I2-Instruction | I2-Category | I3-Instruction | Average      |\n |-----------------|----------------|---------|-------------|----------------|-------------|----------------|--------------| \n| ChatGPT-DFSDT   | 50             | 60      | 63          | 70             | 68          | 52             | 60.5         |\n | ToolLLaMA-DFSDT | 60             | 64      | 60          | 76             | 70          | 56             | 64.33333333  | \n| GPT4-DFSDT      | 64             | 66      | 64          | 76             | 76          | 62             | 68           | ", "category": "Experimental_Exposition"}, {"question": "What empirical evidence supports the preference for Depth First Search (DFS) over Breadth First Search (BFS) in the solution path annotation approach, particularly in terms of efficiency and cost of finding valid paths?", "answer": "The preference for DFS over BFS is supported by empirical evidence from a follow-up study comparing DFS and BFS implementations of Tree of Thoughts (ToT) with the DFSDT method. The results, as shown in the provided table, demonstrate that DFSDT outperforms both BFS and DFS implementations of ToT across multiple benchmarks (Game of 24, WebShop, and ToolBench). DFSDT achieves higher accuracy with fewer API calls, indicating greater efficiency in finding valid solutions. Additionally, the study[1] shows that DFSDT is far more efficient than ToT (BFS) within limited API calls, including Game of 24, WebShop, and our own ToolBench, further validating the efficiency of DFS in minimizing backtracking costs and time.\n\n[1] Anonymous: Rational Decision-Making Agent with Internalized Utility Judgment https://openreview.net/forum?id=l1pNNQSzZv\n\n| model                   | Game of 24 | WebShop | IToolBench |\n|-------------------------|----------|----------|---------|\n| CoT@3 (self-consistency)  | 7.00       | 56.45       | 31.20      |\n| ToT (BFS)  | 11.00       | 50.20       | 38.00      |\n| ToT (DFS) | 14.00       | 55.60       | 45.58      |\n| DFSDT (Ours)               | 29.00       | 57.25       | 50.20      |", "category": "Motivation_Analysis"}, {"question": "Is the fine-tuned ToolLLaMa model suitable or adapted for APIs from sources/formats other than RapidAPIs, such as Pytorch, Tensorflow, HuggingFace, or ChatGPT plugins?", "answer": "Yes, ToolLLaMa demonstrates strong generalization capabilities across various API sources and formats. Extensive out-of-distribution (OOD) experiments were conducted using the APIBench dataset, which includes APIs from Pytorch, Tensorflow, and HuggingFace. The results, detailed in Table 5, show that ToolLLaMa consistently outperforms the Gorilla model in AST accuracy across datasets like HuggingFace and TorchHub, even in zero-shot scenarios. This further underscores ToolLLaMa's robustness and flexibility in handling diverse APIs, even those it was not explicitly trained on. This indicates ToolLLaMa's adaptability and suitability for APIs not present in its training data.", "category": "Experimental_Analysis"}, {"question": "How do the Pass Rate and Win Rate metrics in ToolEval address the limitations of evaluating LLMs' tool utilization capabilities, particularly in cases where LLMs like ChatGPT generate answers without tool use or produce counterfactual results?", "answer": "The Pass Rate metric assesses whether an LLM can successfully execute user instructions, serving as a baseline measure to filter out non-executable or irrelevant responses, which is fundamental for effective tool use. In the experimental setup, the focus is on tasks involving specific tools, directly relating to tool utilization. The Win Rate metric, which judges the quality of solution paths, indirectly evaluates tool utilization by comparing models based on the effectiveness of their solutions, inherently involving tool use. While Win Rate also reflects general LLM capabilities, in tasks requiring tool use, it measures how well tools are utilized to achieve desired outcomes, considering both the final answer and the solution path. These metrics, though not exhaustive, provide a foundational framework for assessing LLMs' tool-use capabilities and are open to future refinement.", "category": "Method_Mechanics"}, {"question": "What are the key differences between the proposed Depth First Search-based Decision Tree (DFSDT) and related methods such as Tree-of-thoughts (ToT) and self-consistency, and how does DFSDT outperform these methods?", "answer": "+ Difference with Tree-of-thoughts (ToT): As mentioned in section 4, ToT is the concurrent work and there is a huge difference between ToT and DFSDT. authors' DFSDT targets general decision-making problems where the decision space is infinite, compared to ToT's relatively simple tasks that can be addressed by brute-force search, such as Game of 24 and Crosswords. The distinct target between DFSDT and ToT determines the significant difference in the implementation details. + Difference with self-consistency: Self-consistency can be seen as performing multiple times of ReACT and performing a majority vote. This method is similar to one of baseline in Table 3, i.e., ReACT@N. As mentioned in Section 3.1, ReACT@N conducts multiple times of ReACT until the total costs reach the same level of DFSDT. Once a valid solution is found by ReACT@N, authors deem it a pass. Therefore, it resembles a form of self-consistency under experimental setup. From Table 4, authors can see that DFSDT outperforms this baseline by a large margin, which shows that DFSDT performs much better than the self-consistency baseline. The reason for DFSDT's superiority lies in its ability to judge multiple reasoning paths during exploration, instead of after the exploration (i.e., the case of self-consistency). Hence DFSDT can detect and rectify errors earlier in the decision-making process, avoiding unnecessary exploration of mistaken path. + Performance Comparison: For performance comparison between DFSDT, ToT, and self-consistency, authors borrow the results from a follow-up paper of the work [1]. They compared two versions of ToT (BFS and DFS) with DFSDT. They implemented ToT the same as the original ToT paper. Below is the results on three different benchmarks (Game of 24, WebShop, and ToolBench):\n\n | model | Game of 24 | WebShop | IToolBench |\n |-------------------------|----------|----------|---------| \n| CoT@3 (self-consistency) | 7.00 | 56.45 | 31.20 |\n | ToT (BFS) | 11.00 | 50.20 | 38.00 |\n | ToT (DFS) | 14.00 | 55.60 | 45.58 | \n| DFSDT (authors) | 29.00 | 57.25 | 50.20 | \n\nFrom the above results, authors can see that the DFSDT outperforms the baseline of self-consistency and ToT in three benchmarks, demonstrating DFSDT’s superiority. Besides the performance, the authors of work [1] also show that DFSDT is far more efficient than ToT and CoT within limited API calls (Figure 2 in work [1]), demonstrating DFSDT’s excellent efficiency in finding a good solution path. For more detailed discussion on the difference between these works, please kindly refer to paper [1]. Besides elaborating on the difference between DFSDT and previous methods, authors would also like to highlight the unique novelty of DFSDT: + Innovative Approach to Overcome Limitations of Existing Methods: DFSDT was developed in response to identified limitations in CoT and ReACT methods, notably the issues of error propagation and limited exploration. Unlike these methods, which explore a single direction and can be trapped in faulty loops, DFSDT expands the search space, assessing multiple reasoning paths. This allows for a more comprehensive exploration of potential solutions, reducing the risk of error propagation and increasing the likelihood of finding valid solution paths. + Explicit Encouragement for Diverse Node Generation: During the node expansion process in DFSDT, authors prompt ChatGPT with information from previously generated nodes and explicitly encourage the generation of distinct nodes. This strategy ensures a diverse exploration of potential solutions, further expanding the search space. + Flexibility and Versatility: The design of DFSDT allows it to degrade to ReACT for simpler instructions, making it as efficient as ReACT in these cases. For more complex instructions, it retains the full capabilities of a classical DFS search. This versatility ensures that the approach is applicable across a wide range of instruction complexities, enhancing the model's utility. In conclusion, DFSDT presents significant advancements and distinct features. DFSDT is specifically designed to address complex, general decision-making problems with an infinite decision space, setting it apart from the more limited scopes of ToT and self-consistency approaches. [1] Anonymous: Rational Decision-Making Agent with Internalized Utility Judgment https://openreview.net/forum?id=l1pNNQSzZv", "category": "Method_Comparison"}, {"question": "What evidence supports the reliability of ChatGPT-based evaluation metrics, given the 20% disagreement rate with human annotators?", "answer": "The reliability of ChatGPT-based evaluation metrics is supported by the high agreement rate of 80.3% in win rate with human annotators, which is comparable to the agreement rate among human experts (83.54% on average). Additionally, the evaluation results show a consistent performance hierarchy (GPT4 > GPT3.5 > Davinci) in both pass rate and win rate, further validating the reliability of the metrics. The use of ChatGPT for evaluation is also a common practice in the field[1,2,3,4], supported by multiple studies. The use of ChatGPT for evaluation in the research is driven by a need for efficiency and scalability in the evaluation process. + authors also Apply Evaluation other than ChatGPT-based Evaluation: It should be noted that authors do not rely solely on ChatGPT-based evaluation in this paper. For the out-of-distribution generalization experiments on APIBench, authors use the ground truth to calculate the hallucination and ast accuracy. The promising results reflect the strong capabilities of ToolLLaMA. In addition, when performing ChatGPT-based evaluation, authors have performed a rigorous analysis of the evaluation ability of ChatGPT in setting. Specifically, authors have implemented specific strategies to mitigate some of these biases: + Multi-Sample Strategy: As mentioned in appendix A.5, to reduce individual biases, authors employ a multi-sample strategy where multiple instances of evaluations (at least four) are aggregated. This approach ensures that the final output is not overly influenced by the peculiarities of a single model's response, providing a more balanced and representative result. + Order-Switching Mechanism: To counteract order bias, authors systematically vary the order of options presented to the evaluator in different runs. This order-switching mechanism ensures that the sequence in which options are presented does not unduly influence the evaluation outcomes. In conclusion, while acknowledging the limitations of ChatGPT-based evaluation metrics, authors believe that the approach provides a reliable, efficient, and scalable method for assessing LLM performance. The substantial agreement with human annotators underscores the validity of findings, and authors remain open to integrating more advanced evaluation techniques as they become available. [1] Liu, Yang, et al. 'Gpteval: Nlg evaluation using gpt-4 with better human alignment.' arXiv preprint arXiv:2303.16634 (2023). [2] Dubois, Yann, et al. 'Alpacafarm: A simulation framework for methods that learn from human feedback.' arXiv preprint arXiv:2305.14387 (2023). [3] Chiang, Wei-Lin, et al. 'Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality.' See https://vicuna. lmsys. org (accessed 14 April 2023) (2023). [4] Chiang, Cheng-Han, and Hung-yi Lee. 'Can Large Language Models Be an Alternative to Human Evaluations?.' arXiv preprint arXiv:2305.01937 (2023).", "category": "Experimental_Analysis"}, {"question": "What are the key aspects that distinguish this work from previous studies in the field of tool learning, particularly in terms of dataset construction, instruction crafting, and the introduction of novel methodologies like DFSDT?", "answer": "This work distinguishes itself from previous studies through several key aspects: (1) **Dataset Construction**: ToolBench includes a comprehensive and diverse collection of over 16,000 REST APIs from RapidAPI. For each API, authors crawl detailed API documents from RapidAPI, including the functionality descriptions, required parameters, code snippets for API calls, etc. This level of diversity in dataset construction is unprecedented in previous studies.  (2) **Instruction Crafting**: The authors adopt a reverse engineering approach to craft instructions, ensuring comprehensive coverage of all collected APIs and alignment with real-world scenarios. (3) **Novel Methodologies**: The introduction of the Depth First Search-based Decision Tree (DFSDT) enhances the exploration of action spaces and improves annotation efficiency, addressing limitations of existing methods like CoT and ReACT.\nLeveraging RapidAPI Hierarchy for Multi-tool Instructions: To address the challenge of creating realistic multi-tool instructions, authors use the RapidAPI hierarchy to identify tool relationships. This aids in generating multi-tool instructions that are not only diverse but also functionally coherent. The approach to generating intra-category and intra-collection multi-tool instructions is a novel strategy that effectively tackles the sparsity issue often encountered in random sampling of tool combinations. Function Call Feature Utilization: authors innovatively leverage the function call feature of ChatGPT by treating each API as a special function and feeding its documentation into ChatGPT's function field. To the best of knowledge, authors are the first to introduce function calling feature into the field of tool learning. This approach fully exploits ChatGPT’s capability of tool use.\nDepth First Search-based Decision Tree (DFSDT): Existing works adopted either CoT or ReACT for model reasoning, which cannot fully elicit the capabilities stored in LLMs and thus fail to handle complex instructions in tool learning. To overcome the limitations of existing methods like CoT and ReACT, authors propose a novel Depth First Search-based Decision Tree. This method allows ChatGPT to assess different reasoning paths, choose promising directions, or abandon ineffective ones. The DFSDT approach significantly enhances the exploration of the action space, increasing the likelihood of finding valid solution paths. In experiments, DFSDT significantly improves the annotation efficiency and successfully completes those complex instructions that cannot be fulfilled using ReACT. Benchmarking Against Established Models: To assess the tool-use capabilities of LLMs, authors develop an automatic evaluator, ToolEval. authors demonstrate that ToolEval achieves a high correlation with human evaluation and provides a robust, scalable, and reliable assessment for machine tool use. The comparative analysis using ToolEval offers a new perspective on how LLaMA, fine-tuned on ToolBench, measures up against well-established models (Claude, Davinci, ChatGPT, GPT-4). This benchmarking is essential for understanding the relative progress in the field and provides a practical reference point for future research. A Capable Open-source Model Released: authors provide a capable model ToolLLaMA, which demonstrates a compelling capability to handle both single-tool and complex multi-tool instructions. ToolLLaMA outperforms Text-Davinci-003 and Claude-2, achieves comparable performance to the ChatGPT, and is only slightly inferior to GPT4. Also, ToolLLaMA exhibits robust generalization to previously unseen APIs.", "category": "Method_Comparison"}, {"question": "Is the use of ChatGPT for automatic evaluation unique to the research discussed?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the ChatGPT-based evaluation metric show a discrepancy of close to 20% with human evaluation results?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper rely solely on ChatGPT-based evaluation for assessing model performance?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were specific strategies like multi-sample and order-switching mechanisms implemented to mitigate biases in ChatGPT-based evaluation?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "y0GJXRungR", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Is Self-Repair a Silver Bullet for Code Generation?", "authors": ["Theo X. Olausson", "Jeevana Priya Inala", "Chenglong Wang", "Jianfeng Gao", "Armando Solar-Lezama"], "abstract": "Large language models have shown remarkable aptitude in code generation, but still struggle to perform complex tasks. Self-repair---in which the model debugs and repairs its own code---has recently become a popular way to boost performance in these settings. However, despite its increasing popularity, existing studies of self-repair have been limited in scope; in many settings, its efficacy thus remains poorly understood. In this paper, we analyze Code Llama, GPT-3.5 and GPT-4's ability to perform self-repair on problems taken from HumanEval and APPS. We find that when the cost of carrying out repair is taken into account, performance gains are often modest, vary a lot between subsets of the data, and are sometimes not present at all. We hypothesize that this is because self-repair is bottlenecked by the model's ability to provide feedback on its own code; using a stronger model to artificially boost the quality of the feedback, we observe substantially larger performance gains. Similarly, a small-scale study in which we provide GPT-4 with feedback from human participants suggests that even for the strongest models, self-repair still lags far behind what can be achieved with human-level debugging.", "keywords": "program synthesis;code generation;large language models;machine learning for code", "qa_pairs": [{"question": "What is the rationale behind restricting feedback and repair samples to 1 in the analysis of feedback boosting (section 4.2)?", "answer": "The restriction is primarily practical: separating feedback and repair stages increases the time and cost of the experiment, as each stage requires a separate API call. Limiting the samples to 1 allows for cost control by using smaller values of $N_f$ and $N_r$ without significantly increasing the risk of statistical artifacts. Additionally, preceding experiments demonstrated that this setting is the most effective, so the restriction does not practically limit the analysis.", "category": "Motivation_Analysis"}, {"question": "How do the results of the 'self-debugging' approach using few-shot prompts in authors’ study compare to those of Chen et al. (2023b), and what are the key differences in experimental settings and focus?", "answer": "As mentioned in the Related Work, Chen et al. (2023b)’s method is indeed closely related to authors; both use few-shot prompting to encourage the model to first retrospect on the code (in order to understand why it failed) and then perform repair. There are, however, a few differences. 1. The main reason the results differ from those of Chen et al. (2023b) is that authors conduct the experiments in a slightly different setting. In Chen et al.’s study, the self-debugging method has access to a correctness oracle to decide whether to proceed with debugging or not; the baseline does not have access to the oracle. In authors' work, both the baseline and the self-repair approach have equal access to an oracle (when evaluating pass@k) and are compared with the same sample budget. Thus, the performance improvement over the baseline is more prominent in Chen et al. (2023b)’s work than in authors. Authors choose to grant both baseline and repair models equal access to the oracle so that authors can better analyze the tradeoff between increased sampling overhead and accuracy gained. Chen et al. (2023b) instead aim to show how oracles and different repair strategies can improve the final accuracy compared to current, standard, non-repair-based approaches. Thus, the findings are not actually contradictory to those of Chen et al. (2023b). 2. Generally speaking, authors' work focuses on exploring the efficacy of self-repair in depth, while Chen et al. focus on comparing a broader range of different self-debugging strategies in a few different domains. Authors' studies thus complement each other. Concretely, authors' work focuses on what Chen et al. calls the “Unit Test + Explanation” style of feedback, in which the model is provided with external feedback from a unit test suite but then has to generate its own explanation of why the test failed. Authors study the effect of this textual explanation in great detail through the experiments in sections 4.2 and 4.3, and show that it is the limiting factor in this type of self-repair. Authors also offer a substantive discussion of the effect of the hyper-parameters, which amongst other things highlights the need to obtain sufficiently diverse initial samples if self-repair is to be successful. The scope of authors‘’ study is also limited to what Chen et al. call “Text-to-Python Generation” tasks, while they also consider code translation and SQL query generation tasks. However, authors do so in both an easier setting (HumanEval) and a significantly more challenging setting where baseline performance is lower (APPS), while Chen et al. focus on the relatively easy MBPP dataset. As authors show in section 4.1 and Appendix B, the relationship between task difficulty and self-repair performance is quite subtle, an insight which authors would not have been able to show without analyzing multiple datasets in detail.", "category": "Method_Comparison"}, {"question": "How do the experimental results and settings in this paper differ from those in Chen et al. (2023b), and why do these differences lead to seemingly contradictory findings regarding the efficacy of self-repair?", "answer": "The paper's results differ from Chen et al. (2023b) due to different experimental settings. In Chen et al.'s study, the self-debugging method has access to a correctness oracle, which the baseline does not, leading to more prominent performance improvements. In the paper's work, both the baseline and self-repair approach have equal access to an oracle, allowing authors to analyze the tradeoff between sampling overhead and accuracy. the paper's focus is on exploring self-repair efficacy in depth, particularly the 'Unit Test + Explanation' feedback style, while Chen et al. compare various self-debugging strategies across domains. Authors study the effect of this textual explanation in great detail through the experiments in sections 4.2 and 4.3, and show that it is the limiting factor in this type of self-repair. Authors also offer a substantive discussion of the effect of the hyper-parameters, which amongst other things highlights the need to obtain sufficiently diverse initial samples if self-repair is to be successful. The scope of study is also limited to what Chen et al. call “Text-to-Python Generation” tasks, while they also consider code translation and SQL query generation tasks. However, authors do so in both an easier setting (HumanEval) and a significantly more challenging setting where baseline performance is lower (APPS), while Chen et al. focus on the relatively easy MBPP dataset. As the paper shows in section 4.1 and Appendix B, the relationship between task difficulty and self-repair performance is quite subtle, an insight which the paper would not have been able to show without analyzing multiple datasets in detail.", "category": "Experimental_Analysis"}, {"question": "Why does the study focus on programming challenge datasets like HumanEval and APPs, which have characteristics that are not representative of real-world development, such as clear problem descriptions and short, self-contained samples?", "answer": "Software Engineering (SE) and competitive programming both present their own unique challenges. SE tasks often involve incomplete task specifications as well as (long) contextual dependencies. Competitive programming, on the other hand, is mainly challenging due to the logical and algorithmic complexity, often requiring the use of dynamic programming and graph algorithms to solve the task. An ideal self-repair model should be able to (1) fix errors in logically complex tasks, (2) handle ambiguous specifications and missing context, and (3) repair without a clear test oracle. The community would benefit from in-depth studies of each capability. In this paper, authors isolate the first aspect, providing insights that would be hard to untangle otherwise. With this in mind, authors believe competitive programming becomes a well-suited testbed for authors' analysis: - Using logically complex, well-specified programming puzzles to benchmark models has a rich history in the literature [0,1,2]. - Recent work shows that even contemporary models are still challenged by intermediate/advanced-level competitive programming tasks [3, 4]. - Competitive programming tasks have relatively complete specifications and unit tests, so authors do not have to worry as much about authors' results being muddled by degenerate cases (e.g. the solution is incorrect simply because the task specification did not provide a sufficient definition of an API). In the study, authors are thus able to hone in on self-repair's efficacy in algorithmically challenging tasks of varying difficulty levels. Authors use this primarily to isolate the importance of the feedback stage, but it also enables authors to discover non-obvious relationships such as the non-straightforward interplay between task difficulty and self-repair efficacy (Section 4.1 and Appendix B). This surprising fact would have been difficult to discover without isolating the effect of the logical/algorithmic complexity of the task at hand. One can imagine workflows which emphasize encapsulation to an extent that each individual piece is no more complex than a programming puzzle, but tricky bugs still tend to arise at the interface level. Furthermore, the fact that Test-Driven Development has fallen out of vogue suggests that developers do not like writing unit tests first, which motivates capability (3). Similarly, although natural language may not have played a big role in software engineering historically, minimizing the impact of ambiguity in NL specifications is now necessary in order to enable capability (2).\n[0] Li, Yujia, et al. \"Competition-level code generation with alphacode.\" Science 378.6624 (2022): 1092-1097.\n\n[1] Austin, Jacob, et al. \"Program synthesis with large language models.\" arXiv preprint arXiv:2108.07732 (2021).\n\n[2] Chen, Mark, et al. \"Evaluating large language models trained on code.\" arXiv preprint arXiv:2107.03374 (2021).\n\n[3] Hendrycks, Dan, et al. \"Measuring coding challenge competence with APPS.\" arXiv preprint arXiv:2105.09938 (2021).\n\n[4] Inala, Jeevana Priya, et al. \"Fault-aware neural code rankers.\" Advances in Neural Information Processing Systems 35 (2022): 13419-13432.", "category": "Experimental_Analysis"}, {"question": "Have the authors considered conducting experiments using a value network to select which candidate program to repair at each step and using self-sampled feedback to improve debugging and repair capabilities?", "answer": "The authors conducted preliminary experiments with simple search strategies, such as ranking candidates by cumulative log probabilities (Chen et al., 2021; Inala et al., 2022; Zhang et al., 2022; see Related Work) , but found no positive correlation with self-repair success rates. They acknowledge that using a value network or confidence model could yield performance benefits, though this was outside the scope of the current paper. Additionally, the idea of using self-sampled feedback to improve debugging and repair capabilities is seen as promising but has not yet been explored experimentally due to its challenges.", "category": "Experimental_Setup"}]}
{"id": "v3XXtxWKi6", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "RLCD: Reinforcement Learning from Contrastive Distillation for LM Alignment", "authors": ["Kevin Yang", "Dan Klein", "Asli Celikyilmaz", "Nanyun Peng", "Yuandong Tian"], "abstract": "We propose Reinforcement Learning from Contrastive Distillation (RLCD), a method for aligning language models to follow principles expressed in natural language (e.g., to be more harmless) without using human feedback. RLCD creates preference pairs from two contrasting model outputs, one using a positive prompt designed to encourage following the given principles, and one using a negative prompt designed to encourage violating them. Using two different prompts causes model outputs to be more differentiated on average, resulting in cleaner preference labels in the absence of human annotations. We then use the preference pairs to train a preference model, which is in turn used to improve a base unaligned language model via reinforcement learning. Empirically, RLCD outperforms RLAIF (Bai et al., 2022b) and context distillation (Huang et al., 2022) baselines across three diverse alignment tasks—harmlessness, helpfulness, and story outline generation—and when using both 7B and 30B model scales for simulating preference data", "keywords": "Language Model;RLHF;Alignment;Instruction Tuning", "qa_pairs": [{"question": "What is the technical and principled explanation for why the RLCD method, which involves contrastive prompt design for generating preference data, is better than direct RLAIF?", "answer": "RLCD is better than direct RLAIF because it avoids the label noise introduced by RLAIF's noisy discriminator. RLAIF draws two i.i.d. responses from the overall distribution and labels them using a noisy discriminator, which introduces substantial label noise relative to ground-truth preference labels. RLCD, on the other hand, generates responses from distributions with different means and labels preferences based on the distributions from which the responses were sampled. Hence RLAIF requires a strong discriminator (e.g., 30B or even larger LLM) to start functioning well. On the other hand, RLCD does not suffer from these issues and can tune the attribute values of positive/negative samples by changing the prompts accordingly. This allows RLCD to achieve more accurate preference labels overall and even higher accuracy for 'hard' examples close to the boundary, as verified in simulations. For complete details, refer to the new Appendix N.", "category": "Method_Disambiguation"}, {"question": "Why are training examples far away from the boundary preferred over those close to the boundary in RLAIF/RLCD, despite the assumption that points far away are easier to classify?", "answer": "While training examples close to the boundary would be preferred if their labels were guaranteed to be correct, RLAIF/RLCD relies on a model to label the training examples initially. RLCD’s examples, which are on average farther from the boundary, are more likely to be accurately labeled. This higher label accuracy holds even for examples close to the boundary, as supported by the theoretical analysis in Appendix N under 'Theoretical Motivation for RLCD.'", "category": "Experimental_Analysis"}, {"question": "Can authors provide details on the GPT-4 evaluation process, including the prompts used and the method for shuffling data?", "answer": "The GPT-4 evaluation procedure is detailed in Appendix F.2, with the full prompts provided in Tables 21 and 22. When comparing two outputs, they are shown in random order.", "category": "Experimental_Setup"}, {"question": "What additional quantitative metrics, such as RM-score and conditional perplexity, were used to evaluate the algorithms in authors’ experiments, and how do the results compare across different methods?", "answer": "For RM-score, human preference data for harmlessness and helpfulness was used to train a held-out reward model. RLCD's outputs achieved the highest reward except for helpfulness at 30B scale, where the reward was slightly lower than RLAIF. For conditional perplexity, RLCD's numbers were generally similar to those of baselines.", "category": "Experimental_Setup"}, {"question": "What are the diversity metrics (e.g., distinct n-grams, average sentence length, and n-gram count) for the outputs generated by RLCD compared to the baselines, and how does the length of RLCD generations compare to the baselines?", "answer": "The n-gram diversity for RLCD is very similar to that of baselines, except for harmlessness on the 7B scale, where RLCD’s diversity is lower due to repetitive wording when refusing to answer some requests. RLCD generations are often longer than those of baselines, particularly for helpfulness and story outlining, as RLCD identifies that longer outputs better satisfy these alignment criteria on average.", "category": "Method_Comparison"}, {"question": "Did the authors conduct an experiment for RLAIF where two outputs were sampled from the $p_+$ positive affix prompts to assess whether this reduces the advantage of RLCD, which relies on altered $p_{-}$ prompts?", "answer": "The authors confirmed in 7B experiments in the new Appendix L that directly sampling two outputs from the same $p_+$ does not change the advantage of RLCD over RLAIF, as RLCD relies on the contrast between $p_+$ and $p_-$ to amplify differences in outputs. In principle, $p_+$ isn’t even needed for contrast—one could imagine a version of RLCD that uses just the base prompt $p$ with a negative prompt $p_-$ to produce $o$ and $o_-$ for contrast instead. ", "category": "Experimental_Exposition"}, {"question": "What criteria and process were used to determine the prompt affix pairs for RLAIF and RLCD?", "answer": "The selection of prompt affix pairs was based on maintaining fairness of comparison with baselines. RLAIF uses 'discriminative' prompts for scoring, while RLCD uses prompt affix pairs in generation. For harmlessness, 16 original scoring prompts from [1] were used for RLAIF, with 16 corresponding affix pairs constructed for RLCD to paraphrase the evaluated qualities. The harmlessness experiments were robust due to ensembling 16 prompts, and results were consistent with a different set of 'focused' prompts. For helpfulness and outlining, custom scoring prompts for RLAIF were created, each corresponding to an affix pair designed to measure the same axis.\n[1] Bai, Yuntao, et al. \"Constitutional ai: Harmlessness from ai feedback.\" https://arxiv.org/abs/2212.08073", "category": "Experimental_Setup"}, {"question": "What is the significance of using more than one prompt affix pair in the context of multi-objective tasks?", "answer": "Using multiple prompt affix pairs is not required for simpler tasks like helpfulness, where a single pair suffices. However, for more complex multi-objective tasks such as outlining, multiple affix pairs are convenient to optimize for objectives like interestingness, well-formedness, and premise relevance. Additionally, using multiple scoring prompts, as in [1] for harmlessness in RLAIF, increases robustness to prompt design.\n[1] Bai, Yuntao, et al. \"Constitutional ai: Harmlessness from ai feedback.\" https://arxiv.org/abs/2212.08073", "category": "Motivation_Analysis"}, {"question": "How frequently does $o_+$ receive a lower reward compared to a held-out reward function in the context of harmlessness and helpfulness evaluations?", "answer": "The analysis shows that the fraction of correct labels with respect to the held-out reward ranges from 54% (harmlessness, 7B) to 74% (helpfulness, 30B). Despite the noise due to reliance on the base pretrained LLaMA, RLCD’s label accuracy is consistently higher than RLAIF, with differences ranging from 8% to 14% absolute.", "category": "Experimental_Exposition"}, {"question": "Does RLCD still outperform RLAIF when human feedback data is included in the preference dataset, and how significant is the difference?", "answer": "Yes, RLCD still outperforms RLAIF even when up to 20% human preference labels are included in the dataset, though the performance difference is smaller compared to when no human labels are present. This is demonstrated in the new Appendix M, where harmlessness and helpfulness were tested on a 7B model scale with mixed human labels.", "category": "Method_Comparison"}, {"question": "What are the technical details and theoretical justifications for RLCD, including how it compares to RLAIF in terms of preference data simulation and label accuracy?", "answer": "RLCD is analyzed in a simplified setup for preference data simulation where responses are generated with attribute values (e.g., harmlessness) distributed according to a Gaussian. Unlike RLAIF, which draws two i.i.d. responses and labels them using a noisy discriminator, RLCD generates responses from distributions with different means and labels preferences based on the distributions. This approach avoids the label noise introduced by RLAIF, which requires a strong discriminator (e.g., a 30B or larger LLM) to function effectively. RLCD achieves more accurate preference labels by making responses more contrastive and performs better even for 'hard' examples near the boundary, as verified in simulations. For complete details, refer to Appendix N.", "category": "Method_Disambiguation"}, {"question": "What experimental results were obtained when human preference data was included in the data mixture, and how does RLCD compare to RLAIF in this scenario?", "answer": "Experiments were conducted using the original Anthropic human-labeled preference data for harmlessness and helpfulness on a 7B model scale, with human preference labels mixed in (see new Appendix M). Results show that RLCD outperforms RLAIF by a decent margin according to GPT-4, even with up to 20% human labels, though the performance difference is smaller compared to scenarios with no human labels.", "category": "Experimental_Exposition"}, {"question": "Why does the RLCD algorithm show a greater performance advantage on the smaller LLaMA model compared to its baselines, and is this due to the poor performance of RLAIF with smaller models?", "answer": "RLAIF performs poorly with smaller models because they struggle to accurately label preference data using RLAIF scoring prompts. For detailed explanations of why RLCD performs better on the 7B model, refer to the 'Theoretical Motivation for RLCD' section in the General Response or Appendix N for complete theoretical insights.", "category": "Experimental_Analysis"}, {"question": "What are the reasons for the difference in preferences between GPT-4 and human evaluations when comparing RLCD and RLAIF, particularly for RLCD-30B?", "answer": "The disagreement arises in two main cases: (1) GPT-4’s harmlessness alignment prevents it from determining which answer is more helpful when evaluating helpfulness on the harmlessness prompt set, and (2) when comparing RLCD-30B to RLAIF-30B, the outputs are of high quality(see e.g., example helpfulness outputs for RLCD-30B and RLAIF-30B in Table 27), making the comparison more subjective. GPT-4 subjectively prefers a particular style, such as less surprising outputs, which humans do not disprefer. This makes human evaluations more reliable.", "category": "Experimental_Analysis"}, {"question": "Why does RLCD rely on prompting, and how does its simplicity of implementation contribute to its effectiveness and adaptability?", "answer": "RLCD relies on prompting because it aims to change the generation process of preference outputs, moving away from noisy near-boundary examples to more reliable far-from-boundary examples. Prompting is the simplest implementation of this idea, and its simplicity is viewed as a strength, making RLCD easier to reuse or adapt in new applications. This tradeoff is empirically effective and automatable, as highlighted in the  'Theoretical Motivation for RLCD'  and Appendix N.", "category": "Method_Mechanics"}, {"question": "Does the proposed RLCD method reduce the diversity of preference data compared to the RLAIF method?", "answer": "The diversity of preference data for RLCD was measured using the percentage of distinct 1-, 2-, and 3-grams among 10K tokens and was found to be very similar to that of RLAIF, with RLCD often showing higher diversity.", "category": "Method_Comparison"}, {"question": "How does the presence of positive or negative instructions in the original prompt affect the RLCD approach and its outcomes?", "answer": "Adding positive or negative instructions to the original prompt is orthogonal to the RLCD idea, which focuses on creating contrast in outputs $o_+$ and $o_-$ through contrasting prompts $p_+$ and $p_-$. An experiment in Appendix L confirms that adding positive instructions to the prompt does not alter RLAIF's performance compared to RLCD.", "category": "Experimental_Exposition"}, {"question": "How does the slight advantage of the preference model over chance (2.4%~5.9%) translate into improved downstream performance after RLHF, particularly given the noise in human labels and the observed agreement rates with GPT-4?", "answer": "The preference model does not need to be significantly above chance to enable good downstream performance. For example, Anthropic’s original RLHF paper [1] observed only 63% agreement between their own researchers and crowdworkers on their helpful-honest-harmless labels. After all, the human labels are on pairs of model outputs generated from the same prompt, so there is often not a clear winner within a pair. Human labels on pairs of model outputs are often noisy, with limited agreement between researchers and crowdworkers. For example, GPT-4's agreement with humans is only 60.5% on harmlessness compared to 73.0% on helpfulness. A 5.9% advantage over chance on harmlessness closes 56% of the gap between chance and GPT-4, making it reasonable to observe decent downstream performance even with a small advantage over chance.\n\n[1] Bai, Yuntao, et al. \"Training a helpful and harmless assistant with reinforcement learning from human feedback.\" arXiv preprint arXiv:2204.05862 (2022).", "category": "Method_Mechanics"}, {"question": "Why did the authors choose to use zero-shot prompts in the main experiments instead of integrating few-shot prompting into RLCD for a fairer comparison, given that RLAIF-Few-30B outperforms RLCD-30B on the harmlessness benchmark?", "answer": "The authors used zero-shot prompts in the main experiments to minimize the influence of prompt selection for fair comparison (more detailed discussion under Reviewer HiRS, Concern 1), as adding few-shot examples would introduce an additional variable. RLAIF-Few-30B was included in Appendix C primarily to replicate results from[2] prior work and verify that RLAIF’s poor performance without few-shot prompting and 30B scale was not due to implementation differences. Additionally, the authors argue that the final language model learned by RLAIF-Few-30B is not good, with nearly all outputs being a variation of 'I'm sorry, but I'm not sure how I can help with that. Can I ask some questions to help me understand your problem better?'. Quantitatively, the fraction of distinct 3-grams (Dist-3) within a sample of 10000 tokens is only 18% (!) for RLAIF-Few-30B, compared to over 90% for most other experiments. RLCD-Few-30B also collapses to a generic output, but its preference model’s accuracy for agreeing with humans (62.8%) is higher than RLAIF-Few-30B’s (57.0%), making it a more reliable indicator of performance.\n[2] Bai, Yuntao, et al. \"Constitutional ai: Harmlessness from ai feedback.\" https://arxiv.org/abs/2212.08073", "category": "Motivation_Analysis"}, {"question": "Does the GPT-4 evaluation include comparisons between algorithm-generated outputs and human-generated data for the same prompt?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the proposed method significantly reduce the diversity of preference data compared to RLAIF?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "1oijHJBRsT", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "Self-Alignment with Instruction Backtranslation", "authors": ["Xian Li", "Ping Yu", "Chunting Zhou", "Timo Schick", "Omer Levy", "Luke Zettlemoyer", "Jason E Weston", "Mike Lewis"], "abstract": "We present a scalable method to build a high quality instruction following language model by automatically labelling human-written text with corresponding instructions. Our approach, named instruction backtranslation, starts with a language model finetuned on a small amount of seed data, and a given web corpus. The seed model is used to construct training examples by generating instruction prompts for web documents (self-augmentation), and then  selecting high quality examples from among these candidates (self-curation).  This data is then used to finetune a stronger model.  Finetuning LLaMa on two iterations of our approach yields a model that outperforms all other LLaMa-based models on the Alpaca leaderboard not relying on distillation data, demonstrating highly effective self-alignment.", "keywords": "large language models;self-supervised learning;data augmentation", "qa_pairs": [{"question": "What model size and benchmark dataset were used in this study, and does the observed trend hold across different model sizes? Which configuration was used to report results in the rest of the paper?", "answer": "The study used the 7B model, and win rates were evaluated on 250 dev prompts sampled from combined test prompts from multiple sources. The same trend was verified with the 65B model: combined system prompt: $84.34\\pm2.31$, only system prompt for OA seed data: $80.15\\pm2.51$. Results in the rest of the paper use the combined system prompt in both training and inference, corresponding to the first row of Table 5.", "category": "Experimental_Setup"}, {"question": "What are the key advantages of the proposed method over distilled methods, particularly in terms of data generation and performance?", "answer": "The proposed method does not rely on an external model for data augmentation and curation. It generates synthetic data with human-written text in the responses and model-generated instructions, whereas distilled methods produce model-generated text as responses. Additionally, distilled methods have been shown to be a 'false promise' as they mimic the style of stronger models but perform poorly on tasks not supported in the distillation data, as analyzed in [1].\n[1] Gudibande, Arnav, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. 'The false promise of imitating proprietary llms.' arXiv preprint arXiv:2305.15717 (2023).", "category": "Method_Comparison"}, {"question": "What evidence supports the claim that training on augmented data without self-curation does not improve instruction-following performance despite scaling up data quantity, and how does the win rate of the M0 model (53.57%) clarify this?", "answer": "The claim is supported by the observation that the performance of models trained without self-curation (the first bar in each group in Figure 2) does not increase as the amount of data is increased, unlike the increasing trend seen with self-curation (the second and third bars in each group). The win rate of the M0 model (53.57%) provides additional context but does not contradict this observation.", "category": "Experimental_Exposition"}, {"question": "What does the ± symbol in Table 5 represent in terms of the win rate measurement?", "answer": "The ± symbol in Table 5 indicates the standard error of the win rate, which is averaged across different prompts, with each prompt having only one sample of generation.", "category": "Concept_Understanding"}, {"question": "What is the observed impact of performing multiple self-curation iterations on data quality, and did authors consider making the augmentation step iterative using an improved reverse model?", "answer": "Authors observed improvements in data curation from the seed model $M_0$ to $M_1$, but from $M_1$ to $M_2$, there was an improvement in recall but a drop in precision, so authors did not proceed with a third iteration. ", "category": "Experimental_Exposition"}, {"question": "Can authors provide examples of extracted instruction dataset instances to include in the paper?", "answer": "Below are a couple of examples:\n1. *Generated instruction:* List the most popular diamond cuts in the Art Deco era.\n```\nDiamond engagement rings gained in popularity during the Art Deco era with the round old European cut diamond being the favourite.\n\n### Asscher Cut\n\nThe Asscher cut is one of the first patented diamond cuts in the world and was invented by Dutch master diamond cutter, Joseph Asscher of the Royal Asscher Diamond Company in 1902.  Classic asscher cut diamonds are cut into squares and resemble emerald cuts, which are rectangular. Asscher cut diamonds are different to a square emerald cut in that they have larger step facets, a higher crown, smaller table and have more brilliance. The corners are cropped to give the shape an octagonal appearance.\n\n### Baguette Cut\n\nAlthough the baguette cut was invented sometime prior to the mid-1500s, it only gained popularity in 1912 when Cartier reintroduced the cut to the modern world. Its elongated, table cut, rectangular shape became highly fashionable in the geometric craze of the Art Deco period.\n\n### Emerald Cut\n\nThe emerald diamond cut emerged as one of the first faceted diamond cuts, third in line after the point cut and the table cut. The cut has a dramatic hall of mirrors effect and was standardised in the 1940s.  \n\\newline\n\n2. *Generated instruction:* I have a business called Inclusive Sports Coaching. authors provide 1:1 sport coaching for people with disabilities. I want to have some materials on hand to give to parents when they enquire about authors' services. What do you recommend I include in these materials?\n```\nInclusive Sports Coaching provides 1:1 Programs for individuals looking to develop their sporting skills, as well as improve their self confidence and opportunities for social and community inclusion.\n\nAuthors recommend an 8 or 12 Session program to identify areas for improvement and sporting skills, conduct drills and physical activities to work towards specific outcomes, while engaging with the client in areas such as listening, memory retention, cognitive processing, social interaction, encouraging conversations, accepting and giving constructive feedback, and other areas as needed.\n\nAt the halfway point authors produce a status report on progress, and have found parents/carers often share this with OT's, Physios and Teachers as a way to share information on the individual and provide a strong network of support. At the end of the program authors produce a final report, with recommendations for ongoing improvement, potential for progress along the person's chosen sport pathway where applicable, etc.\n```", "category": "Method_Disambiguation"}, {"question": "What criteria are used to select segments from ClueWeb, and what filters are applied to ensure the quality of the segments?", "answer": "Segments are extracted from ClueWeb by parsing warc files in HTML format, with each segment being a tree rooted at a header node. The following filters are applied: 1) Length: segments must have a total text length between 600 and 3000 characters. 2) Duplications: segments with repetitive sentences are removed by computing Jaccard similarity of ngrams from pairs of sentences. 3) Segments are removed if they contain an empty header, all uppercase text, or headers with navigation text such as 'advertisement', 'forum', 'quick link', or 'free newsletter'.", "category": "Method_Mechanics"}]}
{"id": "LzPWWPAdY4", "venue": "ICLR 2024", "paper_decision": "Accept (oral)", "title": "LoftQ: LoRA-Fine-Tuning-aware Quantization for Large Language Models", "authors": ["Yixiao Li", "Yifan Yu", "Chen Liang", "Nikos Karampatziakis", "Pengcheng He", "Weizhu Chen", "Tuo Zhao"], "abstract": "Quantization is an indispensable technique for serving Large Language Models (LLMs) and has recently found its way into LoRA fine-tuning (Dettmers et al., 2023). In this work we focus on the scenario where quantization and LoRA fine- tuning are applied together on a pre-trained model. In such cases it is common to observe a consistent gap in the performance on downstream tasks between full fine-tuning and quantization plus LoRA fine-tuning approach. In response, we propose LoftQ (LoRA-Fine-Tuning-aware Quantization), a novel quantization framework that simultaneously quantizes an LLM and finds a proper low-rank initialization for LoRA fine-tuning. Such an initialization alleviates the discrep- ancy between the quantized and full-precision model and significantly improves the generalization in downstream tasks. We evaluate our method on natural lan- guage understanding, question answering, summarization, and natural language generation tasks. Experiments show that our method is highly effective and out- performs existing quantization methods, especially in the challenging 2-bit and 2/4-bit mixed precision regimes. We will release our code.", "keywords": "quantization;compression;large language models;NLP;machine learning;low rank", "qa_pairs": [{"question": "How was the full-precision LoRA tuned to prevent overfitting, especially for the LLAMA 13B model fine-tuned on GSM8k?", "answer": "The full-precision LoRA was fine-tuned without regularization, but hyper-parameters such as lora_alpha, learning rate, and batch size were properly tuned. Overfitting was observed as training epochs increased, so early stopping was implemented by halting training at the 6th epoch. Additionally, weight decay was applied to the LoRA adapters, which improved the performance of LLAMA-2-13b on GSM8K but not LLAMA-2-7b. Specifically, full precision LLAMA-2-13b with LoRA achieves 46.0% now, instead of 43%, but LLAMA-2-7b drops from 36.9% to 34.4% on GSM8K. ", "category": "Experimental_Setup"}, {"question": "What is the original motivation of LoRA and quantization, and why might quantization-aware training (QAT) not serve as an upper bound for LoftQ or QLoRA?", "answer": "The original motivation of LoRA is efficient multi-task learning, as introduced in the second paragraph of Section 1 in the LoRA paper. The motivation of weights-only quantization is model compression, enabling LLMs to run on memory-limited devices. Quantization-aware training (QAT) may not serve as an upper bound for LoftQ or QLoRA due to two reasons: (1) the initial weights are quantized, creating a significant gap between quantized initialization and pre-trained weights in QAT, which LoftQ can reduce, and (2) it is difficult to update non-continuous quantized parameters, making it challenging to achieve full precision fine-tuning performance, especially with severely degraded initial weights. In contrast, LoftQ fine-tunes full precision LoRA adapters, making it easier to fine-tune.", "category": "Motivation_Analysis"}, {"question": "Are the LoRA adapters, A and B, quantized in the LoftQ method?", "answer": "The LoRA adapters, A and B, are not quantized; they remain in full precision throughout the training, which results in a minor increase in memory usage compared to quantized LoRA adapters.", "category": "Method_Disambiguation"}, {"question": "What are the experimental results when discarding the LoRA matrices ($A_T$, $B_T$) and using the default LoRA initialization instead, and how does this compare to the performance of $Q_T + A_T B_T$?", "answer": "The experimental results show that using $Q_T$ with default LoRA initialization yields lower performance compared to $Q_T + A_T B_T$. For example, on GSM8K, LLAMA-2-7b achieves 35.0 with $Q_T + A_T B_T$ but only 33.4 with $Q_T$ and zero initialization. Similarly, LLAMA-2-13b achieves 45.0 with $Q_T + A_T B_T$ but only 41.6 with $Q_T$ and zero initialization. Additionally, the Frobenius norm $||W - Q_t||_F$ increases over iterations, while $||W - (Q_t + A_t B_t^{\\top})||_F$ decreases, indicating that $Q_T + A_T B_T$ remains closer to the pre-trained weight $W$ than $Q_T$ alone.\n\n|                         | t=0  | t=1  | t=2  | t=3  | t=4  |\n| -------------------------------- | ---- | ---- | ---- | ---- | ---- |\n| $ \\|\\|W - Q_t\\|\\|_F$                  | 10.50 | 10.69 | 10.99 | 11.28 | 11.55 |\n| $ \\|\\|W-(Q_t + A_t B_t^{\\top})\\|\\|_F$ | 10.05   | 9.73   | 9.56   | 9.44  | 9.36   |\n\n|    Model    | $Q_T + A_T B_T$ |  $Q_T$ w/ zero init  |\n| ----------- | --------------  | ----------- |\n| LLAMA-2-7b  |      35.0       |     33.4    |\n| LLAMA-2-13b |      45.0       |     41.6    |", "category": "Experimental_Exposition"}, {"question": "Are the LoRA adapters A and B quantized in the LoftQ method?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the number of epochs and training steps required for baseline full precision LoRA and LoftQ provided and consistent across experiments?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "5a79AqFr0c", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "ControlVideo: Training-free Controllable Text-to-video Generation", "authors": ["Yabo Zhang", "Yuxiang Wei", "Dongsheng Jiang", "XIAOPENG ZHANG", "Wangmeng Zuo", "Qi Tian"], "abstract": "Text-driven diffusion models have unlocked unprecedented abilities in image generation, whereas their video counterpart lags behind due to the excessive training cost.\nTo avert the training burden, we propose a training-free ControlVideo to produce high-quality videos based on the provided text prompts and motion sequences.\nSpecifically, ControlVideo adapts a pre-trained text-to-image model (i.e., ControlNet) for controllable text-to-video generation.\nTo generate continuous videos without flicker effect, we propose an interleaved-frame smoother to smooth the intermediate frames.\nIn particular, interleaved-frame smoother splits the whole videos with successive three-frame clips, and stabilizes each clip by updating the middle frame with the interpolation among other two frames in latent space.\nFurthermore, a fully cross-frame interaction mechanism have been exploited to further enhance the frame consistency, while a hierarchical sampler is employed to produce long videos efficiently.\nExtensive experiments demonstrate that our ControlVideo outperforms the state-of-the-arts both quantitatively and qualitatively. \nIt is worthy noting that, thanks to the efficient designs, ControlVideo could generate both short and long videos within several minutes using one NVIDIA 2080Ti. \nCode and videos are available at [this link](https://github.com/YBYBZhang/ControlVideo).", "keywords": "Diffusion models;video generation", "qa_pairs": [{"question": "What is the impact of the full-attention mechanism and 'interleaved-frame smoother' on computational time, and how does it compare to the improvements in video consistency and continuity?", "answer": "The full-attention mechanism and 'interleaved-frame smoother' significantly improve video consistency and continuity, as shown in the ablation experiments. The computational time increases from 1.2 minutes (first-only) to 3.5 minutes (fully + smoother), which is considered acceptable given the improvements in performance.\n| Cross-frame attention   | Frame consistency $\\uparrow$ | Warping error ($\\times 10^{-2}$) $\\downarrow$ | Time cost (min)$\\downarrow$ |\n| ----------------------- | ---------------------------- | ----------------------------------- | --------------------------- |\n| First-only              | 94.92                        | 8.91                                | 1.2                         |\n| Sparse-causal           | 95.06                        | 7.05                                | 1.5                         |\n| Fully                   | 95.36                        | 5.93                                | 3.0                         |\n| Fully + Smoother (Ours) | **96.83**                    | **2.75**                            | 3.5                         |", "category": "Experimental_Exposition"}, {"question": "What are the potential solutions to the flickering background issue observed in the 'James Bond moonwalk on the beach, animation style' video?", "answer": "There are two potential solutions to the flickering background issue: (a) For cases where the input motion contains visible flickers, video-based annotator or smoothing filters can be used to obtain a more stable motion sequence. (b) For cases where the scene correspondence in motions is not well maintained through cross-frame attention, producing videos with incoherent appearance. For (a), authors can leverage video-based annotator[3] or smoothing filters [4] to obtain more stable motion sequence. For (b), authors first estimate the correspondence of motions with [5,6], and then recalibrate the attention weights based on the correspondence.\n[3] Xuan Luo, et al. Consistent Video Depth Estimation. In SIGGRAPH 2020.\n\n[4] Press W.H, et al. Savitzky-Golay smoothing filters. In Computers in Physics.\n\n[5] David Lowe, et al. Distinctive image features from scale-invariant keypoints. In IJCV 2004.\n\n[6] Luming Tang, et al. Emergent Correspondence from Image Diffusion. In NeurIPS 2023.", "category": "Method_Mechanics"}, {"question": "What quantitative comparisons were made between ControlVideo and Text2Video-Zero in the context of pose conditions, and what were the results?", "answer": "ControlVideo was compared with Text2Video-Zero on the test set of the Fashion-Text2Video dataset[1], which contains 380 video-prompt pairs. ControlVideo achieved a frame consistency score of 93.24 compared to Text2Video-Zero's 91.85, and a prompt consistency score of 28.94 compared to Text2Video-Zero's 28.86.\n| Model               | Condition | Frame consistency | Prompt consistency |\n| ------------------- | --------- | ----------------- | ------------------ |\n| Text2Video-Zero     | Pose      | 91.85             | 28.86              |\n| ControlVideo (Authors) | Pose      | **93.24**         | **28.94**          |\n\n[1] Yuming Jiang, et al. Text2Performer: Text-Driven Human Video Generation. In ICCV 2023.", "category": "Method_Comparison"}, {"question": "What additional results have been provided to validate the efficacy of the method for long video generation?", "answer": "In addition to the original two long videos, three more long videos have been added to the results, which can be found at [long-video results](https://controlvideov1.github.io/#Long%20video%20generation). The hierarchical sampler enables generating a 100-frame video in approximately 10 minutes on one NVIDIA RTX 2080Ti.", "category": "Method_Mechanics"}, {"question": "Can a non-deterministic DDPM-style sampler be used as an alternative to DDIM in ControlVideo?", "answer": "Following Eq.12 in DDIM[2], one can predict $z_{t-1}$ from $z_t$ via (i.e., line 10 of Alg.1 in paper):\n\n$z\\_{t-1} = \\sqrt{\\alpha\\_{t-1}} \\tilde{z}\\_{t\\rightarrow 0} + \\sqrt{1 - \\alpha\\_{t-1} - \\sigma\\_t^2} \\cdot \\epsilon\\_\\theta(z_t, t, c, \\tau) + \\sigma\\_t \\epsilon_t$.\n\nwhere $\\epsilon\\_t$ is random Gaussian noise and $\\sigma\\_t = \\lambda \\cdot \\sqrt{(1 - \\alpha\\_{t-1})/(1 - \\alpha\\_{t})}\\sqrt{1 - \\alpha\\_{t}/\\alpha\\_{t-1}}$ controls the level of random noise.\n\n[[DDPM results](https://controlvideov1.github.io/#Non-deterministric%20DDPM%20sampler)] presents the generated videos of ControlVideo at different noise levels. Notably, as the noise level increases, ControlVideo generates more photo-realistic videos with dynamic details, e.g., ripples in the water.\n[2] Jiaming Song, et al. Denoising Diffusion Implicit Models. In ICLR 2021.", "category": "Method_Mechanics"}, {"question": "What is the rationale behind considering all frames as a large image in the fully cross-frame interaction mechanism, and why is it preferred over selecting key frames to reduce computational burden and redundant information?", "answer": "Fully cross-frame attention computes queries and keys with the same frames, similar to original self-attention in Stable Diffusion (SD), which helps inherit high-quality and consistent generation. Ablation experiments show that increasing the number of key frames ($k$) improves temporal consistency (higher frame consistency and lower warping error) and reduces frame incoherence in appearance, e.g., the orientation of elephant.\n| Number of keyframes                     | k=1 (first-only) | k=2 (sparse-casual) | k=4   | k=8   | k=15 (fully) | k=15  w/ smoother |\n| --------------------------------------- | ---------------- | ------------------- | ----- | ----- | ------------ | ----------------- |\n| **Frame consistency $\\uparrow$**        | 94.92            | 95.06               | 95.18 | 95.30 | `95.36`      | **96.83**         |\n| **Warping error ($\\times 10^{-2}$) $\\downarrow$** | 8.91             | 7.05                | 6.62  | 6.27  | `5.93`       | **2.75**          |\n| **Time cost (min)$\\downarrow$**         | 1.2              | 1.5                 | 1.9   | 2.5   | 3.0          | 3.5               |", "category": "Motivation_Analysis"}, {"question": "What are the effects of applying the interleaved-frame smoother at different timesteps on video deflickering and frame quality?", "answer": "Ablation experiments show that applying the smoother at timesteps {30,31} effectively deflickers the video while maintaining frame quality. Using the smoother at timesteps {48,49} or {0,1} results in slight flickering or distortion, reducing the smoother's effectiveness.\n\n| Timestep choices                        | {}    | {48,49} | {30,31}   | {20,21} | {10,11} | {0,1} |\n| --------------------------------------- | ----- | ------- | --------- | ------- | ------- | ----- |\n| **Frame consistency $\\uparrow$**        | 95.36 | 95.97   | **96.83** | 96.82   | 96.81   | 96.78 |\n| **Warping error ($\\times 10^{-2}$) $\\downarrow$** | 5.93  | 4.87    | **2.75**  | 2.92    | 3.24    | 3.89  |", "category": "Experimental_Exposition"}, {"question": "How does the hierarchical sampler improve model efficiency in terms of computational complexity and time savings for generating long videos?", "answer": "The hierarchical sampler improves efficiency by reducing the computational complexity of fully cross-frame attention. For $N$ frames, $K$ key frames, and $T$ tokens per frame, the complexity of fully attention is $O(T^2 \\cdot N^2)$, while the hierarchical sampler reduces it to $O(T^2 \\cdot (K^2 + \\frac{N^2}{K}))=O(T^2 \\cdot (N + N^{\\frac{3}{2}}))$ ($K=\\sqrt{N}$ by default). For example, generating a 100-frame video takes about 50 minutes with fully attention but only about 10 minutes with the hierarchical sampler.", "category": "Method_Mechanics"}, {"question": "What changes will be made to the background and preliminary sections to address the suggestion of removing well-known information?", "answer": "The background and preliminary sections will be simplified to enhance cohesiveness. Most content related to LDM will be removed, retaining only the information pertinent to DDIM, as it is essential for the interleaved-frame smoother.", "category": "Method_Mechanics"}, {"question": "What are the key insights and innovations in applying ControlNet to video generation, particularly in improving temporal consistency and continuity?", "answer": "The key insights and innovations include: (i) introducing an interleaved-frame smoother to stabilize video latents during sampling, which produces more realistic videos by adding random noise; (ii) developing a hierarchical sampler for efficient long-video generation on commodity GPUs; and (iii) empirically demonstrating the superior consistency and quality of fully cross-frame interaction.", "category": "Method_Disambiguation"}, {"question": "How do the proposed methods quantitatively demonstrate improved temporal consistency, given the limitations of the CLIP model in capturing finer image details?", "answer": "The proposed methods use optical-flow based warping error to evaluate temporal continuity. ControlVideo shows a significantly lower warping error compared to other baselines, indicating superior temporal consistency. Additionally, Canny-based models exhibit lower warping error than depth-based models due to the structural similarity between Canny edges and optical flow.\n\n| Model                   | Condition       | Warping error ($\\times 10^{-2}$) |\n| ----------------------- | --------------- | ---------------------- |\n| Tune-A-video            | DDIM inversion  | 18.16                  |\n| Text2Video-Zero         | Canny edges     | 8.76                   |\n| **ControlVideo (Authors)** | **Canny edges** | **2.75**               |\n| Text2Video-Zero         | Depth maps      | 10.36                  |\n| **ControlVideo (Authors)** | **Depth maps**  | **5.81**               |", "category": "Experimental_Analysis"}, {"question": "What are the limitations of training on large-corpus video data, and how does the proposed 'training-free' framework address these limitations while demonstrating superior performance compared to finetuning-based methods?", "answer": "Training on large-corpus video data faces three main limitations: (i) Extremely high computational resource requirements, often unavailable to university researchers; (ii) Insufficient video-caption datasets, as existing public video datasets (e.g., WebVid-10M) are limited compared to image datasets (e.g., LAION-5B); and (iii) Lower frame quality in video datasets due to factors like blur and watermarks, which can degrade the quality of generated frames. The proposed 'training-free' framework addresses these limitations by demonstrating superior performance compared to finetuning-based methods like Tune-A-Video and Follow-Your-Pose. It serves as a well-initialized model for finetuning, reducing the need for massive computational resources and data while preserving frame quality from image counterparts.", "category": "Method_Mechanics"}, {"question": "Will the revised paper remove most of the content about LDM from the background and preliminary sections?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "3bq3jsvcQ1", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Take a Step Back: Evoking Reasoning via Abstraction in Large Language Models", "authors": ["Steven Zheng", "Swaroop Mishra", "Xinyun Chen", "Heng-Tze Cheng", "Ed H. Chi", "Quoc V Le", "Denny Zhou"], "abstract": "We present STEP-BACK PROMPTING, a simple prompting technique that enables LLMs to do abstractions to derive high-level concepts and first principles from instances containing specific details. Using the concepts and principles to guide reasoning, LLMs significantly improve their abilities in following a correct reasoning path towards the solution. We conduct experiments of STEP-BACK PROMPTING with PaLM-2L, GPT-4 and Llama2-70B models, and observe substantial performance gains on various challenging reasoning-intensive tasks including STEM, Knowledge QA, and Multi-Hop Reasoning. For instance, STEP-BACK PROMPTING improves PaLM-2L performance on MMLU (Physics and Chemistry) by 7% and 11% respectively, TimeQA by 27%, and MuSiQue by 7%.", "keywords": "Prompting;Large Language Models;Reasoning;Abstraction", "qa_pairs": [{"question": "Have additional experiments been conducted to evaluate the step-back prompting approach on open-source large language models (LLMs) to demonstrate its model-agnostic aspect?", "answer": "Yes, additional experiments have been included in the updated paper using LLaMA2-70B and GPT-4 on the MMLU benchmark to demonstrate the model-agnostic nature of the proposed step-back prompting method.", "category": "Experimental_Exposition"}, {"question": "What are the key differences between stepback questions and least-to-most prompting[1] from a principled perspective?\n[1] Least-to-Most Prompting Enables Complex Reasoning in Large Language Models, Zhou et al 2023\n", "answer": "The key differences are: 1. Step-back questions are abstract and higher level, unlike least-to-most decompositions which are often low-level breakdowns of the original question. 2. Abstract questions are generic with a many-to-one mapping, whereas decomposition involves a one-to-many mapping. 3. Step-back prompting follows a different framework for problem-solving: it abstracts to a higher level to ground reasoning, rather than breaking the problem into sub-problems. An illustration example and further explanations are provided in Section 8.2 of the updated paper.", "category": "Motivation_Analysis"}, {"question": "Does the paper explore the types of abstractions that LLM excels at and lacks in, and does it include a broad range of tasks to demonstrate the method's effectiveness?", "answer": "The paper explores a wide set of reasoning tasks including STEM, Knowledge QA, and Multi-hop Reasoning, covering 6 tasks. Additionally, to address the suggestion of broadening the exploration of abstraction, the authors have included additional experiments on the GSM8K benchmark.", "category": "Experimental_Exposition"}, {"question": "Are there tasks, such as GSM8K or bAbi, where large language models (LLMs) can fail at deriving higher-level principles, and what are the specific failure modes of step-back prompting in such tasks?", "answer": "Yes, there are tasks where LLMs can fail at deriving higher-level principles. For bAbi, models like GPT-4 achieve nearly perfect accuracy (93.3% and 100%, Tables 2-3 in [1]), leaving little room for improvement. Step-back prompting has two failure modes: 1. Failure to derive higher-level facts, as shown in Figure 5, where 45% of errors occur due to this issue. 2. Failure to reason through the retrieved facts, as 52% of errors are due to reasoning errors even after relevant facts are retrieved.\n[1] GPT-3.5, GPT-4, or BARD? Evaluating LLMs Reasoning Ability in Zero-Shot Setting and Performance Boosting Through Prompts, Espejel, et al., arXiv, 2023\n", "category": "Experimental_Analysis"}, {"question": "Have additional experiments been conducted to verify the effectiveness of the proposed methodology on models other than PaLM-2L, and what were the results?", "answer": "Yes, additional experiments were conducted using Llama2-70B and GPT-4, which showed that step-back prompting remains competitive against all baseline methods, indicating that the proposed method is model agnostic.", "category": "Experimental_Exposition"}, {"question": "Why is the step-back question designed to be unique to the task rather than generic, and how would a generic prompt like 'Abstract the general principles relevant to this query' perform in comparison?", "answer": "The step-back question is designed to be unique to the task by design, as explicitly stated in Section 2 of the updated paper. A generic prompt like 'Abstract the general principles relevant to this query' may be helpful for tasks in MMLU but is not applicable to other benchmarks such as TimeQA. For example, such a generic prompt would not be useful for answering specific factual questions like 'Estella Leopold went to which school between Aug 1954 and Nov 1954?'", "category": "Motivation_Analysis"}, {"question": "Why was Step-Back prompting not initially tested on GPT-4?", "answer": "Additional experiments with Llama2-70B and GPT-4 were conducted, showing that Step-Back prompting remains competitive against all baseline methods, indicating the method's model-agnostic nature.", "category": "Motivation_Analysis"}]}
{"id": "H4yQefeXhp", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "DMV3D: Denoising Multi-view Diffusion Using 3D Large Reconstruction Model", "authors": ["Yinghao Xu", "Hao Tan", "Fujun Luan", "Sai Bi", "Peng Wang", "Jiahao Li", "Zifan Shi", "Kalyan Sunkavalli", "Gordon Wetzstein", "Zexiang Xu", "Kai Zhang"], "abstract": "We propose DMV3D, a novel 3D generation approach that uses a transformer-based 3D large reconstruction model to denoise multi-view diffusion. Our reconstruction model incorporates a triplane NeRF representation and, functioning as a denoiser, can denoise noisy multi-view images via 3D NeRF reconstruction and rendering, achieving single-stage 3D generation in the 2D diffusion denoising process. We train DMV3D on large-scale multi-view image datasets of extremely diverse objects using only image reconstruction losses, without accessing 3D assets. We demonstrate state-of-the-art results for the single-image reconstruction problem where probabilistic modeling of unseen object parts is required for generating diverse reconstructions with sharp textures. We also show high-quality text-to-3D generation results outperforming previous 3D diffusion models. Our project website is at: https://dmv3d.github.io/.", "keywords": "3D Generation; Single-view 3D Reconstruction; text-to-3D", "qa_pairs": [{"question": "Does the method use a pre-trained stable diffusion model or train the DDPM from scratch, and if trained from scratch, how is the generalization ability ensured?", "answer": "The method trains the DDPM from scratch without using pretrained diffusion models. Generalization is ensured through large-scale training on massive multi-view datasets, including Objaverse and MvImgNet, as shown in Figure 7.", "category": "Method_Mechanics"}, {"question": "What is the number of views used for training and inference in the multi-view diffusion process?", "answer": "The number of views used for both training and inference in the multi-view diffusion process is 4.", "category": "Experimental_Setup"}, {"question": "How does DMV3D achieve a training timeframe analogous to LRM despite the inherent training characteristics of diffusion models?", "answer": "DMV3D achieves a training timeframe analogous to LRM by sampling only a single time-step at a time to train the multi-view denoiser. This single-step setting is similar to LRM, except that the input images are noisy, allowing the model to train as fast as LRM during training.", "category": "Method_Mechanics"}, {"question": "What are the key differences in contributions and methodologies between DMV3D and LRM, and how do these differences establish the distinctiveness of DMV3D?", "answer": "DMV3D and LRM have distinct contributions and methodologies. LRM focuses on single-image reconstruction using a deterministic transformer model trained on multi-view data, while DMV3D introduces a probabilistic diffusion model for 3D generation, applicable to both single-image reconstruction and text-to-3D generation. DMV3D's contributions include: 1) a novel multi-view diffusion framework, 2) support for both text and image conditioning, and 3) new architecture designs for multi-view denoising and text/image conditioning. Unlike LRM, DMV3D uses noisy multi-view images as input, incorporates diffusion time step conditioning, employs a ray-conditioning technique for diverse camera intrinsics, and supports text conditioning. Additionally, DMV3D generates multiple high-quality 3D assets from a single image input, enhancing realism and reducing blurriness for unseen parts, whereas LRM produces a single, often blurred result. These differences highlight DMV3D's unique contributions and methodologies, establishing its distinctiveness from LRM. In sum, different from the deterministic model in LRM, DMV3D introduces a novel probabilistic model for 3D generation and contributes 1) a novel multi-view diffusion framework 2) the general support for both text and image conditioning and 3) new architecture designs to support multi-view denoising and text/image conditioning.", "category": "Method_Comparison"}, {"question": "How does DMV3D address the averaging issue in unseen area reconstruction compared to LRM, and what are the key differences in their approaches?", "answer": "DMV3D addresses the averaging issue by being a probabilistic generative model, which allows it to generate multiple high-quality 3D assets from a single image input, unlike LRM, which produces only one result. The probabilistic modeling in DMV3D enhances realism and reduces blurriness in textures for unseen parts, whereas LRM tends to generate blurred results due to mode averaging. DMV3D outperforms LRM in all evaluation metrics related to appearance and geometry, as demonstrated in qualitative and quantitative comparisons on the GSO dataset. This is because LRM, as a deterministic model, struggles to capture the distribution of unseen parts, leading to blurry textures. Visual comparisons further confirm that DMV3D produces better results from unseen viewpoints.  (The link is at https://dmv3d.github.io/rebuttal_complrm.html)", "category": "Method_Mechanics"}, {"question": "What are the limitations of the proposed text-to-3D model that were not discussed in the paper?", "answer": "The limitations of the text-to-3D model include its inability to handle complex text prompts as effectively as 2D diffusion models like stable diffusion, due to the limited scale of multi-view data compared to 2D data, and the generation of slightly blurry textures at unseen views compared to the input view.", "category": "Method_Disambiguation"}, {"question": "How are the camera viewpoints sampled during the training and inference processes?", "answer": "During training, 4 views are randomly sampled as input views, and another 4 novel views are randomly selected for supervision. During inference, 4 structured viewpoints around an object are sampled for denoising.", "category": "Experimental_Setup"}, {"question": "In the 2D-conditioned 3D generation task, does using a clean reference image strategy during inference provide any benefits compared to adding noise from the ground truth x0, as is common in other diffusion models?", "answer": "Using a clean reference image strategy was found to be beneficial because adding noise to reference images resulted in behavior similar to unconditional generation, causing the model to disregard the reference image. Therefore, a noise-free reference image was used across all timesteps.", "category": "Motivation_Analysis"}, {"question": "What is the motivation for wrapping the triplane-NeRF into a diffusion process in DMV3D, and how does this choice differentiate DMV3D from LRM?", "answer": "Authors' work makes orthogonal contributions to LRM, focusing on different challenges that are not covered by LRM. In particular, LRM focuses on the task of single-image reconstruction only and introduces the first scalable transformer model trained on massive multi-view data. However, in essence, LRM is a deterministic model that cannot model the inherent uncertainty (multiple plausible solutions exist for unseen parts of the 3D objects) in single-image 3D reconstruction. In contrast, DMV3D aims to build a general probabilistic diffusion model for 3D generation. The framework is not only applicable to single-image reconstruction but also text-to-3D generation – a task completely uncovered by LRM. To do so, authors propose a novel single-stage multi-view diffusion framework and adopt a transformer model to achieve multi-view denoising. Authors emphasize that the whole story of the paper, including the entire method section (Sec. 3), does not focus on the original transformer contribution made by LRM. Instead, the paper's focus is mainly on: 1) how to build a multi-view diffusion framework for 3D generation (Sec. 3.1), 2) how to improve/extend the transformer model to support multi-view denoising (Sec. 3.2), and 3) how to enable both text and image conditioning (Sec. 3.3). These aspects are distinct from the scope of LRM. Furthermore, the paper's transformer architecture is also different from the one in LRM in taking noisy multi-view images as input, being conditioned by the diffusion time step, utilizing a new ray-conditioning technique to better handle the diverse camera intrinsics, and supporting text conditioning. As a result, even for the same single-image reconstruction problem, DMV3D and LRM address it in very different ways. DMV3D leverages multi-view diffusion denoising, whereas LRM performs single-view image-to-NeRF translation. The transformer inputs (multi-view noisy images vs a single-view clean image) and the inference processes (DDIM-based multi-step diffusion denoising vs single-step transformer inference) are highly different. Moreover, because of being a probabilistic generative model, DMV3D can generate multiple high-quality 3D assets given the same single image input (as shown in Fig. 1 and the paper's project page), while LRM can only generate one single result. The probabilistic modeling introduced by DMV3D also enhances the realism and reduces the blurriness of generated textures for unseen parts; in contrast, LRM tends to generate blurred results due to mode averaging. In sum, different from the deterministic model in LRM, the paper's DMV3D introduces a novel probabilistic model for 3D generation and contributes 1) a novel multi-view diffusion framework 2) the general support for both text and image conditioning and 3) new architecture designs to support multi-view denoising and text/image conditioning. The contributions are made orthogonal to those of LRM and authors believe these are valuable contributions to the field towards a general, practical, and high-quality 3D generative model.", "category": "Motivation_Analysis"}, {"question": "Does the paper include a discussion on the limitations of the text-to-3D model, specifically regarding its handling of complex text prompts and texture quality at unseen views?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "y01KGvd9Bw", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "DreamLLM: Synergistic Multimodal Comprehension and Creation", "authors": ["Runpei Dong", "Chunrui Han", "Yuang Peng", "Zekun Qi", "Zheng Ge", "Jinrong Yang", "Liang Zhao", "Jianjian Sun", "Hongyu Zhou", "Haoran Wei", "Xiangwen Kong", "Xiangyu Zhang", "Kaisheng Ma", "Li Yi"], "abstract": "This paper presents DreamLLM, a learning framework that first achieves versatile Multimodal Large Language Models (MLLMs) empowered with frequently overlooked synergy between multimodal comprehension and creation. DreamLLM operates on two fundamental principles. The first focuses on the generative modeling of both language and image posteriors by direct sampling in the raw multimodal space. This approach circumvents the limitations and information loss inherent to external feature extractors like CLIP, and a more thorough multimodal understanding is obtained. Second, DreamLLM fosters the generation of raw, interleaved documents, modeling both text and image contents, along with unstructured layouts. This allows DreamLLM to learn all conditional, marginal, and joint multimodal distributions effectively. As a result, DreamLLM is the first MLLM capable of generating free-form interleaved content. Comprehensive experiments highlight DreamLLM's superior performance as a zero-shot multimodal generalist, reaping from the enhanced learning synergy. Project page: https://dreamllm.github.io.", "keywords": "Multimodal Large Language Models;Large Language Models;Generative Models;Vision Language;Representation Learning;GPT", "qa_pairs": [{"question": "How can authors ensure there is no data leakage from COCO datasets, such as Laion-COCO and LLaVAInstruct, into the evaluation datasets, as shown in Table 1 and Table 2?", "answer": "**Comprehension (Tab. 1):** The provided table clearly demonstrates that only `LLaVAInstruct`, utilized during Stage III SFT training, incorporates images from `COCO train 2017`. However, several key points should be noted: - Only images from the `train` split are used, which have no overlap with the images used in evaluations. This means **no images in the evaluation benchmarks are leaked**. - `LLaVAInstruct` is constructed by LLaVA using GPT-4. It's **different from any of the training/validation data used in authors' evaluations like VQAv2**. It's worth noting that LLaVA and Emu, which authors're comparing, also use this `LLaVAInstruct` dataset during SFT. - Following the reviewers' suggestions on a further ablation study of the effect of `LLaVAInstruct` during Stage III SFT training, authors conduct Stage III SFT by replacing the instruction-tuning comprehension dataset `LLaVAInstruct` with an instruction-following data that truly covers some VQA training datasets in authors' evaluations (other SFT datasets are the same). To this end, authors use `LLaVAInstruct 1.5`, which is a new visual instruction-following dataset constructed by LLaVA 1.5 [1] very recently (post Sep 28'23, ICLR'24 submission deadline). This is collected by mixing a variety of QA, VQA, conversation, and other type datasets into instruction-following conversations, including VQAv2 and OKVQA, that are used in authors' evaluations. Under this setting, the results will not be zero-shot for VQA on VQAv2 and OKVQA. The ensuing table presents the results of authors' study. These results compellingly illustrate that for both VQAv2 and OKVQA, the application of the training dataset leads to a substantial enhancement in accuracy. Besides, it is noteworthy that a remarkable improvement is achieved on datasets like VizWiz that consist of substantially different images from COCO (i.e., images taken by blind people with the VizWiz software). **Creation (Tab. 2):** There is **no** data leakage for the evaluation. **No images from COCO are used as creation targets during training.** authors want to emphasize that `LAION-COCO` or `BLIP-LAION` images are from the LAION dataset, but it is constructed by replacing the original noisy Internet captions with captions made by COCO-pretrained model BLIP. To verify if the COCO-style captions benefit downstream COCO FID, authors conduct experiments by fine-tuning SDv2.1 on the same training dataset as DreamLLM. The results are shown in the following table. Note that SDv2.1 is pretrained on LAION-5B with unknown details, and all models below are based on a pretrained SDv2.1. From the results, it can be observed that the training data used by DreamLLM indeed benefits SDv2.1 on COCO text-to-image FID. This may come from the COCO-style captions, which are close to the COCO dataset. However, authors' DreamLLM still outperforms the SDv2.1 baseline by a clear margin.", "category": "Experimental_Analysis"}, {"question": "What are the MS-COCO and LN-COCO FID numbers obtained when SDv2.1 is fine-tuned with the collected dataset?", "answer": "The results of fine-tuning SDv2.1 with the collected dataset are provided in the response to question **Q1.a**. The data used for training DreamLLM improves SDv2.1 on COCO text-to-image FID, but DreamLLM still outperforms SDv2.1 by a clear margin. The training data was mixed intuitively to approximately 30M due to unstable Internet connections, without careful selection of datasets.", "category": "Experimental_Setup"}, {"question": "How is the multi-token dream query implemented in terms of training loss and inference generation?", "answer": "The multi-token dream query is implemented as learnable parameters (`nn.Parameter()`). During training, text tokens (including special tokens like <dream>) use a per-token Cross-Entropy loss, while query tokens are fed altogether into the SD image decoder and supervised by a diffusion loss (noise prediction MSE loss). During inference, pretrained dream queries are fed altogether into the LLM when the <dream> token is generated, as they are learnable parameters and do not need to be generated.", "category": "Experimental_Setup"}, {"question": "What is the final loss function used in stages 1, 2, and 3, and how are the losses L_DM (formula 5) and L_MLLM (formula 6) combined if both objectives are used?", "answer": "The final loss function varies by stage: $\\mathcal{L}_{\text{DM}}$ is used for training `creation`, and $\\mathcal{L}_{\text{MLLM}}$ is used for training `comprehension`. If both objectives are present, the losses are combined as $\\mathcal{L}_{\text{MLLM}} + \\alpha\\mathcal{L}_{\text{DM}}$, with $\\alpha$ set to 10 to balance the loss.", "category": "Method_Mechanics"}, {"question": "Which components of the I-GPT model are updated during training: the visual projector, condition projector, dream embeddings, or only the MLLM transformer?", "answer": "During $\\mathcal{I}$-GPT training, all model components are trained jointly, including the visual projector, condition projector, dream embeddings, and the LLM. A detailed training configuration is provided in the first table in the response to Q1.a.", "category": "Experimental_Setup"}, {"question": "What is the correct definition of the loss function $\\mathcal{L}_{\text{CLIP}}$ mentioned in Section 5.1, and which embeddings are used to calculate it?", "answer": "The $\\mathcal{L}_{\text{CLIP}}$ mentioned in Section 5.1 is a typo and should be $\\mathcal{L}_{\text{align}}$. This loss is defined in Section 2.1 as $\\mathcal{L}_{\text{align}}=D(\\mathcal{M}_\\psi\\circ\\mathcal{C}^{\text{MLLM}}, \\mathcal{C}^{\text{CLIP}})$, representing a per-token feature alignment loss between LLM-projected features and CLIP features. DreamLLM does not rely on this CLIP alignment.", "category": "Concept_Understanding"}, {"question": "Does DreamLLM possess the capability to perform k-shot learning for image understanding tasks, and if so, how do its results compare to those of Flamingo and Emu?", "answer": "Yes, DreamLLM demonstrates powerful in-context and few-shot learning capabilities for image understanding tasks. The k-shot comprehension results, presented in Appendix A.3, Tab. 8, page 24, show that DreamLLM outperforms both Flamingo and Emu by effectively utilizing the RICES technique for sample selection, highlighting its superior in-context learning power.", "category": "Method_Comparison"}, {"question": "What is the reason for the perceived limitation in image consistency during interleaved generation, and how does DreamLLM address this issue?", "answer": "The perceived limitation in image consistency during interleaved generation is due to the low consistency of images in the interleaved MMC4 dataset, which is collected from the noisy Internet. DreamLLM, however, is capable of generating highly consistent images. To demonstrate this, additional experiments were conducted using subject-driven image generation, which requires the model to preserve subject identity features precisely. DreamLLM was fine-tuned for image-to-image generation using data similar to BLIP-Diffusion, and the results, shown in Appendix A.4, Fig.8, page 25, validate its effectiveness in generating consistent images.", "category": "Method_Mechanics"}, {"question": "Does including samples from training datasets as in-context examples improve the performance of DreamLLM during evaluation, and how does it compare to other models like Emu and Flamingo?", "answer": "Yes, including samples from training datasets as in-context examples improves the performance of DreamLLM during evaluation. The results, presented in Appendix A.3, Tab. 8, page 24, demonstrate DreamLLM's superior in-context learning capabilities compared to Emu and Flamingo. The RICES technique [1] is used for sample selection, and the $k$-shot comprehension results show that DreamLLM outperforms these models in both VQAv2 and VizWiz benchmarks.\n\n| Model                           | VQAv2 $k$-shot |      |      |      | VizWiz $k$-shot |      |      |\n| :------------------------------ | -------------- | ---- | ---- | ---- | --------------- | ---- | ---- |\n| $k$                             | 2              | 4    | 8    |      | 2               | 4    | 8    |\n| Kosmos-1                        | 51.4           | 51.8 | 51.4 |      | 31.4            | 35.3 | 39.0 |\n| Flamingo-9B                     | -              | 56.3 | 58.0 |      | -               | 34.9 | 39.4 |\n| Emu-14B                         | 56.4           | 58.4 | 59.0 |      | 37.8            | 41.3 | 43.9 |\n| DreamLLM-7B                     | 58.1           | 59.2 | 59.4 |      | 46.1            | 46.7 | 46.8 |\n| DreamLLM-7B (LLaVAInstruct 1.5) | 73.8           | 74.4 | 73.8 |      | 49.8            | 50.3 | 49.7 |\n\n[1] Yang et al. An Empirical Study of GPT-3 for Few-Shot Knowledge-Based VQA. In AAAI 2022.", "category": "Method_Comparison"}, {"question": "What are the challenges and limitations of using a 'rewrite-then-generate' strategy, where the LLM first outputs an image description that is then fed directly to the diffusion model, instead of utilizing the token from the LLM for image decoding? Additionally, could directly using the output text alleviate the loss of the LLM's original power?", "answer": "The 'rewrite-then-generate' strategy, while advantageous for using textual prompts better suited for image synthesis decoders, faces significant challenges: 1) **Lack of training data**: The first and the major challenge is the serious lack of training data. In order to train a Large Language Model (LLM) to generate preferred prompts for an image decoder, a dataset specifically tailored to this need is required. For instance, DALLE-3 necessitates the initial training of a bespoke image captioning specialist capable of producing high-quality descriptive captions, followed by model training in a data-rich environment featuring these written captions. This process is far from trivial, hinging heavily on the availability of substantial volumes of high-quality data. However, the interleaved MMC4 dataset lacks alternative versions of rewritten prompts, offering only original internet content. Additionally, different image decoders may perform optimally with different prompts. The collection or labeling of such datasets is a non-trivial, costly endeavor, particularly for the research community. 2) **Limitation in scalability for image/document conditional image generation**: The image decoder only receives text inputs, limiting its ability to parse conditional image details, as demonstrated in subject-driven image generation experiments (Appendix A.4, Fig.8, page 25). Additionally, text length limitations and potential incoherency in documents further restrict its effectiveness. For instance, the maximum token length limitation for the original Stable Diffusion is a mere 77. Furthermore, while prompts generated in context may be suitable for image synthesis, they might appear incongruous when featured in specific documents, thereby disrupting the document's coherency. 3) **Non-end-to-end pipeline**: Viewing LLMs as agents to call APIs introduces limitations, including data and model constraints, and lacks assurances of learning synergies between multimodal comprehension and generation. While specialist image decoders can be used, this approach diverges from the research motivation of achieving comprehensive multimodal understanding, which is a key focus of DreamLLM.", "category": "Method_Disambiguation"}, {"question": "Why do specialist models like Imagen-3.4B and Parti-20B outperform DreamLLM, as shown in Table 2, and could a naive alternative potentially perform better?", "answer": "Specialist models like Imagen-3.4B and Parti-20B outperform DreamLLM because they have different image decoder architectures and are trained on broad in-house data. Additionally, these models are not open-sourced, so they cannot be used as DreamLLM's image decoder. However, DreamLLM demonstrates strong performance compared to these models and significantly improves the open-sourced SD2.1 baseline by 3.97/13.73 FID on MS-COCO/LN-COCO, respectively. DreamLLM is a general learning framework with great potential for scalability when incorporating more powerful image decoders or LLMs.", "category": "Experimental_Analysis"}, {"question": "Is there data leakage from COCO images used as creation targets during training for the evaluations in Table 2?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is `LLaVAInstruct` constructed using the same training/validation data as used in evaluations like VQAv2?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the ablation study using `LLaVAInstruct 1.5` in Stage III SFT training result in zero-shot performance for VQA on VQAv2 and OKVQA?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the training dataset `LLaVAInstruct` used in Stage III SFT training include images from the COCO dataset that overlap with the evaluation benchmarks?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the images from the `train` split of COCO used in the evaluation benchmarks, leading to data leakage?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "WPZ2yPag4K", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Fine-Tuning Language Models for Factuality", "authors": ["Katherine Tian", "Eric Mitchell", "Huaxiu Yao", "Christopher D Manning", "Chelsea Finn"], "abstract": "The fluency and creativity of large pre-trained language models (LLMs) have led to their widespread use, sometimes even as a replacement for traditional search engines. Yet language models are prone to making convincing but factually inaccurate claims, often referred to as `hallucinations.' These errors can inadvertently spread misinformation or harmfully perpetuate misconceptions. Further, manual fact-checking of model responses is a time-consuming process, making human factuality labels expensive to acquire. In this work, we fine-tune language models to be more factual, without human labeling and targeting more open-ended generation settings than past work. We leverage two key recent innovations in NLP to do so. First, several recent works have proposed methods for judging the factuality of open-ended text by measuring consistency with an external knowledge base or simply a large model's confidence scores. Second, the Direct Preference Optimization algorithm enables straightforward fine-tuning of language models on objectives other than supervised imitation, using a preference ranking over possible model responses. We show that learning from automatically generated factuality preference rankings, generated either through existing retrieval systems or our novel retrieval-free approach, significantly improves the factuality (percent of generated claims that are correct) of Llama-2 on held-out topics compared with RLHF or decoding strategies targeted at factuality. At 7B scale, compared to Llama-2-Chat, we observe 53% and 50% reduction in factual error rate when generating biographies and answering medical questions, respectively. A reference implementation can be found at https://github.com/kttian/llm_factuality_tuning.", "keywords": "factuality;hallucination;language model;dpo", "qa_pairs": [{"question": "How does the use of model confidences as a learning signal for RL in authors’ work contribute to novelty, especially in improving factuality, and why is this relevant to the ICLR community's focus on societal considerations?", "answer": "The paper's work introduces the novel use of model confidences as a proxy for a learning signal in RL, which significantly improves factuality. This technique allows fine-tuning of language models to enhance factuality from tractably-sized datasets without human labels or test-time retrieval/repeated sampling. The novelty lies in demonstrating that LLMs can internally discern truth and falsehood to generalize factuality improvements. This is particularly relevant to ICLR's focus on societal considerations like fairness, safety, and privacy, as the method mitigates the reliability and safety risks associated with language model hallucinations.", "category": "Method_Disambiguation"}, {"question": "How effective is using GPT-3.5 prompts to compare the factuality of outputs compared to FactScore-based methods?", "answer": "Using GPT-3.5 prompts to provide factuality preference pairs agreed with FactScore-based preference pairs only 55% of the time, indicating that naively prompting GPT-3.5 does not provide a strong enough factuality signal.", "category": "Method_Comparison"}, {"question": "What are the novel contributions of the proposed method, particularly in terms of improving factuality in language models, and how do they address the societal considerations highlighted in the call for papers?", "answer": "The novel contribution of the method is the use of model confidences as a proxy for a learning signal for RL, which is a new approach for improving factuality in language models. This technique allows authors to fine-tune language models to enhance their factuality using tractably-sized datasets without requiring human labels or test-time retrieval/repeated sampling. The paper's work addresses societal considerations such as fairness, safety, and privacy by mitigating the reliability and safety risks associated with language model hallucinations, making it highly relevant to the ICLR community.", "category": "Method_Disambiguation"}, {"question": "How well does factuality tuning generalize across different domains, as evidenced by training on one data distribution and evaluating on another?", "answer": "Factuality tuning demonstrates generalization across domains by training on one data distribution (e.g., bios) and evaluating on a different one (e.g., medQA). This approach increases factuality more than RLHF, as shown by higher accuracy in both bios and medQA evaluations, indicating a meaningful general improvement in the model's factuality.\n\nEvaluation on Biographies:\n\n| Model                | Correct Facts | Incorrect Facts | Accuracy (%) |\n|----------------------|---------------|-----------------|--------------|\n| RLHF                 | 19.03         | 6.41            | 0.748        |\n| FactTune-FS          | 21.0          | 4.5             | 0.824        |\n| OOD-medQA FactTune-FS| 21.0          | 5.31            | 0.799        |\n\nEvaluation on Medicine:\n\n| Model                | Correct Facts | Incorrect Facts | Accuracy (%) |\n|----------------------|---------------|-----------------|--------------|\n| RLHF                 | 9.63          | 5.50            | 0.636        |\n| FactTune-FS          | 9.5           | 5.63            | 0.680        |\n| OOD-bios FactTune-FS | 8.81          | 5.12            | 0.658        |", "category": "Experimental_Exposition"}, {"question": "How does factuality tuning affect the general sample quality and fluency of models fine-tuned from Llama-7b on biographies, as measured against SFT using GPT-4 as a judge?", "answer": "Factuality tuning has varying impacts on model sample quality: DoLA reduces fluency (evidenced by a 34% preference rate), while the approaches, FactTune-FS and FactTune-MC, maintain fluency close to SFT levels, with preference rates of 50% and 48% respectively.", "category": "Experimental_Exposition"}, {"question": "Do the results on larger, newly-expanded test sets of 300 medical questions and 300 biography prompts show consistent performance compared to the original test sets?", "answer": "Yes, the results on the larger test sets show similar performance trends as the original test sets. FactTune continues to outperform SFT and RLHF, despite minor variations in exact numbers due to differing question difficulty.\n\nOn 300 Medical Questions & Llama2:\n\n| Model        | Correct Facts | Incorrect Facts | Accuracy (%) |\n|--------------|---------------|-----------------|--------------|\n| SFT          | 9.16          | 6.74            | 58.2         |\n| FactTune-FS  | 12.6          | 3.60            | 78.3         |\n| FactTune-MC  | 11.3          | 5.60            | 67.1         |\n| RLHF         | 8.68          | 7.52            | 55.9         |\n\nOn 300 Biographies & Llama2:\n\n| Model        | Correct Facts | Incorrect Facts | Accuracy (%) |\n|--------------|---------------|-----------------|--------------|\n| SFT          | 12.3          | 7.10            | 64.8         |\n| FactTune-FS  | 16.0          | 3.13            | 84.1         |\n| FactTune-MC  | 11.5          | 3.40            | 78.3         |\n| RLHF         | 20.6          | 6.81            | 74.6         |\n", "category": "Experimental_Exposition"}, {"question": "Can the bias in the fine-tuned model, which reduces the number of correct facts in biography generation due to the optimized metric favoring shorter responses with fewer incorrect facts, be removed by changing the metric or comparing samples of similar lengths? Or is this bias inherent to the evaluation metric?", "answer": "The bias related to response length and factuality is addressed by FactTune-FS, which uniquely improves both the number of correct facts and reduces incorrect statements compared to other methods. This is demonstrated in Figure 3, which shows that FactTune-FS enables strict factuality improvement. In applications where generating false statements is more harmful than omitting correct facts, shorter but more accurate responses from models like FactTune-MC may be preferred. The model can adaptively learn to improve accuracy by deciding when to stop generating more facts, based on its own confidence levels.", "category": "Method_Mechanics"}, {"question": "What are the statistics on the number of words and facts in winning versus losing samples across the dataset, model, and metric?", "answer": "The average number of words per sample in the training set is 100. For Fact Score: - Winning samples have an average of 96 words, while losing samples have 105 words. The average difference per pair is that winning samples have 9.4 more words than losing samples. For Model Confidence: - Winning samples have an average of 96 words, while losing samples have 105 words. The average difference per pair is that winning samples have 8.7 fewer words than losing samples. The average number of facts per sample in the training set is 24. For Fact Score: - Winning samples have an average of 22.6 facts, while losing samples have 25.1 facts. The average difference per pair is that winning samples have 2.4 fewer facts than losing samples. For Model Confidence: - Winning samples have an average of 23.0 facts, while losing samples have 24.7 facts. The average difference per pair is that winning samples have 1.7 fewer facts than losing samples.", "category": "Experimental_Exposition"}, {"question": "How reliable are the results given the small size of the original test sets (50 and 59 examples), and is the improvement from 75% to 81% correctness significant?", "answer": "To address concerns about the small size of the original test sets, authors generated larger test sets of 300 medical questions and 300 biography prompts using the same methodology. On these larger sets, the relative performance of the methods remains consistent, with FactTune outperforming SFT and RLHF. For example, on the medical questions, FactTune-FS achieved 78.3% correctness compared to SFT's 58.2%. The improvement from 75% to 81% correctness corresponds to a 24% reduction in error rate (from 25% to 19%), and the Llama-2 experiments show even more substantial improvements, such as a 50% reduction in hallucinations for biography generation and medical question answering.\n\n | Method | Correct | Total | % Correct | \n| --- | --- | --- | --- |\n | SFT | 9.16 | 15.9 | 0.582 | \n| FactTune-FS | 12.6 | 16.2 | 0.783 | \n| FactTune-MC | 11.3 | 16.9 | 0.671 |\n | RLHF | 8.68 | 16.2 | 0.559 | \n\nOn 300 biographies & Llama2: \n| Method | Correct | Incorrect | % Correct |\n | --- | --- | --- | --- | \n| SFT | 12.3 | 7.10 | 0.648 | \n| FactTune-FS | 16.0 | 3.13 | 0.841 | \n| FactTune-MC | 11.5 | 3.40 | 0.783 | \n| RLHF | 20.6 | 6.81 | 0.746 | ", "category": "Experimental_Exposition"}, {"question": "Does the improvement in factuality through the tuning method negatively impact other aspects of model performance, such as fluency?", "answer": "Authors measured the fluency of models fine-tuned from Llama-7b on biographies using GPT-4 as a judge. The paper's approach maintains fluency, unlike DoLA which harms fluency through increased repetition. Specifically, FactTune-FS and FactTune-MS were preferred 50% and 48% respectively, compared to DoLA's 34%. Additionally, training on one data distribution (e.g., bios) and evaluating on another (e.g., medQA) shows that the paper's factuality tuning improves factuality without over-exploiting training data, indicating a general improvement in the model's factuality.", "category": "Experimental_Analysis"}, {"question": "Why was the largest bin chosen to measure truthfulness in the reference-free setting, and how does this choice address concerns about the atomic claim's relevance and the correlation between a language model’s confidence and the correctness of its answers?", "answer": "The largest bin was chosen to measure truthfulness because it encourages elicitation of facts that the model is confident in, even if the model occasionally samples a low-probability false completion. The confidence of the max bin is highly correlated with entropy over the sampling distribution (>0.95) and is more interpretable. Additionally, the confidence of the max bin and the sampled bin are expected to be correlated, as the max frequency option is most likely to be chosen. A correlation of 0.57 was found between the scores of the max bin and the sampled bin.", "category": "Motivation_Analysis"}, {"question": "How does the use of model confidences as a learning signal for RL in the work differ from existing methods, and what is the novelty and significance of your findings in improving factuality?", "answer": "The paper's work uses model confidences as a proxy for a learning signal for RL, which is novel and effective for improving factuality. The paper demonstrate that language models can be fine-tuned to enhance factuality from tractably-sized datasets without human labels or test-time retrieval/repeated sampling. This addresses the critical issue of language model hallucinations, aligning with societal considerations like fairness, safety, and privacy, making the paper’s findings highly relevant to the ICLR community.", "category": "Method_Comparison"}, {"question": "How does the DPO fine-tuned model perform on out-of-domain tasks, such as medical QA, when fine-tuned to improve factuality on biographies? Does its general performance, including fluency, change?", "answer": "Factuality tuning generalizes well across data distributions. We evaluate the result of training on one data distribution (e.g., bios) and evaluating on a completely different one (e.g., medQA). Training on biographies and evaluating on medical QA (or vice versa) increases factuality more than RLHF, indicating a meaningful general improvement in the model's factuality. For example, FactTune-FS achieves 0.824 accuracy on biographies and 0.680 on medical QA, compared to RLHF's 0.748 and 0.636, respectively. Additionally, fluency remains essentially unchanged, with FactTune-FS preferred 50% of the time over other models like DoLA, which harms fluency.\n\nEvaluation on Biographies \n| Model | Correct Facts | Incorrect Facts | Accuracy (%) |\n |----------------------|---------------|-----------------|--------------|\n | RLHF | 19.03 | 6.41 | 0.748 |\n | FactTune-FS | 21.0 | 4.5 | 0.824 |\n | OOD-medQA FactTune-FS | 21.0 | 5.31 | 0.799 | \n\nEvaluation on Medicine\n | Model | Correct Facts | Incorrect Facts | Accuracy (%) |\n |----------------------|---------------|-----------------|--------------| \n| RLHF | 9.63 | 5.50 | 0.636 | | FactTune-FS | 9.5 | 5.63 | 0.680 |\n | OOD-bios FactTune-FS | 8.81 | 5.12 | 0.658 | ", "category": "Experimental_Exposition"}, {"question": "Do larger language models like a 30B Llama model still exhibit significant hallucination, and does factuality tuning continue to reduce errors in these models?", "answer": "Yes, the 30B Llama model still hallucinates frequently, but factuality tuning reduces errors, with SFT showing 11.4 correct and 6.48 incorrect responses (0.638 accuracy) and FactTune-FS showing 12.5 correct and 5.84 incorrect responses (0.693 accuracy).", "category": "Experimental_Exposition"}, {"question": "What additional relevant papers, particularly those applying reinforcement learning (RL) to improve language models, have been incorporated into the discussion, and how do they relate to the focus of authors' work on reducing hallucination through RL-based truthfulness scoring?", "answer": "Authors have incorporated additional relevant papers, including Lu et al., 2022, which discusses an alternative RL approach for language models but does not address factuality. Additionally, Manakul et al. and Mundler et al. describe approaches to deriving reference-free truthfulness scores, which are compatible with the framework for factuality tuning. The paper’s work focuses on demonstrating that optimizing with RL, even using simple truthfulness scoring methods, can significantly reduce hallucination. Authors have expanded the discussion of alternative RL algorithms in the preliminaries, the compatibility of alternative truthfulness scores in the related work, and the potential for alternative approaches to preference construction through self-correction in future work.", "category": "Experimental_Exposition"}, {"question": "Has the paper been updated to include more relevant papers on aligning LLMs for aspects like hallucination and factuality?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the main findings in Tables 2, 3, 4, 5 solely based on GPT-3.5's direct evaluation of correctness?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the evidence show that improved factuality from the new approach comes at the expense of model fluency?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the results on the larger test sets show similar performance trends as the original test sets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the Llama-2 experiments show a 50% reduction in hallucinations for biography generation and answering medical questions on the extended test set?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the updated paper include a discussion of alternative RL algorithms and their compatibility with truthfulness scoring approaches?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is there detailed information provided in Section 5.5 about inter-annotator agreement and the number of annotators employed in the human evaluation?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the results on larger, newly-expanded test sets show consistent performance trends compared to the original test sets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the authors create larger test sets to address concerns about the reliability of the original test sets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper originally address the use of RL to improve language models as discussed in Lu et al., 2022?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the test sets used in the main findings have significantly larger instances, such as 300 medical questions and 300 biography prompts, to support reliability?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper indicates that the newer, larger test sets show similar performance trends as the original smaller test sets?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "lKK50q2MtV", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "TokenFlow: Consistent Diffusion Features for Consistent Video Editing", "authors": ["Michal Geyer", "Omer Bar-Tal", "Shai Bagon", "Tali Dekel"], "abstract": "The generative AI revolution has recently expanded to videos. Nevertheless, current state-of-the-art video models are still lagging behind image models in terms of visual quality and user control over the generated content. In this work, we present a framework that harnesses the power of a text-to-image diffusion model for the task of text-driven video editing. Specifically, given a source video and a target text-prompt, our method generates a high-quality video that adheres to the target text, while preserving the spatial layout and motion of the input video. Our method is based on a key observation that consistency in the edited video can be obtained by enforcing consistency in the diffusion feature space. We achieve this by explicitly propagating diffusion features based on inter-frame correspondences, readily available in the model. Thus, our framework does not require any training or fine-tuning, and can work in conjunction with any off-the-shelf text-to-image editing method. We demonstrate state-of-the-art editing results on a variety of real-world videos.", "keywords": "Computer Vision Applications;Video Editing and Synthesis;Text-to-Image Diffusion Models;Text-driven Video Editing", "qa_pairs": [{"question": "Are self-attention features the only features replaced by features of neighboring frames, and have other features such as ResBlock features or attention masks been considered for replacement?", "answer": "Only self-attention tokens are replaced, as they capture fine-grained spatial information and are suitable for controlled image generation[1]. Residual block features are also overridden due to their connection with self-attention blocks, but attention masks are not replaced as cross-attention maps only capture rough layout and are unsuitable for fine-grained temporal correspondence[2].\n[1] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023. [2] Hertz, Amir, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. ICLR (2022).", "category": "Motivation_Analysis"}, {"question": "What is the comparison of the runtime of the method with other methods when evaluated on a video of 40 frames using 50 diffusion steps?", "answer": "The method's sampling time is faster than per-frame editing (PnP). The average runtime comparisons are as follows: TAV: 2684.70s, Text2video-zero: 198.52s, Rerender-a-video: 285.85s, fatezero: 349.07s, PnP: 208.43s, authors (preprocess): 50.66s, authors (sampling): 187.20s, authors (total): 237.87s.\n\n| TAV | Text2video-zero | Rerender-a-video | fatezero  | PnP | authors (preprocess) | authors (sampling) | authors (total)  |\n |---|---|---|---|---|---|---|---| \n| 2684.702563 | 198.5167881 | 285.848508 | 349.0715174 | 208.4318189 | 50.6622076 | 187.204452 | 237.8666596  |\n | 578.8067357 | 4.807203365 | 26.62864091 | 26.86996169 | 2.176882622 | 11.11061152 | 3.114096461 | 9.716412365 |", "category": "Method_Comparison"}, {"question": "What are the main differences between the proposed framework and prior methods, and what specific novel components or modifications are introduced in the proposed framework?", "answer": "The proposed framework introduces a method that performs keyframe editing and propagation directly in the feature space of the model during the generation process, which is a novel approach. Key novel components include TokenFlow propagation, which directly extracts and enforces inter-frame correspondence in the feature space, and a dedicated generation process that alternates between keyframe editing and feature propagation. This contrasts with prior methods that either rely solely on extended attention (which is insufficient for temporal consistency, as the paper thoroughly demonstrated in Sec. 5, (e.g. [1], [2]) ) or edit keyframes using optical flow estimation and propagate edits in RGB space(e.g., [3]), leading to long-range inconsistencies and artifacts.\n[1]Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv:2303.09535, 2023.\n\n[2] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. ArXiv, abs/2303.13439, 2023a.\n\n[3] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation, 2023.", "category": "Method_Comparison"}, {"question": "What is the overall runtime for generating a new video using TokenFlow, and how does it compare to the runtimes of other methods listed in Table 1?", "answer": "The overall runtime for TokenFlow to generate a new video of 40 frames using 50 diffusion steps is 237.87 seconds (preprocess: 50.66 seconds, sampling: 187.20 seconds). This is faster than per-frame editing (PnP) with a total runtime of 208.43 seconds and other methods such as TAV (2684.70 seconds), Text2video-zero (198.52 seconds), Rerender-a-video (285.85 seconds), and fatezero (349.07 seconds).\n| TAV | Text2video-zero | Rerender-a-video | fatezero  | PnP | ours (preprocess) | ours (sampling) | ours (total)  |\n|---|---|---|---|---|---|---|---|\n| 2684.702563 | 198.5167881 | 285.848508 | 349.0715174 | 208.4318189 | 50.6622076 | 187.204452 | 237.8666596  |\n| 578.8067357 | 4.807203365 | 26.62864091 | 26.86996169 | 2.176882622 | 11.11061152 | 3.114096461 | 9.716412365 |", "category": "Method_Comparison"}, {"question": "What is the effect of varying the interval size in keyframe sampling on runtime and temporal smoothness, and how does the random sampling scheme handle occlusion cases?", "answer": "Varying the interval size in keyframe sampling affects runtime and temporal smoothness. Using fewer keyframes decreases runtime and improves temporal consistency, but too few keyframes may result in inaccurate correspondences and artefacts. Empirical evaluation on a subset of videos showed that an interval size of 8 balances runtime and quality. Random sampling of keyframes is robust to occlusions, as it is unlikely for an object to be occluded in all sampled keyframes across timesteps. Occlusions were not observed to be an issue in experiments.\n| interval size | warp error (x 10^-3) | runtime |\n|---|---|---|\n| 4 | 2.3 | 1656 |\n| 8 | 2.2 | 656 |\n| 12 | 2.1 | 435 |\n| 16 | 2.0 | 359 |\n", "category": "Experimental_Exposition"}, {"question": "How does the proposed TokenFlow method address the challenges of motion-based or composition-based video editing, given its reliance on feature correspondences in the original video?", "answer": "The proposed method is designed to preserve the motion and object structure of the original video. To support video edits involving significant structure deviations or editing of object movements, a motion prior would be required to ensure consistent application across frames. For instance, changing a wolf to a rabbit necessitates synthesizing new motion for the ears, which is not present in the original video. Current image diffusion models do not provide such motion information. However, with advancements in text-to-video diffusion models, generative motion priors suitable for these tasks are anticipated to become available, potentially extending the method's capabilities.", "category": "Method_Mechanics"}, {"question": "Why is the temporal consistency metric measured by optical flow and warping considered valid despite its potential to over-penalize shape changes and favor color/texture changes?", "answer": "The method and competing methods (except TAV) are designed to preserve object structure, making this metric valid. Additionally, the paper includes a comprehensive user study to evaluate temporal consistency.", "category": "Experimental_Analysis"}, {"question": "How does the proposed method handle scenarios where the target of editing is partially occluded for a few frames in the video?", "answer": "The method is robust to occlusions due to random sampling of keyframes, making it unlikely for an object to be occluded in all sampled keyframes across timesteps. Empirical observations confirm that occlusions are not a significant issue.", "category": "Method_Mechanics"}, {"question": "Why does TokenFlow exhibit flickering in high-frequency patterns compared to baselines like Text2LIVE and Gen1, as observed in the 'Comparisons to Baselines' and 'Additional Qualitative Comparisons' sections?", "answer": "TokenFlow's flickering in high-frequency patterns is due to the inherent nature of the latent diffusion model, which uses a VAE with lossy compression. This results in the decoder hallucinating some high frequencies in the final RGB frames, a limitation also present when encoding-decoding natural videos without manipulations, as discussed in Section 6 of the paper.", "category": "Experimental_Analysis"}, {"question": "What is the maximum number of frames that TokenFlow can process on an A100 GPU, and are there potential optimizations to increase this limit?", "answer": "On an A100 GPU, TokenFlow can process approximately 300 frames. Potential optimizations, such as computing extended attention on random subsets of keyframes instead of all keyframes, could increase this limit.", "category": "Method_Mechanics"}, {"question": "What are the runtime comparisons for the proposed method and other baseline methods on a video of 40 frames using 50 diffusion steps?", "answer": "The table provided reports average runtime comparisons for various methods, including authors. The paper's sampling time is faster than that of per-frame editing (PnP). The runtime values for each method are as follows: TAV (2684.702563, 578.8067357), Text2video-zero (198.5167881, 4.807203365), Rerender-a-video (285.848508, 26.62864091), fatezero (349.0715174, 26.86996169), PnP (208.4318189, 2.176882622), authors (preprocess: 50.6622076, 11.11061152; sampling: 187.204452, 3.114096461; total: 237.8666596, 9.716412365).\n\n| TAV | Text2video-zero | Rerender-a-video | fatezero  | PnP | authors (preprocess) | authors (sampling) | authors (total)  |\n|---|---|---|---|---|---|---|---|\n| 2684.702563 | 198.5167881 | 285.848508 | 349.0715174 | 208.4318189 | 50.6622076 | 187.204452 | 237.8666596  |\n| 578.8067357 | 4.807203365 | 26.62864091 | 26.86996169 | 2.176882622 | 11.11061152 | 3.114096461 | 9.716412365 |", "category": "Method_Comparison"}, {"question": "Are the methods used in the study, including the competing methods (except TAV), designed to preserve the structure of objects, thereby justifying the use of the temporal consistency metric?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "JePfAI8fah", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "iTransformer: Inverted Transformers Are Effective for Time Series Forecasting", "authors": ["Yong Liu", "Tengge Hu", "Haoran Zhang", "Haixu Wu", "Shiyu Wang", "Lintao Ma", "Mingsheng Long"], "abstract": "The recent boom of linear forecasting models questions the ongoing passion for architectural modifications of Transformer-based forecasters. These forecasters leverage Transformers to model the global dependencies over temporal tokens of time series, with each token formed by multiple variates of the same timestamp. However, Transformers are challenged in forecasting series with larger lookback windows due to performance degradation and computation explosion. Besides, the embedding for each temporal token fuses multiple variates that represent potential delayed events and distinct physical measurements, which may fail in learning variate-centric representations and result in meaningless attention maps. In this work, we reflect on the competent duties of Transformer components and repurpose the Transformer architecture without any modification to the basic components. We propose iTransformer that simply applies the attention and feed-forward network on the inverted dimensions. Specifically, the time points of individual series are embedded into variate tokens which are utilized by the attention mechanism to capture multivariate correlations; meanwhile, the feed-forward network is applied for each variate token to learn nonlinear representations. The iTransformer model achieves state-of-the-art on challenging real-world datasets, which further empowers the Transformer family with promoted performance, generalization ability across different variates, and better utilization of arbitrary lookback windows, making it a nice alternative as the fundamental backbone of time series forecasting. Code is available at this repository: https://github.com/thuml/iTransformer.", "keywords": "Time Series Forecasting;Transformer", "qa_pairs": [{"question": "How effective is the proposed method when the dataset is not exchangeable, and what adjustments can be made to retain the test's validity?", "answer": "The effectiveness of the method depends on the source of non-exchangeability. In cases where non-exchangeability arises from ascending IDs (e.g., in SQuAD and HumanEval), the dataset can be adjusted by removing these IDs or permuting the examples while keeping IDs constant. These adjustments allow the test to remain valid.", "category": "Experimental_Analysis"}, {"question": "What evidence and explanation support the claim that iTransformer is better at handling fluctuating series variations compared to PatchTST?", "answer": "The evidence is provided in the PEMS prediction visualization in Appendix E.2 of the paper, which shows that PEMS has more fluctuating series variations compared to others. The explanation is that PatchTST may lose focus on specific locality to handle rapid fluctuations, as its patch length needs to be adaptively tuned for the series (e.g., frequency). In contrast, iTransformer embeds the whole series as a token to describe the underlying process globally.", "category": "Experimental_Analysis"}, {"question": "What steps were taken to verify the stability of iTransformer's performance and its significance compared to previous state-of-the-art methods like PatchTST and SCINet?", "answer": "Each experiment was repeated five times with different random seeds, and the standard deviations of iTransformer's performance in Table 5 of the paper showing stable performance. Significance tests were conducted to compare iTransformer with previous state-of-the-art methods (PatchTST and SCINet), including all subsets of ETT and PEMS datasets. The performance was averaged from four prediction lengths, each containing five runs of random seeds. The standard deviations and test statistics are provided, demonstrating that iTransformer's performance is on par with PatchTST and SCINet within the margin of error.\n\n| MSE     | iTransformer | Previous SOTA | test statistic |\n| ------- | ------------ | ------------- | -------------- |\n| ETT     | 0.383±0.001  | 0.381±0.001   | 2.124          |\n| Weather | 0.258±0.001  | 0.259±0.001   | -2.913         |\n| PEMS    | 0.119±0.001  | 0.121±0.002   | -0.419         |\n\nThe benchmark of time series forecasting has been excavated for a long time. And PatchTST and SCINet as the SOTA on specific datasets has surpassed concurrent works by a large margin. iTransformer can be on par with on their advantaged datasets while further go ahead on other challenging datasets (such as ECL and Traffic), and thus achieves the comprehensive SOTA.", "category": "Method_Comparison"}, {"question": "How does the efficiency of the proposed efficient training strategy compare to linear models in terms of training speed, memory footprint, and performance?", "answer": "Figure 10 in the paper comprehensively compares the training speed, memory footprint, and performance of iTransformer, iTransformer with the efficient training strategy, iTransformer with an efficient attention module, linear forecasters (DLinear and TiDE), and other Transformers (Transformer, PatchTST, and Crossformer). The observations are:\n1. iTransformer exceeds other Transformers in efficiency on datasets with a small number of variates (e.g., Weather with 21 variates). On datasets with numerous variates (e.g., Traffic with 862 variates), memory footprints are similar, but iTransformer trains faster.\n2. iTransformer achieves better performance on datasets with numerous variates due to explicit utilization of multivariate correlations.\n3. With an efficient attention module or the proposed efficient training strategy on partial variates, iTransformer achieves the same level of speed and memory footprint as linear forecasters.", "category": "Method_Comparison"}, {"question": "What are the reasons for the differences in the TiDE results between the original paper and the current study?", "answer": "The differences in the TiDE results are due to two main factors: 1) The lookback window size used in the original TiDE paper is 720, while the current study uses a lookback length of 96, following the unified long-term forecasting protocol of TimesNet. 2) The original TiDE paper trains the model with 100 epochs, whereas the current study reproduces results with 10 epochs. Despite these differences, the study ensures fairness by keeping other hyperparameters, such as hidden dimension and the number of layers, comparable. Additionally, when tested under the same lookback length as TiDE, iTransformer still achieves better results due to its ability to benefit from an enlarged lookback window.\n\n| MSE          | ECL       | ETTh2     | Exchange  | Traffic   | Weather   | Solar-Energy | PEMS03    |\n| ------------ | :-------- | :-------- | :-------- | :-------- | :-------- | :----------- | :-------- |\n| TiDE         | 0.160     | 0.309     | 0.092     | 0.445     | 0.171     | 0.303        | 0.216     |\n| iTransformer | **0.129** | **0.301** | **0.086** | **0.349** | **0.158** | **0.191**    | **0.071** |", "category": "Experimental_Analysis"}, {"question": "How is the train-validation-test split handled in the experiments to ensure there are no data leakage issues in time series forecasting?", "answer": "The train-validation-test split follows the same protocol as previous works, with datasets strictly divided according to chronological order to prevent data leakage. The detailed data processing and split protocols are provided in Appendix A.1 of the paper.", "category": "Experimental_Setup"}, {"question": " Does the paper prioritize the consideration of variate heterogeneity over temporal dependency in multi-time series forecasting (MTSF)?", "answer": "The paper considers both variate heterogeneity and temporal dependency crucial for good MTSF performance. However, the problem is that the heterogeneity of variates can hardly be considered in the vanilla Transformer. After embedding, the variates are projected into the channels of embedding. It ignores the problem of inconsistent physical measurements and can fail to maintain the independence of variates, let alone capture and utilize the multivariate correlation, which is essential for forecasting with numerous variates, and complicated systems driven by the latent physical process (such as meteorological systems). In addition, even if the FFN and layernorm seem simpler than attention blocks, they are efficient and competent in learning the temporal dependency of a series, which can be traced back to statistical forecasters such as ARIMA and Holt-Winter. They also have no problems with inconsistent measurements since they work on the time points of the same variate, and have an enlarged respective field as the whole lookback series can be embedded as the variate token.", "category": "Method_Disambiguation"}, {"question": "What explains the poor performance observed in the second row (both attention) of Table 3?", "answer": "The poor performance in the second row of Table 3, as well as in the third row for the Traffic dataset, is attributed to the application of the attention module to the temporal dimension. This approach assumes that time points of each timestamp are aligned, which is not always the case due to potential systematic delays in road occupancy data from sensors installed in different highway areas. This can lead the model to learn meaningless attention maps unless it has a larger receptive field to capture decay or causal processes, which is why we propose embedding the whole series as a variate token.", "category": "Experimental_Analysis"}, {"question": "What are the results of experiments conducted with 20% partial variates, and how do they compare to full-variate training? Specifically, how does taking the first 20% variates compare to taking the last 20% variates in terms of generalization error?", "answer": "The authors conducted 60 experiments with 20% partial variates, calculating the generalization error (Comp) as the increasing ratio of MSE from full-variate training (Full) to the average (Avg) of five runs. The results show that taking the first 20% variates results in slightly better transfer on ECL compared to the last 20% variates. However, the model still generalizes well on unseen variates with a small increase in forecasting error on Traffic and Solar Energy. \n\n\n|  MSE (iTransformer) | ECL    | Traffic | Solar Energy |\n|----------------|--------------------|--------|---------|\n| Full (0-100%)  | 0.162              | 0.443  | 0.202   |\n| 0-20%          | 0.192              | 0.444  | 0.214   |\n| 20%-40%        | 0.210              | 0.448  | 0.210   |\n| 40%-60%        | 0.210              | 0.448  | 0.242   |\n| 60%-80%        | 0.222              | 0.445  | 0.239   |\n| 80%-100%       | 0.222              | 0.445  | 0.225   |\n| Avg ± Std (Comp)| 0.211±0.011 (+30.2%)| 0.445±0.001 (+0.6%)| 0.226±0.013 (+11.8%)|\n\n| MSE (iInformer) | ECL    | Traffic | Solar Energy |\n|----------------|-----------------|--------|---------|\n| Full (0-100%)  | 0.169           | 0.459  | 0.200   |\n| 0-20%          | 0.204           | 0.466  | 0.233   |\n| 20%-40%        | 0.211           | 0.465  | 0.228   |\n| 40%-60%        | 0.218           | 0.465  | 0.237   |\n| 60%-80%        | 0.217           | 0.477  | 0.238   |\n| 80%-100%       | 0.215           | 0.468  | 0.230   |\n| Avg ± Std (Comp)| 0.213±0.005 (+26.0%)| 0.468±0.005 (+2.0%)| 0.233±0.004 (+16.4%)|\n\n| MSE (iReformer) | ECL    | Traffic | Solar Energy |\n|----------------|-----------------|--------|---------|\n| Full (0-100%)  | 0.170           | 0.459  | 0.198   |\n| 0-20%          | 0.190           | 0.463  | 0.204   |\n| 20%-40%        | 0.202           | 0.462  | 0.203   |\n| 40%-60%        | 0.203           | 0.463  | 0.207   |\n| 60%-80%        | 0.221           | 0.463  | 0.215   |\n| 80%-100%       | 0.202           | 0.466  | 0.208   |\n| Avg ± Std (Comp)| 0.204±0.010 (+20.0%)| 0.463±0.001 (+1.0%)| 0.207±0.004 (+4.8%) |\n\n| MSE (iFlowformer) | ECL    | Traffic | Solar Energy |\n|----------------|-------------------|--------|---------|\n| Full (0-100%)  | 0.168             | 0.449  | 0.204   |\n| 0-20%          | 0.191             | 0.456  | 0.214   |\n| 20%-40%        | 0.208             | 0.449  | 0.217   |\n| 40%-60%        | 0.234             | 0.454  | 0.229   |\n| 60%-80%        | 0.217             | 0.450  | 0.222   |\n| 80%-100%       | 0.234             | 0.448  | 0.229   |\n| Avg ± Std (Comp)| 0.216±0.016 (+28.5%)| 0.451±0.003 (+0.4%)| 0.222±0.006 (+8.6%) |\n\n", "category": "Experimental_Exposition"}, {"question": "How do the concepts of CD/CI and distribution shift relate to the methodology and performance of the iTransformer model in time series forecasting?", "answer": "It is essential to compare these architecture-free techniques addressing the time series properties and go further into bigger questions of how to consider time series variates and tokenize the time series. The paper newly added RLinear to the baseline, which is very competitive on several datasets, such as ETT and Weather. Regarding CI/CD and RevIN, it is important to discuss these techniques: Previous works have discussed the CI/CD tradeoff between model capacity and robustness. For example, CI can greatly increase the performance when data is scarce, while CD benefits from a higher capacity ideally. In the context of the paper’s work, making the variates independent is essential, especially when there are potential risks of delayed series and loosely organized measurements. Regardless of CI/CD, multivariate correlations cannot be explicitly utilized. In contrast to these works, iTransformer repurposes an architecture with the native Transformer modules to cope with it. RevIN or Series Stationarization has been widely applied to address distribution shift (or non-stationarity) for real-world time series. These techniques can boost the performance of both linear models and Transformers. There are Transformers in the baseline already equipped with this technique (PatchTST and Stationary), and the paper newly includes the linear one, RLinear. These works aim to model temporal dependencies. In iTransformer, temporal dependencies are modeled naturally by the FFN and layernorm, and still leaves further improvement for the paper to tackle the distribution shift. Compared with linear forecasters, iTransformer is good at forecasting high-dimensional time series, namely, numerous variates and complicated correlations. Under the univariate forecasting scenarios, iTransformer becomes in effect a stackable linear forecaster (FFN with the laynorm). For correlation between variates, iTransformer keeps the variate independent and the attention module handles this task.", "category": "Experimental_Analysis"}, {"question": "What is the significance of Figure 7 (b), and how should the reader interpret the lines in the figure? How does this demonstrate an improvement over the usual Transformer model?", "answer": "Figure 7 (b) illustrates multivariate correlations, where the x- and y-axis represent variates, and the attention map displays Pearson Correlations among them. The attention maps reflect correlations in the raw data series (visible when comparing subplots in the same column), enhancing the Transformer's interpretability. Additionally, the attention maps change to reflect correlations that vary between the lookback window and forecast window (visible when comparing subplots in the same row), demonstrating an improvement over the usual Transformer by providing clearer insights into data relationships.", "category": "Concept_Understanding"}, {"question": "What are the reasons for the discrepancy in the experimental results between the paper implementation of PatchTST and the published results, specifically for the Electricity dataset with a horizon of 96 where PatchTST reports an MSE of 0.129 and the paper's implementation reports 0.195?", "answer": "The discrepancy in results arises from several differences in experimental settings: 1) PatchTST uses a tunable lookback length of either 336 or 512, while the paper's implementation uses a unified length of 96, adhering to the TimesNet protocol. 2) PatchTST is trained for 100 epochs, whereas the paper trains all models for 10 epochs. 3) The PatchTST paper employs a carefully designed learning rate strategy, which differs from the paper's  approach.", "category": "Experimental_Analysis"}, {"question": "Did the authors conduct significance tests to compare iTransformer with previous SOTA models on the ETT, Weather, and PEMS datasets?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the standard deviations and test statistics provided show that iTransformer's performance is within the margin of error compared to previous SOTA models?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are standard deviations reported for the iTransformer performance in the paper to assess significance of differences?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "nnVO1PvbTv", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph", "authors": ["Jiashuo Sun", "Chengjin Xu", "Lumingyuan Tang", "Saizhuo Wang", "Chen Lin", "Yeyun Gong", "Lionel Ni", "Heung-Yeung Shum", "Jian Guo"], "abstract": "Although large language models (LLMs) have achieved significant success in various tasks, they often struggle with hallucination problems, especially in scenarios requiring deep and responsible reasoning. These issues could be partially addressed by introducing external knowledge graphs (KG) in LLM reasoning. In this paper, we propose a new LLM-KG integrating paradigm ``$\\hbox{LLM}\\otimes\\hbox{KG}$'' which treats the LLM as an agent to interactively explore related entities and relations on KGs and perform reasoning based on the retrieved knowledge. We further implement this paradigm by introducing a new approach called Think-on-Graph (ToG), in which the LLM agent iteratively executes beam search on KG, discovers the most promising reasoning paths, and returns the most likely reasoning results. We use a number of well-designed experiments to examine and illustrate the following advantages of ToG: 1) compared with LLMs, ToG has better deep reasoning power; 2) ToG has the ability of knowledge traceability and knowledge correctability by leveraging LLMs reasoning and expert feedback; 3) ToG provides a flexible plug-and-play framework for different LLMs, KGs and prompting strategies without any additional training cost; 4) the performance of ToG with small LLM models could exceed large LLM such as GPT-4 in certain scenarios and this reduces the cost of LLM deployment and application. As a training-free method with lower computational cost and better generality, ToG achieves overall SOTA in 6 out of 9 datasets where most previous SOTAs rely on additional training.", "keywords": "Knowledge Graph;Chain-of-Thought;Large Language Models", "qa_pairs": [{"question": "What are the advantages of the ToG method despite its reliance on strong foundation LLMs, as observed in Table 2?", "answer": "The performance of ToG relies on the foundation LLM since the LLM plays an important role in the pruning and reasoning steps of ToG. However, ToG has particular advantages: \n(1) ToG with small and poor foundation LLM has an opportunity to perform better than large and powerful LLM-only model, as shown in Table 2 of the paper. It can be seen that, on dataset CWQ, ToG-R with LLama2-70B-Chat reaches an accuracy 57.6%, higher than the accuracy of GPT-4 (Chain-of-thought prompting) 46.0%. Similarly, on dataset WebQSP, ToG-R with LLama2-70B-Chat reaches 68.9%, higher than GPT-4 (Chain-of-thought prompting) 67.3%. This result exhibits that ToG has a significant advantage in low cost deployment. Specifically, ToG with a cheaper and faster small LLM may be good enough in many scenarios, rather than deploying expensive and slow large LLM. \n(2) Although ToG relies on the foundation LLM, its performance also benefits from the contribution of knowledge graph, as well as the interaction of LLM and KG. We can see from Table 2 that, for any specific LLM, ToG always outperforms the corresponding LLM-only model. E.g., on dataset CWQ, ToG with ChatGPT reaches 58.9%, and ChatGPT-only (CoT) reaches 38.8%, so there is a 20.1% accuracy gap. It can be claimed that, ToG has the flexibility of plug-and-play any LLMs and KGs, and this algorithm usually improves the performance of the original LLM by incorporating the structural knowledge from KGs without adding any additional training cost (training-free). (3) All prior SOTA methods belong to training-based methods, which have a natural advantage in prediction accuracy on specific task or specific dataset compared with prompting-based methods (ToG, CoT etc.). However, ToG, a prompting-based method, outperforms those training-based prior SOTA methods on 6 out of 9 datasets. Both LLM and KG contribute to the performance of ToG (from Table 2, we see GPT-4 only is not as good as prior SOTAs without the help of KG). Moreover, prompting-based methods have their own particular advantages, such as avoidance of additional training cost and better explainability.", "category": "Experimental_Analysis"}, {"question": "What are the accuracy results of the self-ask[1] and retrieval augmentation methods[2] compared to ToG on the WebQSP dataset?\n\n[1] Press et al. 2022. Measuring and Narrowing the Compositionality Gap in Language Models. https://arxiv.org/abs/2210.03350 \n[2] Kasai et al. 2022. RealTime QA: What's the Answer Right Now?https://arxiv.org/abs/2207.13332", "answer": "The accuracy of the self-ask method is 50.8%, and the accuracy of Retrieval Augmentation with Wikipage is 61.6% on the WebQSP dataset. These results are based on the latest GPT-3.5-turbo model. In comparison, ToG achieves an accuracy of 76.2%, and ToG-R achieves 75.8%. \n\n|Prompting Methods w. ChatGPT | Accuracy on WebQSP| \n|--- |---| \n|CoT |62.2| \n|Self-ask|50.8| |\nRetrieval Augmentation with Wikipage| 61.6| \n|ToG|76.2| |ToG-R|75.8|", "category": "Experimental_Exposition"}, {"question": "What measures have been taken to address the computational efficiency concerns regarding the beam search and pruning process in the Think-on-Graph algorithm?", "answer": "The authors came up with new ideas to reduce the computational complexity of Think-on-Graph (proportional to the number of calling LLMs) from the original $O(ND)$ to $O(D)$, where $D$ is the depth (or equivalently length) of the reasoning path, and $N$ is the width of the beam-search (how many paths are remained in the pool in each iteration).\n● Solution 1: Reducing computational complexity from $O(ND)$ to $O(D)$ by using a lightweight model in pruning.\n● Solution 2: Reducing computational complexity from $O(ND)$ to $O(D)$ by unifying the prompts in the same pruning step.\n● Solution 3: Optimizing the pruning step to make the actual calls of LLMs much less than the previously estimated $2ND+D+1$ and closer to some common prompting methods such as CoT-SC.\n**Details of Solution 1**: The bottleneck of computation is the pruning step, which contributes to $N*D$ times calling, and it is important to optimize it for computational efficiency. A technical route is to replace LLM with small models such as BM25 and Sentence-BERT in the pruning step since the small models are much faster than LLM calling. In this way, the number of LLM calling can be reduced from $2ND+D+1$ to $D+1$. When $D$=3, for example, there are only 4 times LLM calling. However, this optimization sacrifices accuracy due to the weaker scoring model in pruning. For example, as shown in Table 5 of the paper, the performance of ToG on WebQSP drops from 76.2% to 66.3% after replacing ChatGPT with SentenceBERT for pruning. To alleviate the issue of performance degradation, the search width can be appropriately increased to compensate for the loss because increasing search width can improve the chance of the optimal path being selected in the pool and it doesn't affect the number of LLM calling. To empirically verify this, the authors increase the search width from 3 to 5 and reevaluate ToG with SentenceBERT as the pruning model on WebQSP. The accuracy rises from 66.3% to 68.5% and could be further improved with a greater width since the greater width would not cause an increase in the number of LLM calls.\n**Details of Solution 2**: Another solution to speeding up the pruning step is employing the LLM at once to score all components of N candidate sets for obtaining top-N candidates, instead of calling the LLM N times to score N candidate sets separately. Through this solution, either entity pruning step or relation pruning step only needs 1 LLM call for each iteration. Thus, the maximum number of LLM calls per question needed for ToG and ToG-R would drop to $2D+D+1$ and $D+D+1$.\n**Other Discussion About Efficiency**: As a completely training-free method, ToG actually saves a lot of computational time training models (compared to pre-training or fine-tuning methods).\n", "category": "Method_Mechanics"}, {"question": "Why does ToG not consistently outperform the 'prior finetune SOTA' in some cases, particularly on single-hop reasoning datasets like SimpleQuestion and T-REx, and how does its performance compare on multi-hop reasoning tasks?", "answer": "ToG does not outperform the 'prior finetune SOTA' on single-hop reasoning datasets like SimpleQuestion and T-REx because its strength lies in deep reasoning, which is more suited for multi-hop tasks. On single-hop datasets, the long reasoning paths used by ToG are unnecessary. However, ToG significantly outperforms the prior FT SOTA on Zero-Shot RE (88.3% vs. 74.6%)which is also a single-hop dataset, indicating it can perform well on single-hop tasks. For multi-hop tasks, ToG-R initially underperformed on CWQ (69.5% vs. 70.4%) due to the paper simply sets search width $N=3$ and search depth $D=3$ without any comprehensive hyper-parameter tuning. After tuning with larger search width ($N=4$) and depth ($D=4$), ToG-R's accuracy improved to 72.5%, exceeding the prior FT SOTA's 70.4%. ", "category": "Experimental_Analysis"}, {"question": "What specific aspect of the hallucination problem does ToG aim to address, and how is its effectiveness demonstrated in the paper?", "answer": "ToG aims to address the accurate deep-reasoning problem, a subset of the hallucination problem that accounts for a significant portion of hallucinations in LLM reasoning. The effectiveness of ToG is demonstrated through significant accuracy improvements in deep-reasoning problems, as shown in Table 1, Table 2, and additional tables in the appendix, when compared with LLM-only methods like CoT.", "category": "Method_Disambiguation"}, {"question": "What is the current running-time efficiency of the Think-on-Graph (ToG) algorithm, and what improvements have been proposed to enhance its efficiency?", "answer": "The current running-time efficiency of the Think-on-Graph (proportional to the number of calling LLMs) algorithm has been improved by reducing its computational complexity from $O(ND)$ to $O(D)$, where $D$ is the depth of the reasoning path (or equivalently length) and $N$ is the width of the beam-search(how many paths are remained in the pool in each iteration). The proposed improvements include:\nSolution 1: Reducing computational complexity from $O(ND)$ to $O(D)$ by using lightweight model in pruning. \nSolution 2: Reducing computational complexity from $O(ND)$ to $O(D)$ by unifying the prompts in the same pruning step. \nSolution 3: Optimizing pruning step to make the actual calls of LLMs much less than the previously estimated $2ND+D+1$ and closer to some common prompting methods such as CoT-SC.", "category": "Method_Mechanics"}, {"question": "How do the knowledge graphs (KGs) used in the experiment address the limitation of out-of-date knowledge compared to large language models (LLMs)?", "answer": "Firstly, the knowledge in KG is much easier to be updated than the knowledge in LLM. Specifically, the knowledge in KGs is stored on entities and relations (in the form of triples) which are usually physically stored on graph databases, and can be added, deleted and updated immediately. On the contrary, updating knowledge in LLM needs training with new samples which is usually time-consuming. Secondly, in practical scenarios, many KGs are updated frequently or even in real-time by adding new knowledge and deleting or revising out-of-date knowledge. The knowledge updating process can be automatically performed by a knowledge extraction system or manually done with crowdsourcing. For example, the knowledge stored in Wikidata is extracted automatically from Wikipedia and updated regularly according to the changes in Wikipedia content edited by human. Therefore, the knowledge stored in KGs such as Wikidata often has better timeliness than the inner knowledge of LLMs. Thirdly, the version of Wikidata used in this manuscript was released in Jan. 2023 and the latest version was released in Nov. 2023 where the newest knowledge is updated to Wikipedia. In comparison, the cut-off date of the knowledge of ChatGPT is September 2021, about two years ago.", "category": "Method_Comparison"}, {"question": "What solutions have been proposed to improve ToG's efficiency?", "answer": "The authors propose several solutions to improve efficiency, including reducing the computational complexity from $O(ND)$ to $O(D)$, where $D$ is the depth of the reasoning path(or equivalently length) and $N$ is the width of the beam-search(how many paths are remained in the pool in each iteration). Additionally, lightweight models are suggested for pruning, and prompts are unified in the same pruning step to further reduce costs.Some of these solutions have been empirically verified.", "category": "Method_Mechanics"}, {"question": "Can ToG effectively understand and generalize to diverse relation formats in different knowledge graphs (KGs), such as the hierarchical format in Freebase and the clearer semantics in Yago?", "answer": "Yes, ToG can effectively understand and generalize to diverse relation formats in different KGs. Conceptually, relation names in Freebase, though hierarchical, consist of words and phrases understandable by LLMs (e.g., people.person.occupation, sports.mascot.team, the last word shows the meaning of relation). In Wikidata, most relations are similar to natural language (e.g., instance of, part of, founded by). Empirically, experimental results show that LLMs and light-weight pruning models perform well on both Freebase and Wikidata.", "category": "Experimental_Analysis"}, {"question": "Can ToG or its variants solve complex reasoning problems such as counting, intersection, union, and negation operations, as exemplified by questions like 'How many TV programs has Bob Boyett created?' in GrailQA?", "answer": "Although vanilla ToG may not solve these complex reasoning problems, a slight variant of ToG can solve all these complex problems mentioned above. Specifically, swapping the order of pruning and reasoning in each iteration of vanilla ToG results in a new ToG algorithm (called new-ToG) that can solve all the above complex reasoning problems. For example, when new-ToG answers the question 'How many TV programs has Bob Boyett created?', new-ToG searches all neighbors of 'Bob Boyett' related to 'TV programs' and generates $M$ reasoning paths: {Bob Boyett-producer-program1, Bob Boyett-producer-program2, Bob Boyett-producer-program3... Bob Boyett-producer-programM}. LLMs can 'count' the number of neighbors during the reasoning step and output the answer $M$. Vanilla ToG can't finish this task because these neighbors are pruned in each iteration and only top $N$ (width of beam search) are kept in next iteration, so it can't count all neighbors. For intersection questions, such as 'How many TV programs have Bob Boyett and somebody created together?' According to Bob Boyett - productor - {program1, program2, program3... programN} and someone - productor - {program1, program2, program3... programM}, LLM reasoner can output the count of intersections. Similarly, union and negation operations can be implemented by LLM reasoner if a complete neighbor list (before pruning) can be provided.", "category": "Experimental_Analysis"}, {"question": "What is the performance of ToG when there is a knowledge conflict between external KG knowledge and parametric knowledge stored in LLMs?", "answer": "The authors conducted an experiment on the CWQ dataset to analyze ToG's performance when LLM (GPT-4) and KG outputs conflict. When CoT with GPT-4's answer conflicts with KG's answer, the accuracy of ToG with GPT-4 is 43.4%, which is a 13.7% reduction from its baseline performance of 57.1% on CWQ. This demonstrates the impact of knowledge conflicts on ToG's robustness.", "category": "Experimental_Exposition"}, {"question": "What are the experiment results when testing the performance of CoT and ToG prompting with small LLMs (LLaMA2-7B and LLaMA2-13B) on WebQSP, and what factors contribute to their performance?", "answer": "The results of the performance using LLaMA2-7b and LLaMA2-13b, two very small LLMs, on WebQSP are summarized below:\n\n| Method | LLaMA2-7b | LLaMA2-13b | \n| --- | --- | --- | \n| CoT | 7.1%| 8.6% | \n| ToG | 12.6% | 25.1% | \n\nThe experiment results show that CoT prompting with LLaMA2-7B and LLaMA2-13B achieved 7.1% and 8.6% accuracy, respectively, while ToG achieved 12.6% and 25.1% accuracy, respectively, on WebQSP. ToG performs better than CoT with these two very small LLMs. However, the actual accuracies for CoT and ToG are indeed far below from the authors expectation. The main factor of the bad performance is: since both LLaMA2-7B and LLaMA2-13B are base models which have not been trained with supervised fine-tuning and RLHF, the format of their outputs are really difficult to manage and difficult to exactly match the format required for a 'correct' answer. For example, the requirement that answers from LLMs be placed between { and } is often violated by these two small LLMs. This makes it difficult for the evaluator to locate the answer, leading to answers being misclassified as ‘wrong’. In another example, when asked to prune and return the top 3 relations from candidates (A, B, C, D, E), LLaMA2-7B and LLaMA2-13B sometimes return irrelevant relations, such as W, S, T, etc. In some cases, the LLaMA2 models restate the input questions or repeat text over and over again without responding to the question.", "category": "Experimental_Analysis"}]}
{"id": "2msbbX3ydD", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "Ferret: Refer and Ground Anything Anywhere at Any Granularity", "authors": ["Haoxuan You", "Haotian Zhang", "Zhe Gan", "Xianzhi Du", "Bowen Zhang", "Zirui Wang", "Liangliang Cao", "Shih-Fu Chang", "Yinfei Yang"], "abstract": "We introduce Ferret, a new Multimodal Large Language Model (MLLM) capable of understanding spatial referring of any shape or granularity within an image and accurately grounding open-vocabulary descriptions. To unify referring and grounding in the LLM paradigm, Ferret employs a novel and powerful hybrid region representation that integrates discrete coordinates and continuous features jointly to represent a region in the image. To extract the continuous features of versatile regions,  we propose a spatial-aware visual sampler, adept at handling varying sparsity across different shapes. Consequently, Ferret can accept diverse region inputs, such as points, bounding boxes, and free-form shapes. To bolster the desired capability of Ferret, we curate GRIT, a comprehensive refer-and-ground instruction tuning dataset including 1.1M samples that contain rich hierarchical spatial knowledge, with an additional 130K hard negative data to promote model robustness. The resulting model not only achieves superior performance in classical referring and grounding tasks, but also greatly outperforms existing MLLMs in region-based and localization-demanded multimodal chatting. Our evaluations also reveal a significantly improved capability of describing image details and a remarkable alleviation in object hallucination.", "keywords": "Ferret;Multimodal Large Language Model;Referring;Grounding", "qa_pairs": [{"question": "What are the unique aspects and contributions of the paper's work in jointly solving referring and grounding tasks, especially in comparison to concurrent works (Chen et al., 2023a)[1] and (Peng et al., 2023)[2]?\n\n[1]Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Positionenhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023a.\n\n[2]Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu\nWei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint\narXiv:2306.14824, 2023.", "answer": "The paper's work uniquely demonstrates the mutual benefits of referring and grounding through detailed ablation studies in Table 5. Additionally, the detailed comparison table highlighting distinctions such as input types, output grounding, data construction, and quantitative evaluation of referring/grounding with chat is shown below:\n\n| Model       | Input Types (Point) | Input Types (Box) | Input Types (Free-form) | Output Grounding | Data Construction (Convention) | Data Construction (GPT-Generate) | Data Construction (Robustness) | Quantitative Eval. of Refer/Ground w. Chat |\n|-------------|---------------------|-------------------|-------------------------|------------------|-------------------------------|--------------------------------|-----------------------------|--------------------------------------------|\n| BuboGPT     | ❌                  | ❌                | ❌                      | ✅                | ✅                            | ❌                            | ❌                          | ❌                                         |\n| Vision-LLM  | ❌                  | ❌                | ❌                      | ✅                | ✅                            | ❌                            | ❌                          | ❌                                         |\n| Kosmos-2    | ❌                  | ✅                | ❌                      | ✅                | ✅                            | ❌                            | ❌                          | ❌                                         |\n| Shikra      | ✅                  | ✅                | ❌                      | ✅                | ✅                            | ✅                            | ❌                          | ❌                                         |\n| GPT4-ROI    | ❌                  | ✅                | ❌                      | ❌                | ✅                            | ❌                            | ❌                          | ❌                                         |\n| PVIT        | ❌                  | ✅                | ❌                      | ❌                | ✅                            | ✅                            | ❌                          | ✅                                         |\n| **Ferret**  | ✅                  | ✅                | ✅                      | ✅                | ✅                            | ✅                            | ✅                          | ✅                                         |\n\n\n*Note: `Convention` refers to a comprehensive collection of publicly available data that has been transformed using templates, `GPT-Generate` signifies the generated refer/ground datasets employing GPT, and `Robustness` denotes datasets aimed at mitigating hallucination and improving robustness.*", "category": "Method_Comparison"}, {"question": "What are the limitations and potential failure scenarios of the proposed method?", "answer": "Failure Scenarios:: (1) Referring to too many objects (more than 3) in one question might not be as accurate as referring to each of them in separate conversations due to a relative scarcity of training data that mentions too many objects. (2) The referring and grounding of very small objects is less accurate than large or medium objects, which is a common challenge in object detection, though improving input image resolution could help. \n\nLimitations: (1). Not good at other languages because the training dataset is curated only in English. Although Ferret shows some emergent referring and grounding capability in other languages, its performance in other languages is still worse than in English. Future incorporation of multilingual training data could potentially mitigate this. (2). Similar to many large language models, Ferret has the potential to generate harmful or factually incorrect responses. (3). Ferret is not designed for segmentation tasks requiring mask outputs.", "category": "Method_Disambiguation"}, {"question": "How does the performance of the Spatial-Aware Visual Sampler compare to the Visual Sampler in SEEM (Zou et al., 2023)[1] across different types of regions, as shown in the ablation study in Tab. 6?\n\n[1]Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023.\n", "answer": "The Spatial-Aware Visual Sampler shows a modest improvement of 0.6% in point referring tasks for small and simple regions like circles. For larger and more complex shapes such as boxes or free-form regions, it demonstrates more significant improvements, ranging from 1% to 2%.  The above observation makes sense, as when the region is small, simple average pooling in SEEM might already be enough. Although the enhancements observed in these smaller regions seem modest, the paper's approach demonstrates substantial benefits when addressing the intricacies of larger regions.", "category": "Method_Comparison"}, {"question": "Why does the proposed method significantly outperform Shikra on Ferret-Bench but not as much on existing datasets?", "answer": "The proposed method, Ferret, outperforms Shikra more significantly on Ferret-Bench due to its superior performance in referring-related tasks compared to grounding tasks, which aligns with its performance on existing datasets. The larger difference is attributed to Ferret's better integration of referring and grounding with semantics, knowledge, and reasoning in a global context. Factors contributing to this include GPT-4 data, hierarchical data collection, and model design. Additionally, the performance of existing benchmarks may be close to saturation, whereas Ferret-Bench covers more complex scenarios, making it a more challenging and relevant benchmark for future research.", "category": "Experimental_Analysis"}, {"question": "How should the observed effect in Table 5, where the grounding task appears to benefit the referring task more than vice versa, be interpreted?", "answer": "The absolute values of performance gains on different datasets cannot be treated in the same way. On the one hand, referring and grounding tasks are measured in different metrics. Same as common practice, referring is evaluated by accuracy, while Flickr30k grounding is evaluated by Recall@1. On the other hand, in some datasets, such as Flickr30k, the baselines can already achieve high scores, and thus, further enhancement appears relatively modest in terms of values. For a straightforward example, improving 10% on top of 80% is much more difficult than over 50%. \nDifference in the amount of training data. The paper's GRIT dataset contains a larger quantity of grounding data compared to referring data. This imbalance in data volume may contribute to the observed different benefits for each task type.", "category": "Experimental_Analysis"}, {"question": "Why does Ferret outperform other models in Table 7, given that the Ferret-Bench is via GPT4 as a judge, but some of the data are collected from GPT4?", "answer": "All other models in Tab. 7 also use data collected from ChatGPT/GPT-4, such as LLaVA's instruction data and Shikra's instruction data. Since the tasks in Ferret-Bench (Tab. 7) require a joint capability of accurate referring, grounding, and correct reasoning, it is hypothesized that  Ferret outperforms other models in Table 7 due to two main reasons: (1) Ferret has stronger basic referring and grounding capabilities, as demonstrated in Tables 1, 2, and 3, which provide a better foundation for the tasks in Ferret-Bench. (2) Ferret's GPT-generated data is of higher quality, as it is conditioned on more fine-grained visual symbolic information (e.g., relationships and region descriptions) and is prompted to be more aware of joint referring/grounding and reasoning, making it better suited for complex scenarios.", "category": "Experimental_Analysis"}, {"question": "How does ChatGPT process image input when used for dataset collection, given that it is a language model?", "answer": "In LLaVA's instruction data collection, ChatGPT processes image input by receiving ground-truth bounding boxes (with coordinates) and five ground-truth captions as textual representations of the image. This textual representation, although not covering all image details, is sufficient for GPT to generate precise instruction-tuning data by leveraging its commonsense knowledge in the language domain. This method has proven to be effective in LLaVA and is widely used in recent MLLMs. In GRIT data collection, besides following LLaVA's setup, additional fine-grained information, such as physical relationships between objects and region captions with coordinates, is provided to enhance the quality of the generated data. ", "category": "Method_Mechanics"}, {"question": "Why is the hierarchical nature of the dataset considered practical, and how can custom datasets without a hierarchical structure be effectively used with the Ferret model?", "answer": "The hierarchical nature of the dataset reflects the inherent complexity in the real world and is fundamental in computer vision. It begins with basic elements such as object detection [1] and segmentation [2], progresses to understanding relationships between these objects[3], and culminates in the comprehension of complex scenes through scene graphs[4]. The paper's model, Ferret, is designed to leverage this hierarchical structure, thereby enhancing its robustness and adaptability across various contexts. For finetuning using custom datasets, it's not necessary for these datasets to have a hierarchical structure. People have the flexibility to fine-tune the model with their preferred data. Two viable strategies are as below: (1). fine-tuning the paper’s pre-trained Ferret model with one’s custom dataset; or (2) incorporating one’s dataset into the initial training phase alongside the paper’s existing data.\n\n[1] Girshick, Ross. 'Fast r-cnn.' Proceedings of the IEEE international conference on computer vision. 2015. \n\n[2] He, Kaiming, et al. 'Mask r-cnn.' Proceedings of the IEEE international conference on computer vision. 2017. \n\n[3] Lu C, Krishna R, Bernstein M, Fei-Fei L. Visual relationship detection with language priors. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14 2016 (pp. 852-869). Springer International Publishing. \n\n[4] Xu D, Zhu Y, Choy CB, Fei-Fei L. Scene graph generation by iterative message passing. InProceedings of the IEEE conference on computer vision and pattern recognition 2017 (pp. 5410-5419).", "category": "Experimental_Analysis"}, {"question": "What aspects of the paper contribute to its scientific novelty beyond the engineering aspects?", "answer": "The Spatial-aware Visual Sampler and the hybrid region representation are novel solutions for handling arbitrary referring regions in MLLM. Additionally, the preparation of hierarchical and robust refer-and-ground instruction-tuning data and the introduction of Ferret-Bench as a new and advanced benchmark for this domain also contribute to the scientific novelty of the work. Scientific novelty is not limited to modeling but also includes data, training recipes, and evaluation.", "category": "Method_Disambiguation"}, {"question": "What are the results of the ablation experiment on hybrid region representation for LVIS Referring Object Classification, and how does it compare to the region-only and coordinate-only methods?", "answer": "The ablation experiment on hybrid region representation for LVIS Referring Object Classification shows the following results: \n\n| Region                       | Point | Box  | Free-Form |\n|------------------------------|-------|------|-----------|\n| Hybrid Region Representation | 65.5  | 76.2 | 69.0      |\n| Coordinate Only              | 64.6  | 74.5 | -         |\n| Region Only                  | 60.4  | 70.1 | 64.2      |\n\nCompared to the region-only method, the coordinate-only method performs better in point and box referring but cannot handle free-form shapes. The hybrid region representation further boosts performance across all three types of referring.", "category": "Experimental_Exposition"}, {"question": "Was the evaluation on Flickr30k grounded caption conducted under fine-tuned or zero-shot conditions?", "answer": "The evaluation on Flickr30k grounded caption is not considered zero-shot because approximately 30% of the Flickr30k training data was included in the GRIT dataset, which guided the model to emulate the output style. Additionally, Ferret outperforms other state-of-the-art methods that use the entire Flickr30k training set.", "category": "Experimental_Setup"}, {"question": "Does the method exhibit lower accuracy in grounding very small objects compared to large or medium objects?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the method less accurate when referring to more than three objects in one question due to limited training data?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the evaluation on Flickr30k grounded caption performed in a zero-shot setting?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the method designed for segmentation tasks requiring mask outputs?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the model used for Flickr30k grounded caption evaluation see around 30% of the Flickr30k training data in the GRIT dataset?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do other state-of-the-art methods mentioned in the paper use the entire Flickr30k training set for grounded caption evaluation?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the Ferret model outperform other state-of-the-art methods on Flickr30k grounded caption despite using less training data?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "1tZbq88f27", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models", "authors": ["Deyao Zhu", "Jun Chen", "Xiaoqian Shen", "Xiang Li", "Mohamed Elhoseiny"], "abstract": "The recent GPT-4 has demonstrated extraordinary multi-modal abilities, such as directly generating websites from handwritten text and identifying humorous elements within images. These features are rarely observed in previous vision-language models. However, the technical details behind GPT-4 continue to remain undisclosed.\nWe believe that the enhanced multi-modal generation capabilities of GPT-4 stem from the utilization of sophisticated large language models (LLM). \nTo examine this phenomenon, we present MiniGPT-4, which aligns a frozen visual encoder with a frozen advanced LLM, Vicuna, using one projection layer. \nOur work, for the first time, uncovers that properly aligning the visual features with an advanced large language model can possess numerous advanced multi-modal abilities demonstrated by GPT-4, \nsuch as detailed image description generation and website creation from hand-drawn drafts.\nFurthermore, we also observe other emerging capabilities in MiniGPT-4, including writing stories and poems inspired by given images, teaching users how to cook based on food photos, and so on. \nIn our experiment, we found that the model trained on short image caption pairs could produce unnatural language outputs (e.g., repetition and fragmentation). To address this problem, we curate a detailed image description dataset in the second stage to finetune the model, which consequently improves the model's generation reliability and overall usability.", "keywords": "large language models;vision language models", "qa_pairs": [{"question": "Why does the performance drop significantly when three projection layers are used instead of one?", "answer": "The three-layer variant almost triples the learning capacity (the number of learnable parameters), which may require more training data compared to the original one-layer version. Training this variant on the same amount of data as the original version might not be sufficient to achieve comparable performance.", "category": "Experimental_Analysis"}, {"question": "What are the maximum input and output lengths for the training and inference stages, and how are they determined?", "answer": "The context length is limited by the LLM itself. For the Vicuna version based on Llama 1, the context length is 2048 tokens (input + output). During training, the output is truncated to a maximum of 160 tokens if it exceeds this limit to prevent GPU memory exhaustion.", "category": "Experimental_Setup"}, {"question": "How does MiniGPT-4's performance in the MMBench benchmark[1] compare to other contemporary models, particularly InstructBlip, in various visual reasoning tasks?\n\n[1] MMBench: Is Your Multi-modal Model an All-around Player?", "answer": "In the MMBench benchmark focusing on visual reasoning, MiniGPT-4 demonstrates competitive performance compared to InstructBlip. It surpasses InstructBlip in logical reasoning (LR), fine-grained perception for single instance (FP-S), and fine-grained perception across instances (FP-C). Additionally, MiniGPT-4 achieves competitive results in relation reasoning (RR), attribute reasoning (AR), and coarse perception (CP). Results from MMBenchmark are listed below.\n\n| Model          | Overall |  LR  |  AR  |  RR  | FP-S | FP-C |  CP  |\n|----------------|:-------:|:----:|:----:|:----:|:----:|:----:|:----:|\n| OpenFlamingo   |  4.6%   | 6.7% | 8.0% | 0.0% | 6.7% | 2.8% | 2.0% |\n| VisualGLM      |  38.1%  | 10.8%| 44.3%| **35.7%**| 43.8%| 23.4%| 47.3%|\n| LLaVa          |  38.7%  | 16.7%| 48.3%| 30.4%| 45.5%| 32.4%| 40.6% |\n| InstructBlip   |  **44%**   | 19.1%| **54.2%**| 34.8%| 47.8%| 24.8%| **56.4%**|\n| MiniGPT-4      |  42.3%  | **20.8%**| 50.7%| 30.4%| **49.5%**| **26.2%**| 50.7%", "category": "Method_Comparison"}, {"question": "How does MiniGPT-4 process the output vectors corresponding to the learned queries?", "answer": "MiniGPT-4 applies a projection layer to the Q-former’s output vectors (tokens) corresponding to the Q-former’s internal learnable queries. This layer individually converts each output token to match the input size required by the LLM model.", "category": "Method_Mechanics"}, {"question": "How does MiniGPT-4's performance in the MMBench benchmark compare to other contemporary models, particularly in areas such as logical reasoning (LR), fine-grained perception (FP-S and FP-C), relation reasoning (RR), attribute reasoning (AR), and coarse perception (CP)?", "answer": "In the MMBench benchmark, MiniGPT-4 demonstrates competitive performance compared to contemporary models like InstructBlip. It surpasses InstructBlip in logical reasoning (LR), fine-grained perception for single instance (FP-S), and fine-grained perception across instances (FP-C). Additionally, MiniGPT-4 achieves competitive results in relation reasoning (RR), attribute reasoning (AR), and coarse perception (CP). Results from MMBenchmark are listed below.\n\n| Model          | Overall |  LR  |  AR  |  RR  | FP-S | FP-C |  CP  |\n|----------------|:-------:|:----:|:----:|:----:|:----:|:----:|:----:|\n| OpenFlamingo   |  4.6%   | 6.7% | 8.0% | 0.0% | 6.7% | 2.8% | 2.0% |\n| VisualGLM      |  38.1%  | 10.8%| 44.3%| **35.7%**| 43.8%| 23.4%| 47.3%|\n| LLaVa          |  38.7%  | 16.7%| 48.3%| 30.4%| 45.5%| 32.4%| 40.6% |\n| InstructBlip   |  **44%**   | 19.1%| **54.2%**| 34.8%| 47.8%| 24.8%| **56.4%**|\n| MiniGPT-4      |  42.3%  | **20.8%**| 50.7%| 30.4%| **49.5%**| **26.2%**| 50.7%|", "category": "Experimental_Exposition"}, {"question": "What criteria were used to select the Advanced Abilities tasks, and do the images used for evaluation in these tasks overlap with the dataset used in stage 1 pre-training and stage 2 fine-tuning?", "answer": "The Advanced Abilities tasks were selected based on their requirement for understanding visual content and complex language generation. The images for evaluation were gathered using Google Image Search with keywords like 'funny memes' and 'photos of delicious food.' The training set consists only of image description texts and does not include text data for advanced abilities. MiniGPT4 was not trained with image-text paired data for these skills but gains them compositionally by mapping images to the embedding space of the LLM(frozen by default).", "category": "Experimental_Setup"}, {"question": "Which version of ChatGPT was used for answer revision in the experiments?", "answer": " In Table 2, the paper focuses not on the correctness of the generated captions but on their information coverage. A high score indicates that the generated descriptions encompass more information from the image, rather than 'more correct.' For this evaluation, the paper employed the API gpt-3.5-turbo. Traditional metrics such as BLEU and Rouge, which rely on sentence similarity, fall short when the format of evaluated image descriptions diverges from the ground truth. This discrepancy is particularly noticeable when the generated descriptions are significantly longer than the original ground truth captions. The paper designed a new metric based on GPT-4 that takes all the ground truth captions into account. ", "category": "Experimental_Setup"}, {"question": "Why does MiniGPT-4 fail to generate a sentence with exactly 20 words, and how does this compare to the performance of other models like GPT-4 and Llama2 13B?", "answer": "Generating a sentence with exactly 20 words is constrained by the capabilities of the underlying language model. While state-of-the-art closed-source models like GPT-4 perform well in this task, open-source models like Llama2 13B struggle. A small experiment with Llama2 13B showed an average of 14.8 words over 50 generations when instructed to 'generate a sentence with 20 words.' Since MiniGPT-4 uses a frozen open-source LLM, it cannot precisely generate a 20-word caption. However, this prompt still encourages MiniGPT-4 to significantly reduce the image description length from 175 to 28.8 words.", "category": "Experimental_Analysis"}, {"question": "How does the proposed MiniGPT-4 model compare with recent open-sourced components like LLaVA, InstructBLIP, and mPlug-Owl in terms of performance on multimedia benchmarks such as MMBench?", "answer": "MiniGPT-4 has been evaluated in the MMBench benchmark, as well as in other benchmarks like MME and VisIT-Bench, alongside similar baseline methods like LLaVa and InstructBLIP. In the MMBench benchmark, which focuses on visual reasoning, MiniGPT-4 demonstrates competitive performance compared to contemporary methods such as InstructBlip. It surpasses InstructBlip in key areas including logical reasoning (LR), fine-grained perception for single instance (FP-S), and fine-grained perception across instances (FP-C). Additionally, MiniGPT-4 achieves competitive results in relation reasoning (RR), attribute reasoning (AR), and coarse perception (CP). Results from MMBenchmark are listed below.\n\n| Model         | Overall | LR   | AR   | RR   | FP-S | FP-C | CP   |\n|---------------|---------|------|------|------|------|------|------|\n| OpenFlamingo  | 4.6%    | 6.7% | 8.0% | 0.0% | 6.7% | 2.8% | 2.0% |\n| VisualGLM     | 38.1%   | 10.8%| 44.3%| 35.7%| 43.8%| 23.4%| 47.3%|\n| LLaVa         | 38.7%   | 16.7%| 48.3%| 30.4%| 45.5%| 32.4%| 40.6%|\n| InstructBlip  | 44%     | 19.1%| 54.2%| 34.8%| 47.8%| 24.8%| 56.4%|\n| MiniGPT-4     | 42.3%   | 20.8%| 50.7%| 30.4%| 49.5%| 26.2%| 50.7%|\n", "category": "Method_Comparison"}, {"question": "How do the AOK-VQA and GQA results, along with qualitative analysis, support the claim that Qformer does not play a critical role in MiniGPT-4's advanced skills?", "answer": "The variant of MiniGPT-4 without Q-Former achieved a score 1.2 points higher in GQA and 1.3 points lower in AOK-VQA compared to the original model. This result demonstrates similar performance to the original version, as detailed in Table 5 of Section 4.3. Qualitative analysis in Figures 4, 13, and 14 also shows comparable results between the no-Q-former variant and the original model, substantiating the claim.", "category": "Experimental_Analysis"}, {"question": "Did the paper initially acknowledge the existence of recent multi-modal instruction tuning datasets like LLaVA, InstructBLIP, M3IT, and MultiInstruct in the vision-language domain?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the length of the learned visual queries from the Q-Former in Mini-GPT4 32?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper provide evidence that MiniGPT-4 achieves similar results with and without Qformer in AOK-VQA and GQA tasks?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "Fx2SbBgcte", "venue": "ICLR 2024", "paper_decision": "Accept (spotlight)", "title": "AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning", "authors": ["Yuwei Guo", "Ceyuan Yang", "Anyi Rao", "Zhengyang Liang", "Yaohui Wang", "Yu Qiao", "Maneesh Agrawala", "Dahua Lin", "Bo Dai"], "abstract": "With the advance of text-to-image (T2I) diffusion models (e.g., Stable Diffusion) and corresponding personalization techniques such as DreamBooth and LoRA, everyone can manifest their imagination into high-quality images at an affordable cost. However, adding motion dynamics to existing high-quality personalized T2Is and enabling them to generate animations remains an open challenge. In this paper, we present AnimateDiff, a practical framework for animating personalized T2I models without requiring model-specific tuning. At the core of our framework is a plug-and-play motion module that can be trained once and seamlessly integrated into any personalized T2Is originating from the same base T2I. Through our proposed training strategy, the motion module effectively learns transferable motion priors from real-world videos. Once trained, the motion module can be inserted into a personalized T2I model to form a personalized animation generator. We further propose MotionLoRA, a lightweight fine-tuning technique for AnimateDiff that enables a pre-trained motion module to adapt to new motion patterns, such as different shot types, at a low training and data collection cost. We evaluate AnimateDiff and MotionLoRA on several public representative personalized T2I models collected from the community. The results demonstrate that our approaches help these models generate temporally smooth animation clips while preserving the visual quality and motion diversity. Codes and pre-trained weights are available at https://github.com/guoyww/AnimateDiff.", "keywords": "Deep Learning;Diffusion Model;Video Generation", "qa_pairs": [{"question": "How does MotionLoRA learn motion patterns, and can these patterns be inferred from text clues?", "answer": "MotionLoRA does not learn new motion patterns entirely from scratch but refines and enhances pre-existing motion priors obtained during pre-training. Whether a motion pattern can be triggered via text prompt depends on the accuracy of motion labels in the training dataset. For example, in the WebVid-10M dataset, 'zooming' is a coarse label that includes both zoom-in and zoom-out, so text hints like 'zoom in' result in both motions. More complex motion patterns like 'Rolling' lack specific text labels and cannot be triggered via simple textual clues. MotionLoRA enables the motion module to express desired motion priors at inference.", "category": "Method_Mechanics"}, {"question": "Why does the identity preservation in the image animation results appear poor?", "answer": "The method primarily focuses on generating personalized videos from text prompts, rather than animating a given image, which is not the primary objective of this submission. Temporal consistency, which is related to identity preservation, can be improved through scale-up training as detailed in Appendix B.3 and illustrated in Figures 11.", "category": "Experimental_Analysis"}, {"question": "What is the reason for the apparent visual link between the scale of the domain adaptor in Figure 6 and camera movement?", "answer": "The apparent link is due to a misunderstanding. The domain adaptor is used only during training to reduce the gap between video training data and high-quality training images of T2I, not for domain transfer during inference. This not only results in high-quality video generation but also offers a new usage of domain adapters. The pose change in Figure 6 is caused by the varying image distribution, not camera movement. Adjusting the scale of the domain adaptor from 1.0 to 0.0 shifts the generated appearance from the WebVid domain to the personalized image domain.", "category": "Method_Disambiguation"}, {"question": "What are the visualization results when the domain adapter training (stage 1) is entirely omitted from the process, as opposed to setting the parameter $α$ to zero at the inference state?", "answer": "The visualizations show that when the domain adapter is entirely removed from the training pipeline, the watermarks from the training dataset appear on the synthetic animations. This is evident in the 1st row of the figures. When the adapter has its full impact, the watermarks also appear (2nd row). However, by fitting the visual distribution to a separate domain adapter and dropping it at inference, the full pipeline (3rd row) achieves favorable quality without watermarks. These results are presented in Appendix B.2 and Fig.10.", "category": "Experimental_Exposition"}, {"question": "What is the performance and visual outcome of AnimateDiff when applied directly to the base T2I model used for training the motion module on WebVid, without any enhancements from external models?", "answer": "The generated videos using the original base T2I model, presented in Appendix B.1 and Figures 9, show that the synthesized frames remain smooth over time. However, without enhancements from personalized T2I models, the images closely match the video dataset, sometimes including watermarks. Enhanced base T2I models, like SDXL, are shown in Appendix C.2 and Figures 12, and are recommended for better flexibility and practicality due to the challenges in collecting a larger-scale, higher-quality dataset than WebVid.", "category": "Experimental_Exposition"}, {"question": "How does the dependency of AnimateDiff on the underlying T2I models affect its performance and applicability, especially when the base T2I models face challenges in generating accurate images?", "answer": "AnimateDiff relies on the underlying T2I models, inheriting their appealing generated appearance but also facing potential failure cases due to T2I model limitations. However, AnimateDiff can be improved with advancements in T2I models. Implementation on Stable Diffusion XL (SDXL) enhances visual quality and text conditioning. Comparisons in Appendix C.2 and Fig. 12 show compatibility with other advanced T2I models, mitigating limitations caused by specific T2I models.", "category": "Experimental_Analysis"}, {"question": "What is the technical novelty of the proposed pipeline, given that the domain adapter is based on LoRA, the motion module follows standard feature inflation, and the architecture is based on Timesformer?", "answer": "The technical novelty lies in the unique combination and utilization of the modules, which results in an innovative solution that is more effective than simply combining them. The goal is to turn a T2I model into an animation generator, and feature inflation is the most efficient practice for this purpose. Extensive experiments and high-quality visual results validate the effectiveness of the architecture choices. Additionally, it is not ad-hoc to use these two modules. In Sec. 5.3 and Sec. 4.2, the paper investigates different options for the motion module, e.g., a convolution version, and it turns out the current simple design is strong and efficient. In terms of the domain adapter, as discussed in Sec. 4.1 and Sec. 5.3, it’s worth noting that the paper didn’t use it for domain personalization as in previous work. Instead, the paper only uses it during training to reduce the gap between video training data and high-quality training images of T2I. This not only results in high-quality video generation but also offers a new usage of domain adapters.", "category": "Method_Disambiguation"}, {"question": "How reliable is the use of LoRA for domain adaptation in the motion module, and what steps are taken to ensure generalization to new or complex motion patterns not covered in the training set?", "answer": "The domain adapter, implemented as an image LoRA, bridges the gap between training videos and images for base T2I models and is discarded after training. Sec.5.3 and Fig.6 demonstrate that this design indeed helps alleviate the negative effects of the video dataset. The motion module is proposed to learn the motion pattern regarding the video dataset. For instance, videos in the training set usually involve how humans should act and how cars should move. In this case, the generalization ability largely depends on the motion that the training set covers. After this learning, personalized video generation is enabled with a user-specific domain by inserting a personalized image LoRA. One straightforward solution for some specific complex movements that might not appear in the training set is to collect the larger-scale corresponding videos into the training set and tune the motion module. However, this is costly and impractical. Instead, MotionLoRA is proposed as a complementary method to learn specific motion patterns based on the general motion prior. Training MotionLoRA requires 20-50 reference videos and takes about 1 hour, providing a practical solution. AnimateDiff alleviates the effects of training videos, captures their motion prior. More importantly, it could quickly and efficiently adapt to other new/complex motion patterns that the aforementioned training set cannot cover.", "category": "Method_Mechanics"}, {"question": "How does the technical novelty of the paper's work manifest in the use of transformer blocks, LORA, and adapters, given that these modules are not novel individually and are used in other personalization papers?", "answer": "As discussed in Sec. 1, the paper’s primary goal is to develop an effective pipeline to animate personalized T2I models. The technical novelty of the paper’s work is not in the modules themselves, but in the unique way they are combined and utilized, leading to an innovative solution that is greater than simply putting them together. Furthermore, the use of these two modules is not ad-hoc. In Sec. 5.3 and Sec. 4.2, the paper investigates different options for the motion module, e.g., a convolution version, and the investigation reveals that the current simple design is strong and efficient. In terms of the domain adapter, as discussed in Sec. 4.1 and Sec. 5.3, the paper does not use it for domain personalization as in previous work. Instead, the paper only uses it during training to reduce the gap between video training data and high-quality training images of T2I. This not only results in high-quality video generation but also offers a new usage of domain adapters.", "category": "Method_Disambiguation"}, {"question": "How does the motion module in AnimateDiff learn motions like zoom-ins and rolls when pre-trained on the WebVid dataset, and does it require explicit training or can it infer these motions from textual cues alone?", "answer": "The motion module in AnimateDiff learns motions by utilizing data priors from the WebVid-10M dataset, which includes real-world footage such as outdoor scenes and human activities. This allows the module to learn common real-world motion patterns like ocean waves, car and boat movements, and human body/hand motions, as illustrated in Fig. 1, without the need for explicit training using LORA with a curated dataset.", "category": "Method_Mechanics"}, {"question": "How does the pipeline perform on different motion categories, and what are the limitations when generating complex or rare motions?", "answer": "The pipeline performs well on non-violent motions such as fluid (e.g., ocean waves, fog), rigid objects (e.g., cars, boats), and simple human movements (e.g., walking, facial expressions), as shown in Figures 1, 4, and 9. However, when generating complex or rare motions rarely seen in the training dataset, such as dance or camera movements with drastic scene changes, the model may produce static animation or unnatural deformations.", "category": "Experimental_Exposition"}, {"question": "What steps have been taken to address the issues of minimal motion amplitudes and flickering between frames in the animation results, and how are these improvements demonstrated in the study?", "answer": "The experimental results in Appendix B.3 and Figures 11 demonstrate that *scale-up training* improves motion amplitude, diversity, and temporal consistency in the generated videos. This indicates that the current model design is capable of capturing larger motion. Additionally, enlarging the video-text paired dataset is identified as a potential avenue for further improvement, though it is acknowledged to be challenging and costly. ", "category": "Experimental_Analysis"}]}
{"id": "FQepisCUWu", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate", "authors": ["Chi-Min Chan", "Weize Chen", "Yusheng Su", "Jianxuan Yu", "Wei Xue", "Shanghang Zhang", "Jie Fu", "Zhiyuan Liu"], "abstract": "Text evaluation has historically posed significant challenges, often demanding substantial labor and time cost. With the emergence of large language models (LLMs), researchers have explored LLMs' potential as alternatives for human evaluation. While these single-agent-based approaches show promise, experimental results suggest that further advancements are needed to bridge the gap between their current effectiveness and human-level evaluation quality.\nRecognizing that best practices of human evaluation processes often involve multiple human annotators collaborating in the evaluation, we resort to a multi-agent debate framework, moving beyond single-agent prompting strategies.\nIn this paper, we construct a multi-agent referee team called $\\textbf{ChatEval}$ to autonomously discuss and evaluate the quality of different texts. \nOur experiments on two benchmarks illustrate that ChatEval delivers superior accuracy and correlation in alignment with human assessment. Furthermore, we find that the diverse role prompts (different personas) are essential in the multi-agent debate process; that is, utilizing the same role description in the prompts can lead to a degradation in performance. Our qualitative analysis also shows that ChatEval transcends mere textual scoring, offering a human-mimicking evaluation process for reliable assessments.", "keywords": "Large Language Model;Multi-Agent Debate;LLM evaluators", "qa_pairs": [{"question": "Can the two agents employed in the implementation of ChatEval always reach an agreement at the end of debates?", "answer": "The analysis of inter-annotator agreement experiment results, measured using ICC (Intraclass Correlation Coefficient), shows that the annotators are generally able to reach a consensus by the end of their debate. However, as the number of roles increases, there is a decrease in agreement. It is reasonable considering that it's not also easy for a team of people to reach a consensus. Despite this, the overall ICC values remain high, indicating strong agreement among annotators.\n\n|     | 2 roles | 3 roles | 4 roles | 5 roles |\n|-----|---------|---------|---------|---------|\n| ICC | 0.981   | 0.955   | 0.890   | 0.881   |", "category": "Experimental_Exposition"}, {"question": "What are results available for multiple large language models (LLMs) used in an ensemble manner as described in Table 1?", "answer": "The results for the ensemble method applied to the Topical-Chat benchmark show some improvement compared to the single agent (SA) method. However, the ensemble method still performs less effectively than the ChatEval framework, demonstrating the superiority of the proposed framework.\n\n|                   | Nat| | Coh|       | Eng|       | Grd|       | Avg|       |\n|-------------------|---------|-------|-------|-------|-------|-------|-------|-------|-------|-------|\n|                   |  $\\tau$ |   $\\rho$    |  $\\tau$     |   $\\rho$    |   $\\tau$    |   $\\rho$    |   $\\tau$    |   $\\rho$    |   $\\tau$    |     $\\rho$  |\n| ChatGPT(SA)       | 0.474   | 0.421 | 0.527 | 0.482 | 0.599 | 0.549 | 0.576 | 0.558 | 0.544 | 0.503 |\n| ChatGPT(Ensemble) | 0.421   | 0.359 | 0.486 | 0.442 | 0.611 | 0.551 | 0.661 | 0.628 | 0.545 | 0.495 |\n| ChatGPT(ChatEval)       | 0.441   | 0.396 | 0.500 | 0.454 | 0.664 | 0.607 | 0.602 | 0.583 | 0.552 | 0.510 |\n| GPT-4(SA)         | 0.532   | 0.483 | 0.591 | 0.535 | 0.734 | 0.676 | 0.774 | 0.750 | 0.658 | 0.611 |\n| GPT-4(Ensemble)   | 0.512   | 0.450 | 0.607 | 0.544 | 0.755 | 0.693 | 0.781 | 0.756 | 0.664 | 0.611 |\n| GPT-4(ChatEval)         | 0.630   | 0.571 | 0.619 | 0.561 | 0.765 | 0.695 | 0.722 | 0.700 | 0.684 | 0.632 |", "category": "Experimental_Exposition"}, {"question": "How does the proposed framework's effectiveness vary with different large language models (LLMs), and has there been an evaluation of ChatEval's generalization ability across various LLMs?", "answer": "The paper's method is also applicable to a range of LLMs and it operates orthogonally to SA prompting techniques. It complements and extends these techniques by integrating them within a MA debate context, which can leverage the unique strengths of different LLMs in a collaborative setting. \n\nWhile the paper’s approach was benchmarked against the g-eval method, the comparison may not be entirely straightforward. The g-eval method, which achieved a better 3.5 performance score, does not solely rely on COT prompting; instead, it requires access to the language model's logits. These logits are then used to perform a weighted calculation, which differs from the paper’s methodology. However, the paper's proposed MA approach does not necessitate access to the model's logits—a requirement that is often impractical in real-world applications. Consequently, it is expected that there would be performance discrepancies between the paper’s method and g-eval.\n\nAdditionally,  the paper's experimental results demonstrate that the MA approach (MA ChatEval), surpasses the performance of the SA COT method referenced in the paper’s main text.  ", "category": "Experimental_Analysis"}, {"question": "Does the multiple-agent debating framework (MA ChatEval) perform better than the single-agent COT framework, as indicated by the performance differences in Table 2?", "answer": "Yes, the experimental results demonstrate that the MA approach (MA ChatEval) surpasses the performance of the SA COT method. The MA framework operates orthogonally to SA prompting techniques, integrating them within a collaborative multi-agent debate context, which leverages the unique strengths of different LLMs.", "category": "Experimental_Exposition"}, {"question": "What is the comparative analysis of time and financial costs between human annotation and various LLM-based evaluators in the proposed framework?", "answer": "The appended table shows that LLM-based evaluators, including gpt-4 and ChatGPT in different configurations, significantly reduce both time and financial expenditures compared to human evaluators. While the chateval framework has higher costs than the SA and MA ensemble methods due to multiple inference rounds, it remains more cost-effective than human evaluators, which is considered an acceptable trade-off for the benefits provided.\n\n |                | Cost    | Time   |\n|-------------------|---------|--------|\n|Human             | $90    | 240min |\n|gpt-4 (SA)    | $2.90  | 10min  |\n|gpt-4 (MA ensemble)  | $8.70  | 10min  |\n|gpt-4 (MA chateval)  | $12.30 | 36min  |\n|Chatgpt(SA)   | $0.11  | 3min   |\n|Chatgpt(MA ensemble) | $0.33  | 3min   |\n|Chatgpt(MA chateval) | $0.41  | 11min  |", "category": "Experimental_Exposition"}, {"question": "How does the framework address the challenges of evaluating longer dialogue conversations, as highlighted in Section 4.3?", "answer": "The framework incorporates a summary-style communication strategy to mitigate issues arising from lengthy historical context. At the end of each debate iteration, an extra LLM is prompted to summarize the messages conveyed so far, and this condensed information is used as historical context. The results in Section 4.3 demonstrate that this summary-style approach exhibits better scalability. Additionally, the principal aim of the paper's work is to establish a more comprehensive mechanism for evaluating textual content. The experimental results indicate that the application of a debating format can indeed catalyze improved performance. As more capable LLMs emerge, particularly those adept at managing exceptionally long contexts, the paper's framework is inherently equipped to generalize these advancements. The integration of such LLMs is anticipated to unlock the full potential of the paper's framework further.", "category": "Method_Mechanics"}, {"question": "How do these strategies(One-by-One、Simultaneous-Talk and Simultaneous-Talk-with-Summarizer) compare in performance to the SA method?", "answer": "The data presented in the table indicate that each of the strategies yields a statistically consistent enhancement over the SA method, affirming the efficacy of the paper's approach.\n\n|                             |                       | Acc.        | Kap          |\n|------------------------------|------------------------|-------------|--------------|\n|ChatGPT                      | SA         | $53.7_{\\pm 1.4}$ | $0.27_{\\pm 0.02}$ |\n|ChatGPT                      | one-by-one           | $60.0_{\\pm 0.9}$ | $0.30_{\\pm 0.02}$ |\n|ChatGPT                      | simultaneous-talk | $59.3_{\\pm 1.0}$ | $0.29_{\\pm 0.02}$ |\n|ChatGPT                     | simultaneous-talk-with-summarizer | $58.8_{\\pm 1.3}$ | $0.28_{\\pm 0.03}$ |\n|GPT-4                         | SA         | $60.8_{\\pm 0.7}$ | $0.36_{\\pm 0.01}$ |\n|GPT-4                         | one-by-one           | $63.8_{\\pm 0.9}$ | $0.40_{\\pm 0.01}$ |\n|GPT-4                         | simultaneous-talk | $63.2_{\\pm 0.6}$ | $0.38_{\\pm 0.02}$ |\n|GPT-4                         | simultaneous-talk-with-summarizer | $63.4_{\\pm 0.8}$ | $0.39_{\\pm 0.02}$ |", "category": "Method_Comparison"}, {"question": "What are the specific differences in performance between the single-agent (SA) and multi-agent (MA) approaches across various sub-metrics for ChatGPT and GPT-4 on the Topical-Chat dataset?", "answer": "On average, the paper's MA approach surpasses the SA method in performance when using both ChatGPT and GPT-4. However, there are subtle variations in performance across different sub-metrics between the two LLMs. For ChatGPT, the MA approach slightly underperforms in Naturalness and Coherence but shows improvements in Engagingness and Groundedness. For GPT-4, the MA approach displays a more uniform improvement across three sub-metrics, with only a minor decline in Groundedness. This variability is observed not only in the paper's proposed metrics but also in traditional metrics and previous methods like g-eval, where g-eval-4's performance dips in Engagingness and Groundedness compared to g-eval-3.5. Additionally, GPT-4 (ChatEval) consistently excels over ChatGPT (ChatEval) in both SA and MA configurations, and this trend represents reasonable and scalable behavior.", "category": "Experimental_Exposition"}, {"question": "Can the proposed multi-agent framework generalize to other large language models (LLMs) beyond ChatGPT and GPT-4, and does it require models with decent performance to enable constructive discussion and evaluation?", "answer": "Additional experiments were conducted with Llama2-chat-7b and Vicuna-7b-v1.5 to evaluate the generalizability of the multi-agent framework. Llama2 achieves only marginal performance under single-agent CoT methods. Despite reaching an accuracy of 37.5%, the negative kappa coefficient suggests that the observed agreement is less than what would be expected by chance. However, it's noteworthy that through the application of a multi-agent debate, Llama2 demonstrates a modest improvement, albeit still marginally above chance levels. In contrast, Vicuna exhibits more robust performance improvement compared to Llama2. ChatEval notably enhances its capabilities, achieving an accuracy of 52.3% and a kappa coefficient of 0.19, indicative of fair agreement beyond chance. While both models show limitations in their performance, these results demonstrate the effectiveness of ChatEval and underscore its positive impact. Furthermore, the paper's focus on ChatGPT and GPT-4 stems from their recognition in prior studies as benchmarks for evaluating LLM-based evaluators. They are the most widely used LLM-based evaluators in various scenarios and have been shown effective under SA settings. This established efficacy provides a solid foundation for the paper's  decision to build upon ChatGPT and GPT-4 for our MA framework. In comparison, some other smaller models diverge from human preference even in SA settings, let alone the multi-agent debate framework. To ensure consistency and comparability with previous research, it is appropriate to conduct the paper's experiments with these two models.", "category": "Experimental_Analysis"}, {"question": "What are the limitations of the proposed method in terms of inference speed and cost, particularly for multi-agent systems requiring multiple rounds of generation with the LLM?", "answer": "The cost analysis table shows that while multi-agent (MA) debate inherently incurs higher costs compared to single-agent (SA) systems, it remains significantly more cost-effective than using human annotators in terms of both time and expense. \n\n|                 | Cost    | Time   | \n|-------------------|---------|--------| \n|Human             | $90    | 240min | \n|gpt-4 (SA)    | $2.90  | 10min  | \n|gpt-4 (MA ensemble)  | $8.70  | 10min  | \n|gpt-4 (MA chateval)  | $12.30 | 36min  | \n|Chatgpt(SA)   | $0.11  | 3min   | \n|Chatgpt(MA ensemble) | $0.33  | 3min   | \n|Chatgpt(MA chateval) | $0.41  | 11min  |", "category": "Method_Disambiguation"}, {"question": "What are the roles of the two agents used in the default experiments, and how are specific roles compared against the default setting in Figure 3? Additionally, what are the specific versions of ChatGPT and GPT-4 used in the paper?", "answer": "The two agents used in the default experiments are 'general public' and 'critic'. In Figure 3, specific roles are compared against the default setting, where 'general' on the x-axis represents the default setting. The specific versions used are 'ChatGPT' corresponding to gpt-3.5turbo-0301 and 'GPT4' to gpt-4-0314.", "category": "Experimental_Setup"}, {"question": "Is it necessary for a large language model (LLM) to perform well in a single-agent evaluation setting to succeed in a multi-agent scenario? Additionally, can weaker LLMs achieve performance improvements through multi-agent debate, making them a cost-effective alternative to more powerful models like GPT?", "answer": "The paper's findings demonstrate that ChatEval, a multi-agent debate framework, can significantly enhance the performance of weaker LLMs, even if they perform poorly in single-agent settings. For example, Llama2-chat-7b, which achieves only marginal performance under single-agent CoT methods, demonstrates modest improvement through the application of a multi-agent debate, albeit still marginally above chance levels. In contrast, Vicuna-7b-v1.5 exhibits more robust performance improvement compared to Llama2. ChatEval notably enhances its capabilities, achieving an accuracy of 52.3% and a kappa coefficient of 0.19, indicative of fair agreement beyond chance. These results suggest that even weaker LLMs can benefit from the multi-agent debate framework, which could indeed be a cost-effective choice for certain applications.", "category": "Experimental_Analysis"}, {"question": "How does the resource consumption of the proposed method compare to previous methods, specifically in terms of cost and time?", "answer": "The proposed method substantially lowers both time and labor costs compared to the employment of human annotators. The table provided shows the following comparisons: \n\n|                | Cost    | Time   |\n|-------------------|---------|--------|\n|Human             | $90    | 240min |\n|gpt-4 (SA)    | $2.90  | 10min  |\n|gpt-4 (MA ensemble)  | $8.70  | 10min  |\n|gpt-4 (MA chateval)  | $12.30 | 36min  |\n|Chatgpt(SA)   | $0.11  | 3min   |\n|Chatgpt(MA ensemble) | $0.33  | 3min   |\n|Chatgpt(MA chateval) | $0.41  | 11min  |", "category": "Method_Comparison"}, {"question": "What is the cost and time requirement for baseline methods like G-EVAL, and how does it compare to the SA COT method?", "answer": "The cost and time requirements for baseline methods like G-EVAL are detailed in the provided table.\n\n|                | Cost    | Time   |\n|-------------------|---------|--------|\n|G-eval-3.5 | $2.2   | 3min |\n|G-eval-4    | $58  | 10min  |\n|FairEval-chatgpt | $0.34  | 3min  |\n|FairEval-gpt4  | $6.38 | 10min  |\n\nThese methods require multiple inferences per instance, assumed to be run asynchronously, making their time cost equivalent to that of an SA COT method. Specifically, G-eval incurs a cost 20 times that of an SA COT method due to its need for 20 samples to estimate token probabilities.", "category": "Method_Comparison"}, {"question": "What does the term 'human-inspired' refer to in the qualitative analysis?", "answer": "The term 'human-inspired' refers to four patterns identified in the paper, which are heuristically derived from observing human behavior during debates. ", "category": "Concept_Understanding"}, {"question": "How does the debate framework in ChatEval contribute to the transparency and interpretability of AI systems?", "answer": "The debate framework in ChatEval transcends being merely an evaluative tool; it is instrumental in increasing the transparency and interpretability of AI systems. OpenAI has indeed articulated in their blog[1] that debate is a technique intended to refine alignment research, aiding humans in evaluating complex tasks beyond direct human assessment. While their own work [2] on debate has not extensively covered experiments in natural language and remains preliminary, the paper's framework's utilization of natural language initiates a crucial dialogue on the evolving role of AI in evaluation processes. The paper's framework envisages AI as a collaborative partner in the evaluative discourse, not just the subject of evaluation. \n\n[1] https://openai.com/blog/our-approach-to-alignment-research\n[2]https://arxiv.org/abs/1805.00899", "category": "Method_Mechanics"}, {"question": "How does the proposed MA debate framework differ from contemporary works that also use multiple LLMs to enhance the performance of LLM-evaluators?", "answer": "The contemporary research the paper referenced share the same aims to enhance the performance of llm-evaluators through multiple llms. However, the proposed MA debate framework differs from contemporary works in that it investigates the use of the MA debate framework and emphasizes the potentiality and heuristics of natural language communication, which are overlooked in other works. For example, while one referenced work[1] uses multiple LLMs to consider pairwise preferences of answer pairs, it does not focus on the intricate components of debate frameworks, such as role design and communication strategy.Moreover, their results have been shown to be inferior to those of the paper. Another work[2] aggregates results in an ensemble manner without engaging models in a debate. \n\n[1]https://arxiv.org/abs/2307.02762\n\n[2] https://arxiv.org/abs/2308.01862", "category": "Method_Comparison"}, {"question": "What is the origin of the role design and communication strategy, and how do they contribute to the study of impersonation?", "answer": "The role prompt design originated from existing literature as a foundational reference, which serves as a starting point for studying impersonation. The effectiveness of this design is demonstrated in Table 3. Additionally, alternative designs were developed and assessed, with details in Section 4.1, showing that specifically crafted role prompts perform better in their respective categories, as shown in Figure 3. Regarding communication strategy, the 'one-by-one' method is a common practice for turn-taking debates, adopted from a previous framework. Two additional strategies, 'simultaneous-talk' and 'simultaneous-talk-with-summarizer,' were introduced, inspired by human interaction patterns and inherently have different functionalities in the paper's framework. For example, 'simultaneous-talk' relaxes the rigid sequence of turns, while 'simultaneous-talk-with-summarizer' addresses the challenge of maintaining a lengthy conversation history. These functionalities are discussed in section 4.2.", "category": "Motivation_Analysis"}, {"question": "Are the differences in performance between the methods statistically significant, and how is this demonstrated in the results?", "answer": "The differences in performance between the methods are statistically significant, as demonstrated by running the experiments 5 times and including standard deviations in the results. The consistent stability and statistically significant improvements highlight the robustness of the proposed method. The objective is to show that the method consistently outperforms others, as evidenced by the average and standard deviation values provided in the results table.\n\n|                             |                        | Acc.        | Kap          |\n|------------------------------|------------------------|-------------|--------------|\n|ChatGPT                      | Single-Agent           | $53.7_{\\pm 1.4}$ | $0.27_{\\pm 0.02}$ |\n|ChatGPT                      | Multi-Agent (Ensemble) | $55.5_{\\pm 0.7}$ | $0.29_{\\pm 0.01}$ |\n|ChatGPT                     | Multi-Agent (ChatEval) | $60.0_{\\pm 0.9}$ | $0.30_{\\pm 0.02}$ |\n|GPT-4                         | Single-Agent           | $60.8_{\\pm 0.7}$ | $0.36_{\\pm 0.01}$ |\n|GPT-4                         | Multi-Agent (Ensemble) | $61.5_{\\pm 0.5}$ | $0.38_{\\pm 0.01}$ |\n|GPT-4                         | Multi-Agent (ChatEval) | $63.8_{\\pm 0.9}$ | $0.40_{\\pm 0.01}$ |", "category": "Method_Comparison"}]}
{"id": "4WnqRR915j", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "Llemma: An Open Language Model for Mathematics", "authors": ["Zhangir Azerbayev", "Hailey Schoelkopf", "Keiran Paster", "Marco Dos Santos", "Stephen Marcus McAleer", "Albert Q. Jiang", "Jia Deng", "Stella Biderman", "Sean Welleck"], "abstract": "We present Llemma, a large language model for mathematics. We continue pretraining Code Llama on the Proof-Pile-2, a mixture of scientific papers, web data containing mathematics, and mathematical code, yielding Llemma. On the MATH benchmark Llemma outperforms all known openly released models, as well as the unreleased Minerva model suite on an equi-parameter basis. Moreover, Llemma is capable of tool use and formal theorem proving without any finetuning. We openly release all artifacts, including 7 billion and 34 billion parameter models, the Proof-Pile-2, and code to replicate our experiments.", "keywords": "reasoning;language models;pretraining", "qa_pairs": [{"question": "Did the paper conduct an ablation study to determine the optimal mixture of data sources, and what were the findings?", "answer": "Yes, the authors conducted an ablation study, which is reported in section 3.4 and Table 5 of the paper. The study confirmed that large-scale language modeling performance is sensitive to the precise choice of data mixture. In contrast, many prior works on language modeling [1, 2] provide little justification for their data mixture.\n\n[1] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, & Connor Leahy. (2020). The Pile: An 800GB Dataset of Diverse Text for Language Modeling.\n\n[2] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, & Guillaume Lample. (2023). LLaMA: Open and Efficient Foundation Language Models.", "category": "Experimental_Exposition"}, {"question": "What are the novel aspects of the proposed model in terms of few-shot formal theorem proving and memorization study?", "answer": "1. Few-shot formal theorem proving: Prior work using language models for formal theorem proving has either relied on models fine-tuned for specific theorem proving systems [1, 2] or proprietary models fine-tuned on private datasets [2]. The paper demonstrates the first instance of an open base model achieving strong theorem proving capabilities without fine-tuning. One reason this is a useful capability is that supervised fine-tuning data for formal theorem proving and autoformalization are scarce, which means being able to fine-tune from a strong base model is critical.\n2. Memorization study: In the paper’s memorization study (section 3.5), it is found that the paper’s model does not attain a higher accuracy on problems that appear in the training set compared to those that do not. This is a counterintuitive finding that warrants awareness among the research community and further investigation. Furthermore, Llemma is viewed not just as the outcome of research, but also as research infrastructure that can support future work. For example, many works on machine learning theorem proving [1, 2, 3] may benefit greatly from a stronger mathematics language model.\n\n[1] Kaiyu Yang, Aidan M. Swope, Alex Gu, Rahul Chalamala, Peiyang Song, Shixing Yu, Saad Godil, Ryan Prenger, & Anima Anandkumar. (2023). LeanDojo: Theorem Proving with Retrieval-Augmented Language Models.\n\n[2] Emily First, Markus N. Rabe, Talia Ringer, & Yuriy Brun. (2023). Baldur: Whole-Proof Generation and Repair with Large Language Models.\n\n[3] Amitayush Thakur, Yeming Wen, & Swarat Chaudhuri. (2023). A Language-Agent Approach to Formal Theorem-Proving.", "category": "Method_Disambiguation"}, {"question": "Did the paper conduct experiments to evaluate different combinations of training data, and what were the findings?", "answer": "Yes, an ablation study was conducted to determine the optimal mixture of data sources, as reported in Section 3.4 and Table 5. The study confirmed that large-scale language modeling performance is sensitive to the precise choice of data mixture.", "category": "Experimental_Exposition"}, {"question": "What methodology was used for instruction fine-tuning in the experiments involving Llemma on MATH and GSM8k datasets?", "answer": "All results in the main body of the paper show Llemma’s performance via few-shot prompting, without applying any finetuning. In Appendix G and Table 12 of the paper, the paper also demonstrates that Llemma is able to substantially improve its performance on MATH and GSM8k when fine-tuned on MetaMathQA, a supervised finetuning dataset targeted at mathematical problem solving. For these experiments, the paper uses exactly the same methodology as [1].\n\n[1]Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T. Kwok, Zhenguo Li, Adrian Weller, & Weiyang Liu. (2023). MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models.", "category": "Experimental_Setup"}, {"question": "What unique methodological insights does the Llemma approach provide for adapting language models to mathematics?", "answer": "Llemma offers a comprehensive recipe for adapting language models to mathematics through continued pretraining, detailing data processing, data mixture selection, model training, and evaluation methodologies, including novel approaches for studying data overlap and evaluating models in few-shot formal theorem proving. This full recipe is rare at the considered scale and provides insights into training language models for specialized domains.", "category": "Method_Disambiguation"}, {"question": "Given the differences in datasets, model architecture, and training methods between Llemma and Minerva, How can the good performance of Llemma be attributed to specific components, and why was a detailed ablation study not conducted?", "answer": "The authors used Minerva as a baseline because it is a state-of-the-art methodology for adapting language models to mathematics. A detailed ablation study comparing each component of Minerva’s data and training pipeline with Llemma’s was not conducted because Minerva is entirely closed source, making such a study impossible. However, the authors ensured that, wherever possible, Llemma and Minerva were evaluated on the same evaluation setup, such as using the exact few-shot prompt and answer extraction method used by Minerva, to make the comparisons as detailed as possible given Minerva’s closed-source nature.", "category": "Experimental_Analysis"}, {"question": "What are the reasons for choosing Code Llama initialization over Llama 2 initialization or random initialization for training Llemma ?", "answer": "The choice to initialize Llemma with Code Llama weights instead of Llama 2 weights was made due to the importance of coding and theorem proving tasks for Llemma. Llama 2 has weak coding capabilities, whereas code pretraining, as seen in Code Llama, enhances general reasoning abilities, demonstrated by Code Llama’s superior MATH score compared to Llama 2. Additionally, the compute cost of conducting head-to-head comparisons with random initialization and Llama 2 initialization is prohibitive, with training costs for Llemma 7B and 34B being around 95,000 USD and 195,000 USD respectively on an AWS 8xA100 instance.", "category": "Motivation_Analysis"}, {"question": "Is fine-tuning still necessary for Llemma, and what are its benefits?", "answer": "Yes, fine-tuning Llemma on downstream applications is beneficial. For example, fine-tuning on supervised datasets for mathematical problem solving improves its accuracy on MATH and GSM8k, as shown in Appendix G. Additionally, fine-tuning can align Llemma’s outputs to specific styles or applications, such as enhancing its chat dialog capabilities.", "category": "Method_Disambiguation"}]}
{"id": "zAdUB0aCTQ", "venue": "ICLR 2024", "paper_decision": "Accept (poster)", "title": "AgentBench: Evaluating LLMs as Agents", "authors": ["Xiao Liu", "Hao Yu", "Hanchen Zhang", "Yifan Xu", "Xuanyu Lei", "Hanyu Lai", "Yu Gu", "Hangliang Ding", "Kaiwen Men", "Kejuan Yang", "Shudan Zhang", "Xiang Deng", "Aohan Zeng", "Zhengxiao Du", "Chenhui Zhang", "Sheng Shen", "Tianjun Zhang", "Yu Su", "Huan Sun", "Minlie Huang", "Yuxiao Dong", "Jie Tang"], "abstract": "The potential of Large Language Model (LLM) as agents has been widely acknowledged recently.\nThus, there is an urgent need to quantitatively evaluate LLMs as agents on challenging tasks in interactive environments.\nWe present AgentBench, a multi-dimensional benchmark that consists of 8 distinct environments to assess LLM-as-Agent's reasoning and decision-making abilities.\nOur extensive test over 29 API-based and open-sourced (OSS) LLMs shows that, while top commercial LLMs present a strong ability of acting as agents in complex environments, there is a significant disparity in performance between them and many OSS competitors that are no larger than 70B.\nWe identify the typical reasons of failures in environments and LLMs, showing that poor long-term reasoning, decision-making, and instruction following abilities are the main obstacles for developing usable LLM agents.\nImproving instruction following and training on high quality multi-round alignment data could improve agent performance.\nAnd different from existing assumptions, training on code present ambivalent impacts on different agent tasks.\nDatasets, environments, and an integrated evaluation package for AgentBench are released at https://github.com/THUDM/AgentBench.", "keywords": "Large language models;Autonomous agents;Reasoning;Evaluation;Benchmark", "qa_pairs": [{"question": "What are the specific contributions of the paper, particularly in terms of technical insights and the creation of new environments, as opposed to simply applying LLMs to existing environments?", "answer": "The paper makes significant contributions in several areas: 1. Most environments tested in AgentBench are newly created, with 5 out of 8 environments (OS, DB, KG, LTP, DCG) being created for the first time. 2. The evaluation setup involved challenging algorithmic and engineering efforts to create and integrate new environments for scalable testing on LLMs. 3. The paper is the first to quantitatively benchmark up to 27 LLMs on multiple real-world agent tasks in a unified setting. 4. The techniques developed for creating 5 new environments are useful for future agent environment creation, and the integrated testing toolkit serves as a cornerstone for future agent tasks. 5. The analysis of LLM behaviors and shortcomings during the evaluation provides insights for improving LLMs’ agent abilities.", "category": "Method_Disambiguation"}, {"question": "What is the relationship between 'Task Limit Exceeded (TLE)' and the short fixed context lengths of LLMs, and how does this affect the analysis and conclusions in the paper?", "answer": "The term 'Task Limit Exceeded (TLE)' refers to the unexpected exceeding of maximum rounds of interaction as designed, and it has no relationship with LLMs' short fixed context lengths (CLE), which account for less than 1% of all failure cases in most tasks. The tasks were designed to be mostly solvable within around 1000 tokens, with a harder limit of 3500 tokens when earlier histories begin to be folded. The paper's findings in Appendix J.2.3 prove that 95% of completed results are around this limit. Exceeding this limit typically indicates a divergence from the correct solution path (i.e., a failure solution). Notably, as shown in Figure 9 in Appendix J.2.3, in 90% of instances where Task Limit Exceeded occurred, the LLM began to deadly loop over previously generated content, resulting in repetition of identical contents in the final 10 rounds of interaction. It is the repetition that triggers the exceeding of maximum interaction rounds (i.e., TLE), rather than context length exceeding. ", "category": "Experimental_Analysis"}, {"question": "How were the task prompts chosen for evaluating the performance of different LLMs, and what measures were taken to ensure their universal validity and equal testing conditions?", "answer": "The task prompts were chosen based on a combination of representative LLMs to ensure their robustness. The authors experimented with various prompt paradigms, such as ReAct[1] and Chain of Thought (CoT)[2], and tested them across multiple models, including claude-2, gpt-4, gpt-3.5-turbo, and llama-2, to ensure their universal validity. The prompts were not specifically tailored to any particular model, and all models were tested under relatively equal conditions.\n\n[1] Yao, Shunyu, et al. \"ReAct: Synergizing Reasoning and Acting in Language Models.\" The Eleventh International Conference on Learning Representations. 2022.\n\n[2] Wei, Jason, et al. \"Chain-of-thought prompting elicits reasoning in large language models.\" Advances in Neural Information Processing Systems 35 (2022): 24824-24837.", "category": "Experimental_Setup"}, {"question": "What criteria were used to select the text benchmarks for evaluating LLMs as agents, and why is there no benchmark for testing multi-agent interaction?", "answer": "The selection criteria for the text benchmarks were based on two principles:\n1. **Recognition**: Tasks universally acknowledged as crucial in the language agent community, such as NL2Bash[1] and InterCode[2], underscore the importance of OS tasks. Similarly, Wob[3], MiniWob++[4], Webshop[5], Mind2Web[6], and WebArena[7] highlight the critical role of web-based tasks.\n2. **LLM Applicability**: Another key criterion is that the task should be compatible and of proper difficulty levels for current LLMs' few-shot text prompting with multi-round interactions. For example, though datasets like MineDojo[8] can provide more challenging open-world game environments, they generally require vision information beyond the capabilities of existing LLMs.\n\n**Multi-agent evaluation**: Multi-agent interaction evaluation does count as an aspect in LLM-as-Agent evaluation, but the authors think that it might currently be too hard for existing LLMs, especially those smaller ones with fewer than 30 billion parameters, which cannot even handle single-agent tasks. Additionally, widely-acknowledged metrics and tasks for evaluating LLMs' multi-agent capabilities are still not very clear at this time. Therefore, the authors decide to first place the paper's focus on single-agent evaluation at present.\n\n[1] Lin, Xi Victoria, et al. \"NL2Bash: A Corpus and Semantic Parser for Natural Language Interface to the Linux Operating System.\" Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). 2018.\n\n[2] Yang, John, et al. \"InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback.\" arXiv preprint arXiv:2306.14898 (2023).\n\n[3] Shi, Tianlin Tim, et al. \"World of bits: an open-domain platform for web-based agents.\" Proceedings of the 34th International Conference on Machine Learning-Volume 70. 2017.\n\n[4] Liu, Evan Zheran, et al. \"Reinforcement Learning on Web Interfaces using Workflow-Guided Exploration.\" International Conference on Learning Representations. 2018.\n\n[5] Yao, Shunyu, et al. \"Webshop: Towards scalable real-world web interaction with grounded language agents.\" Advances in Neural Information Processing Systems 35 (2022): 20744-20757.\n\n[6] Deng, Xiang, et al. \"Mind2Web: Towards a Generalist Agent for the Web.\" arXiv preprint arXiv:2306.06070 (2023).\n\n[7] Zhou, Shuyan, et al. \"WebArena: A Realistic Web Environment for Building Autonomous Agents.\" NeurIPS 2023 Foundation Models for Decision Making Workshop. 2023.\n\n[8] Fan, Linxi, et al. \"Minedojo: Building open-ended embodied agents with internet-scale knowledge.\" Advances in Neural Information Processing Systems 35 (2022): 18343-18362.", "category": "Motivation_Analysis"}, {"question": "How were the task prompts chosen for the benchmark to ensure fairness across all models, and what steps were taken to avoid giving an unfair advantage to any specific model?", "answer": "The task prompts were chosen based on a combination of representative large language models (LLMs) to ensure their robustness. The authors experimented with various prompt paradigms, such as ReAct[1] and Chain of Thought (CoT)[2], and tested them across multiple models, including claude-2, gpt-4, gpt-3.5-turbo, and llama-2, to ensure universal validity. Task prompts were not specifically tailored to any particular model, and all models were tested under relatively equal conditions.\n\n[1] Yao, Shunyu, et al. \"ReAct: Synergizing Reasoning and Acting in Language Models.\" The Eleventh International Conference on Learning Representations. 2022.\n\n[2] Wei, Jason, et al. \"Chain-of-thought prompting elicits reasoning in large language models.\" Advances in Neural Information Processing Systems 35 (2022): 24824-24837.", "category": "Experimental_Setup"}, {"question": "What measures are in place to address the risk of data leakage from pretraining data over the internet in AgentBench?", "answer": "AgentBench minimizes the risk of data leakage because solving its tasks requires groundtruth trajectories based on dynamic environments (e.g., filenames and paths in Linux dockers of the OS dataset), which are not available as text corpora on the web. As a result, the risk of data leakage in AgentBench is very minimal.", "category": "Method_Mechanics"}, {"question": "What is the formula used to calculate the success rate, and how are different completion statuses accounted for in this calculation?", "answer": "The paper calculates success rate as follows:\n\n$SR = \\frac{C}{N}$,\n\nwhere $C$ represents the count of correct results, and $N$ denotes the total count of samples.\n\nMore specifically, for tasks whose evaluation metrics are success rate, the following situations may occur after each sample has finished:\n\n- Completed:  the final status is a termination status (e.g., submit an answer in OS or DB)\n    - Correct:  the final status is accepted as a correct solution.\n    - Wrong:  the final status is NOT accepted as a correct solution.\n- Context Limit Exceeded\n- Invalid Format\n- Invalid Action\n- Task Limit Exceeded\n\nOnly **Correct** results are the part of successful ones, meaning:\n\n$SR = \\frac{\\#Correct}{\\#Completed + \\#CLE + \\#IF + \\#IA + \\#TLE}$", "category": "Concept_Understanding"}, {"question": "What specific insights does the benchmark provide for improving model performance, particularly in tasks like web-browsing where the model may underperform?", "answer": "- **Observations from Interaction Trajectories**: The paper offers valuable insights, particularly in scenarios where tasks are not successfully completed. \nThe paper has identified two primary issues: \n- **Instruction Following**: Models often struggle to adhere to given instructions. This leads to outputs that are either format-incompatible or lack a valid action. (Cf. Appendix J.2.1) \n- **Output Repetition**: There is a tendency for models to repeat previous outputs after a certain number of interactions. This repetition prevents task completion and causes interactions to reach the maximum turn limit. (Cf. Appendix J.2.3) \n- **Code Training with Ambivalent Impacts**: The paper shows that code training helps certain tasks (e.g., Database, WebShop). However, its introduction might harm other tasks (e.g., Digital Card Games, Operating System) that require more ability in interacting. Later work should be aware of this double-edged sword. (Cf. Section 4.3 & Appendix J.2.5) \n- **High-Quality Alignment Data**: By comparing models like vicuna-13b and llama-2-13b (which share the same base model), the paper shows that alignment data distilled from GPT-4 (e.g., ShareGPT), significantly enhances LLM performance. Vicuna-13b, aligned with ShareGPT data, notably outperforms its counterpart and rivals larger models. (Cf. Section 4.3) \n- **Implications for Fine-Tuning**: These observations suggest key areas for fine-tuning LLMs-as-Agents: one is enhancing instruction compliance, and another is addressing repetitive outputs. \n- **Future Work and Community Contribution**: While the paper's current focus is on establishing a robust agent benchmark for the community, fine-tuning exploration is important. It is also worth mentioning that recent studies like **FireAct[1]** and **AgentTuning[2]** have already been inspired and begun exploring fine-tuning approaches for LLMs-as-Agents. In summary, the research lays the groundwork for future advancements in fine-tuning LLMs-as-Agents, offering a clear direction for addressing key challenges identified in the research.\n\n[1] Chen, Baian, et al. \"FireAct: Toward Language Agent Fine-tuning.\" arXiv preprint arXiv:2310.05915 (2023).\n\n[2] \"AgentTuning: Enabling Generalized Agent Abilities for LLMs.\" arXiv preprint arXiv:2310.12823 (2023).", "category": "Method_Disambiguation"}, {"question": "Why was ALFWorld chosen as an evaluation task despite its simplification of 2D/3D aspects and potential for mastery by a fine-tuned GPT-2?", "answer": "ALFWorld was chosen because, despite being a text-based environment that simplifies complexity, it remains a relevant and challenging task for language models, particularly those not fine-tuned on the task. It requires strategic planning and decision-making within a given scenario, demanding cognitive engagement and planning abilities beyond basic language understanding, thereby offering a valuable framework for evaluating the depth and adaptability of language models. ALFWorld has been widely adopted in recent LLM-as-Agent studies (e.g., ReAct[1], Reflextion[2]), and its use in AgentBench aims to unify evaluation and results for later LLMs. \n\n[1] Yao, Shunyu, et al. \"ReAct: Synergizing Reasoning and Acting in Language Models.\" The Eleventh International Conference on Learning Representations. 2022.\n\n[2] Shinn, Noah, et al. \"Reflexion: Language agents with verbal reinforcement learning.\" Thirty-seventh Conference on Neural Information Processing Systems. 2023.\n", "category": "Motivation_Analysis"}, {"question": "How does the AgentBench benchmark ensure a balanced representation of coding and non-coding tasks to evaluate language models comprehensively?", "answer": "AgentBench is designed to include a balanced mix of coding and non-coding tasks. Out of 8 tasks, only 3 (Operating System, Database, and Knowledge Graph) involve coding. These tasks are designed to evaluate the agent abilities of language models when they are grounded in coding. The remaining 5 tasks (Digital Card Game, Lateral Thinking, Household, Web Shopping, and Web Browsing) do not involve coding at all. These tasks focus on different aspects such as decision-making, language comprehension, and reasoning, offering a comprehensive evaluation of language models in non-coding contexts. This balance between coding and non-coding tasks ensures a holistic assessment of language models, capturing their versatility and range of capabilities. This approach accurately reflects the diverse applications and potential of contemporary language models.", "category": "Experimental_Setup"}, {"question": "Why was a human expert score not included as a high-score benchmark in the study?", "answer": "A human expert score was not included due to the significant time and cost required. The complete dataset contains about a thousand evaluation entries, with each entry taking 5-20 minutes, amounting to around 200 hours for the entire benchmark. A smaller dev dataset would require about 50 hours. Additionally, some tasks involve complex knowledge(such as Ubuntu operations, SQL commands), requiring high expertise. Collecting reliable human expert data (e.g., 385 people for 95% confidence level and 5% confidence interval) would cost approximately $57,750, which is prohibitive.", "category": "Motivation_Analysis"}, {"question": "How does the focus of the paper's research on multi-turn interactions and code execution differ from the multi-turn interactions in DIAL-SQL and TPTU?", "answer": "1. **TPTU: Large Language Model-based AI Agents for Task Planning and Tool Usage**\n    - TPTU primarily investigates how Large Language Models (LLMs) can be leveraged for task planning and tool usage. Although it encompasses multi-turn interactions, the central aim is not explicitly focused on multi-turn code execution, a key element of the paper's study.\n2. **Text-to-SQL Empowered by Large Language Models: A Benchmark Evaluation**\n    - The primary contribution of Dial-SQL is concerning the selection of examples in the domain of In-Context Learning. This is distinct from the paper's research, which delves deeper into the complexities of multi-turn interactions in practical coding environments.", "category": "Method_Comparison"}, {"question": "What were the criteria for selecting task domains in the benchmark, and how does the current method of scoring tasks address the need for objective quantification of task complexity?", "answer": "1. **Choice of Scenarios**:\n    - It is not feasible to cover every potential scenario in the benchmark. However, well-known, real-world scenarios were carefully selected after research. The decision-making process is grounded in two fundamental principles:\n        - Recognition Criterion: Priority is given to task types universally recognized as pivotal within the language agent community. For instance, benchmarks such as NL2Bash[1] and InterCode[2] emphasize the significance of Operating System (OS) tasks. Similarly, Wob[3], MiniWob++[4], Webshop[5], Mind2Web[6], and WebArena[7] highlight the vital nature of web-based tasks.\n        - LLM Applicability: A crucial factor in the selection process is the suitability of tasks for the existing capabilities of Large Language Models (LLMs), specifically in terms of their ability to handle few-shot text prompts in multi-round interactions. For example, while datasets like Minedojo[2] offer more complex challenges in open-world game environments, they often necessitate visual information processing, which is beyond the scope of current LLM capabilities.\n\n2. **Difficulty Quantification:**\n   Improvements in task settings include:\n-   **Tasks that are objectively controllable:** For many tasks, an objective definition of difficulty within a specific domain can be given. For example, for an MDP with known environment state transition and limited action space, after limiting interaction rounds, the expected score can be calculated in the same way as the score of Curated List in Text World[8] is defined, and the real score can be normalized on this basis.\n    \n-   **Tasks that reflect complex real-world needs:** As the aim is to create diverse tasks across multiple domains, it is infeasible to uniformly quantify difficulty for all considered tasks. For example, tasks in OS operations can span various domains like file manipulation, user management, and network settings, each with inherently different complexity metrics. These tasks cannot be properly assigned difficulty control, or their comprehensiveness to reflect real-world cases would be significantly harmed.\n\n\n[1] Lin, Xi Victoria, et al. \"NL2Bash: A Corpus and Semantic Parser for Natural Language Interface to the Linux Operating System.\" Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). 2018.\n\n[2] Yang, John, et al. \"InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback.\" arXiv preprint arXiv:2306.14898 (2023).\n\n[3] Shi, Tianlin Tim, et al. \"World of bits: an open-domain platform for web-based agents.\" Proceedings of the 34th International Conference on Machine Learning-Volume 70. 2017.\n\n[4] Liu, Evan Zheran, et al. \"Reinforcement Learning on Web Interfaces using Workflow-Guided Exploration.\" International Conference on Learning Representations. 2018.\n\n[5] Yao, Shunyu, et al. \"Webshop: Towards scalable real-world web interaction with grounded language agents.\" Advances in Neural Information Processing Systems 35 (2022): 20744-20757.\n\n[6] Deng, Xiang, et al. \"Mind2Web: Towards a Generalist Agent for the Web.\" arXiv preprint arXiv:2306.06070 (2023).\n\n[7] Zhou, Shuyan, et al. \"WebArena: A Realistic Web Environment for Building Autonomous Agents.\" NeurIPS 2023 Foundation Models for Decision Making Workshop. 2023.\n\n[8]Côté, Marc-Alexandre, et al. \"Textworld: A learning environment for text-based games.\" Computer Games: 7th Workshop, CGW 2018, Held in Conjunction with the 27th International Conference on Artificial Intelligence, IJCAI 2018, Stockholm, Sweden, July 13, 2018, Revised Selected Papers 7. Springer International Publishing, 2019.\n", "category": "Experimental_Setup"}, {"question": "How can the potential bias introduced by using Large Language Models (LLMs) like GPT-3.5-turbo for data augmentation in the database task be addressed, and what evidence supports the claim that the performance advantage of GPT-3.5-turbo over GPT-4 is not due to this bias?", "answer": "The authors acknowledge the potential for bias when using LLMs for data augmentation and have taken steps to minimize it. Taking the data generation process of DB as an example, the augmentation can be divided into two steps: table augmentation and row labeling.\n\n- During the table augmentation phase, the authors ask LLMs to generate similar data entries based on existing rows. After generation, the authors verify the legitimacy of these data and incorporate them into the table. The purpose of this step is to increase the complexity of the task. Since the LLM simply imitates and repeats based on existing entries, this process should not introduce bias.\n- In the INSERT and UPDATE tasks of DB, the paper used LLMs to generate questions and answers. Specifically, LLMs are provided with header information and row data of the table. The LLM is first asked to generate an SQL statement. If the statement is valid and changes the database, this data is retained. On this basis, the LLM is further required to generate an Instruction describing the changes made in the SQL.\n\nIn summary, during the data augmentation phase, as much information as possible is provided to the LLM, enabling it to perform a relatively simple process of generating from a larger to a smaller amount of information. In contrast, during the evaluation phase, limited information is provided to the LLM, requiring it to infer from a smaller to a larger amount of information, a more complex and challenging process. This approach helps to minimize potential biases introduced during the data augmentation stage.\n\nTo add credibility, the authors re-executed a small batch of data labeling for INSERT and UPDATE in DB using claude-2 and re-tested on gpt-4 and gpt-3.5-turbo. The results showed that gpt-3.5-turbo still outperforms gpt-4 in the UPDATE category, as it did originally. This indicates that the superiority of gpt-3.5-turbo over gpt-4 is not due to biases introduced in the data augmentation process, but rather due to other reasons. (Cf. Appendix C.4)", "category": "Method_Mechanics"}, {"question": "Why did the paper choose to omit prompts in its study, and how does this approach impact the experimental results?", "answer": "The majority of tasks designed in this study are intended to be completed within 2000 tokens. In fact, the analysis of interaction trajectories supports this design choice. For completed tasks, an average of 7.95 rounds per task and an average token usage of 2220.1 were observed. Additionally, it was rare for the model to use more than 3000 tokens.\n\nTherefore, if the total token count in an interaction exceeds 3500, it indicates a low likelihood of the task being completed. This approach to omission minimally impacts the experimental results. Moreover, this strategy was consistently applied across all models, ensuring a fair evaluation framework.\n\nFurthermore, another interesting observation is detailed in Appendix J.2.4. It was found that the majority of instances where models failed to complete tasks were due to a tendency to repeat content already generated. This insight further supports the approach, as a summary that condenses only the current state of the task would not necessarily mitigate this issue of content repetition.", "category": "Experimental_Analysis"}]}
{"id": "AL1fq05o7H", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces", "authors": ["Albert Gu", "Tri Dao"], "abstract": "Foundation models, now powering most of the exciting applications in deep learning, are almost universally based on the Transformer architecture and its core attention module. Many subquadratic-time architectures such as linear attention, gated convolution and recurrent models, and structured state space models (SSMs) have been developed to address Transformers' computational inefficiency on long sequences, but they have not performed as well as attention on important modalities such as language.  We identify that a key weakness of such models is their inability to perform content-based reasoning, and make several improvements. First, simply letting the SSM parameters be functions of the input addresses their weakness with discrete modalities, allowing the model to *selectively* propagate or forget information along the sequence length dimension depending on the current token. Second, even though this change prevents the use of efficient convolutions, we design a hardware-aware parallel algorithm in recurrent mode. We integrate these selective SSMs into a simplified end-to-end neural network architecture without attention or even MLP blocks (**Mamba**). Mamba enjoys fast inference (5$\\times$ higher throughput than Transformers) and linear scaling in sequence length, and its performance improves on real data up to million-length sequences. As a general sequence model backbone, Mamba achieves state-of-the-art performance across several modalities such as language, audio, and genomics. On language modeling, our Mamba-1.4B model outperforms Transformers of the same size and matches Transformers twice its size, both in pretraining and downstream evaluation.", "keywords": "Sequence model;language model;state space model;RNN;SSM;S4;Mamba", "qa_pairs": [{"question": "Why did the paper choose not to compare its method with more baselines on the Long Range Arena (LRA) benchmark, which is considered the standard for long sequence understanding?", "answer": "This paper has deliberately focused on more impactful settings than small benchmarks such as LRA. \nSeveral additional points of justification: \n* The emphasis on the importance of pretraining is underscored by recent works such as [Amos], which indicate that **Transformers are competitive with state-of-the-art SSM models on LRA when allowed pretraining**. This supports the paper's claim that small-scale train-from-scratch benchmarks such as LRA may not be the most appropriate for measuring the quality of new foundation model architectures at scale, where thus far only Transformers have been effective. \n* **LRA is not actually long-range compared to the paper's selection of tasks**. The paper's shortest-range task is standard Language Modeling (Figure 4), for which the sequences are of length 8192, which is already longer than (or on the order of magnitude of all LRA tasks). On the other hand, the paper has multiple settings (induction heads, genomics, and audio) involving sequences of length >1000000.\n\n[Amos] “Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors”. https://arxiv.org/abs/2310.02980", "category": "Motivation_Analysis"}, {"question": "Why is memory benchmarking data not included in Section 4.5, and how does Mamba's memory efficiency compare to the most optimized Transformer implementation?", "answer": "The memory usage for Mamba, as with most deep sequence models, is based on the size of the activation tensors. Mamba is highly memory efficient, as demonstrated by additional measurements of training memory requirements for 125M models on a 1 A100 80GB GPU, with batch sizes ranging from 1 to 32.Each batch consists of sequences of length 2048. These measurements show that Mamba's memory requirements are comparable to the most memory-efficient Transformer implementation known, which includes kernel fusion from torch.compile and FlashAttention-2. \n\n| Batch size | Transformer (w/ FlashAttention-2) | Mamba  | \n|------------|-----------------------------------|--------| \n|          1 | 4.6GB                             | 4.8GB  | \n|          2 | 5.2GB                             | 5.8GB  | \n|          4 | 6.9GB                             | 7.3GB  | \n|          8 | 11.5GB                            | 12.3GB | \n|         16 | 20.7GB                            | 23.1GB |\n|         32 | 34.5GB                            | 38.2GB |", "category": "Method_Comparison"}, {"question": "What are the detailed setups of Figure 8 left, including model layers, model sizes, and details of the convolution?", "answer": "The detailed setups for the left side of Figure 8 are reported in Appendix F.5. The measurements include only the core operations (scan, convolution, attention), not the full block (e.g., QKV projections of attention are excluded). There is no concept of layers. The model dimension is 1024, with sequence lengths ranging from 512 to 524288. The convolution is an FFT-based convolution using PyTorch primitives.", "category": "Concept_Understanding"}, {"question": "Does Mamba possess the ability to generalize lengths, similar to attention-based Transformer models like T5 and Alibi, by being trained on shorter sequences and tested on longer sequences?", "answer": "The paper includes length extrapolation results for the Induction Heads task, which is hypothesized to be deeply related to generalization abilities of language models [Anth].  Mamba demonstrates length generalization ability. As shown in Figure 3, Mamba can extrapolate from sequences of length 256 to length 1000000, whereas other models, including those with relative positional embeddings, do not extrapolate beyond length 512. \n\n[Anth] “​​In-context Learning and Induction Heads.” https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html", "category": "Experimental_Analysis"}, {"question": "Why was H3 not directly compared with Mamba in the experiments, and how does Mamba’s performance compare to H3 in downstream evaluations, given the following context: From Table 4 in [1], the ppl of H3 is 8.8 (125M), 7.1 (355M), and 6.0 (1.3B) on the Pile dataset, which are considerably better than Mamba’s?\n\n[1] https://arxiv.org/pdf/2212.14052.pdf", "answer": "H3 is one of the main baselines and is featured prominently in almost all results in the paper. * The synthetic tasks, Selective Copying (Table 1) and Induction Heads (Figure 3). * For Chinchilla scaling laws (Figure 4), the paper used a much stronger version of H3 that called H3++ (details in Appendix F.2.1). * For genomics, the paper compares against Hyena (Figure 5), which is the successor of H3. * The paper has since downloaded the pretrained H3 models from [1] and ran the suite of evaluations.  As mentioned in the caption of the paper's Table 2, perplexity scores are only comparable when trained with the same tokenizer. The H3 paper used the GPT2 tokenizer and is not comparable to The paper which uses the now-standard NeoX tokenizer. Instead, the downstream evaluations are directly comparable, where Mamba performs significantly better than H3.", "category": "Experimental_Analysis"}, {"question": "How does Mamba compare to other models such as S4-diagonal[1], SGConv[2], MEGA[3], SPADE[4], and efficient Transformer models (e.g., [5]) in terms of model performance and efficiency, particularly on the WikiText-103 benchmark? Given that the motivation of Mamba is to address the drawbacks of recurrent models while improving the efficiency of attention-based models, [1,5] follow the same direction.\n\n[1] https://arxiv.org/pdf/2203.14343.pdf\n[2] https://arxiv.org/pdf/2210.09298.pdf\n[3] https://arxiv.org/pdf/2209.10655.pdf\n[4] https://arxiv.org/pdf/2212.08136.pdf\n[5] https://arxiv.org/pdf/2202.10447.pdf", "answer": "Mamba strongly outperforms all suggested models and many more on WikiText-103. In the same setting as the Hyena paper[6, Table 4.3], Mamba improves over the paper's Transformer baseline by 1.7 ppl and the original baseline Transformer by 2.3 ppl. Specifically, Mamba achieves a perplexity of 16.3 on WikiText-103, which is significantly better than the paper's Transformer baseline (18.0 ppl) and other subquadratic architectures.\n\n| Model (125M)   | WikiText-103 ppl | $\\Delta$ to Transf. | \n|----------------|------------------|-----| \n| Transformer    | 18.6             | 0.0 | \n| Hybrid H3      | 18.5             | -0.1 | | Performer      | 26.8             | 8.2 | \n| Reformer       | 25.6             | 7.0 | \n| AFT-conv       | 28.2             | 9.6 | \n| Linear Attn.   | 25.6             | 7.0 | \n| Hyena          | 18.5             | -0.1 | \n| Transf. (ours) | 18.0             | -0.6 | \n| **Mamba** (the paper) | **16.3**         | **-2.3** | \n\n[6] “Hyena Hierarchy: Towards Larger Convolutional Language Models.” https://arxiv.org/abs/2302.10866", "category": "Method_Comparison"}, {"question": "Does Mamba have a quadratic memory requirement during training, similar to Transformers?", "answer": "No, Mamba’s memory requirement scales linearly in sequence length. The memory bottleneck is the size of the activation tensors, which is true of most deep sequence models, including Transformers when using optimized implementations like FlashAttention. Mamba is memory efficient.", "category": "Method_Comparison"}, {"question": "How does the scaling behavior of Mamba beyond 1.4B parameters compare to Transformers at 10B scales?", "answer": "Mamba has been scaled to 2.7-2.8B parameters, where it strongly outperforms open-source Transformers of the same size and matches Transformers of the next larger size (7B parameters). Compared to open-source models at the 3B scale trained to the same token count (300B), Mamba performs better on every evaluation. It is also comparable to models at the 7B scale, such as OPT, Pythia, and RWKV, achieving the best average score and best or second-best score on every benchmark. For example, Mamba-2.8B achieves an average score of 63.3, outperforming Pythia-2.8B (59.1), RWKV-3B (59.6), OPT-6.7B (62.9), and RWKV-7.4B (62.5).", "category": "Method_Comparison"}, {"question": "What are the results of the length extrapolation experiments for the Induction Heads task, and how does Mamba compare to other models in terms of extrapolation capabilities?", "answer": "The paper includes length extrapolation results for the Induction Heads task, which is hypothesized to be deeply related to the generalization abilities of language models. Figure 3 shows that Mamba can extrapolate from sequences of length 256 to length 1000000, while no other model (including Attention with variants of relative positional embeddings) extrapolates beyond length 512. ", "category": "Method_Comparison"}, {"question": "How sensitive are the results of the input selection mechanism to hyperparameters such as the projection rank $R$, the state size $N$, and the initialization, and what evidence supports this?", "answer": "The results are not highly sensitive to hyperparameters such as the projection rank $R$, the state size $N$, and the initialization. Regarding the state size $N$, larger states can perform better but slow down the model; Prior works on SSMs have had many variants of the initialization. Ablations in Tables 7, 8, and 9 of the paper show that increasing $R$ and $N$ provides small improvements for a modest increase in parameters but does not significantly affect performance. All experiments were conducted without tuning these hyperparameters, with $R$ fixed to $1/16$ of the model dimension, $N$ fixed to 16, and the initialization fixed. This suggests that Mamba can be a promising drop-in replacement for attention in general foundation models.", "category": "Experimental_Analysis"}, {"question": "What is the complete ablation analysis of Table 6?", "answer": "This table includes the full ablations:\n\n| Selective $\\Delta$ | Selective $B$ | Selective $C$ | Perplexity |\n|--------------------|---------------|---------------|------------|\n|                    |               |               | 10.93      |\n|                    | +             |               | 10.15      |\n|                    |               | +             | 9.98       |\n| +                  |               |               | 9.81       |\n|                    | +             | +             | 9.49       |\n| +                  | +             |               | 9.21       |\n| +                  |               | +             | 9.07       |\n| +                  | +             | +             | 8.71       |", "category": "Experimental_Exposition"}, {"question": "How do the FLOPs of Transformers and SSMs compare at different sequence lengths, and does this difference affect the applicability of Chinchilla's scaling laws to the paper's model's performance?", "answer": "The FLOPs of Transformers and SSMs do not differ significantly at shorter sequence lengths (e.g., length 2048), so the standard Chinchilla scaling laws apply. However, at longer sequence lengths, the FLOPs of Transformers and SSMs do differ, as shown in Figure 4 (Left) and 4 (Right). ", "category": "Experimental_Analysis"}, {"question": "How were the pretrained H3 models evaluated, and what are the comparative performance results between H3 and Mamba models on language evaluations?", "answer": "The pretrained H3 models (sizes 125M-2.7B parameters) were downloaded from [1] and evaluated using a suite of language tasks. Mamba, designed to address H3's linear time invariance weakness, outperforms H3 on all language evaluations. \n\n |Model           |Lambada |HellaSwag |PIQA |Arc-E |Arc-C |WinoGrande |Avg  | \n|----------------|--------|----------|-----|------|------|-----------|-----| \n|Hybrid H3-130M  |25.8    |31.7      |64.2 |44.4  |24.2  |50.6       |40.2 | \n|Mamba-130M      |44.2    |35.2      |64.5 |48.0  |24.2  |52.3       |44.7 | \n|Hybrid H3-360M  |48.0    |41.5      |68.1 |51.4  |24.7  |54.1       |48.0 | \n|Mamba-370M      |55.6    |46.5      |69.5 |55.1  |28.0  |55.3       |51.7 | \n|Hybrid H3-1.3B  |49.6    |52.6      |71.3 |59.2  |28.1  |56.9       |53.0 | \n|Mamba-1.4B      |65.0    |59.1      |74.2 |65.5  |32.8  |61.5       |59.7 | \n|Hybrid H3-2.7B  |55.7    |59.7      |73.3 |65.6  |32.3  |61.4       |58.0 | \n|Mamba-2.8B      |69.2    |66.1      |75.2 |69.7  |36.3  |63.5       |63.3 |\n\n Notably, Mamba's pure linear-time layers significantly surpass the hybrid H3 models using quadratic attention.\n\n[1] 'Hungry Hungry Hippos: Towards Language Modeling with State State Models.' https://arxiv.org/abs/2212.14052", "category": "Experimental_Exposition"}, {"question": "How does the 3B parameter Mamba model compare to other open-source models at the same scale (3B) and larger scales (7B) in terms of evaluation results?", "answer": "The 3B parameter Mamba model outperforms all other open-source models at the 3B scale trained on the same token count (300B) on every evaluation metric. Additionally, Mamba (2.8B) is comparable to models at the 7B scale, achieving the best average score and either the best or second-best score on every benchmark when compared to OPT, Pythia, and RWKV (7B).\n\n |Model           |Lambada |HellaSwag |PIQA      |Arc-E    |Arc-C    |WinoGrande |Avg      | \n|----------------|--------|----------|----------|---------|---------|-----------|---------| \n|GPT-Neo 2.7B    |62.2    |55.8      |72.1      |61.1     |30.2     |57.6       |56.5     | \n|Hybrid H3-2.7B  |55.7    |59.7      |73.3      |65.6     |32.3     |61.4       |58.0     | \n|OPT-2.7B        |63.6    |60.6      |74.8      |60.8     |31.3     |61.0       |58.7     | \n|Pythia-2.8B     |64.7    |59.3      |74.0      |64.1     |32.9     |59.7       |59.1     | \n|RWKV-3B         |63.9    |59.6      |73.7      |67.8     |33.1     |59.6       |59.6     | \n|**Mamba-2.8B**  |**69.2**|*66.1*    |*75.2*    |**69.7** |*36.3*   |*63.5*     |**63.3** | \n|OPT-6.7B        |*67.7*  |**67.2**  |**76.3**  |65.6     |34.9     |**65.5**   |*62.9*   | \n|Pythia-6.9B     |67.1    |64.0      |*75.2*    |67.3     |35.5     |61.3       |61.7     | \n|RWKV-7.4B       |67.2    |65.5      |76.1      |*67.8*   |**37.5** |61.0       |62.5     |", "category": "Method_Comparison"}]}
{"id": "tgjGR7eY5H", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "RL4CO: a Unified Reinforcement Learning for Combinatorial Optimization Library", "authors": ["Federico Berto", "Chuanbo Hua", "Junyoung Park", "Minsu Kim", "Hyeonah Kim", "Jiwoo Son", "Haeyeon Kim", "Joungho Kim", "Jinkyoo Park"], "abstract": "Deep reinforcement learning offers notable benefits in addressing combinatorial problems over traditional solvers, reducing the reliance on domain-specific knowledge and expert solutions, and improving computational efficiency. Despite the recent surge in interest in neural combinatorial optimization, practitioners often do not have access to a standardized code base. Moreover, different algorithms are frequently based on fragmentized implementations that hinder reproducibility and fair comparison. To address these challenges, we introduce RL4CO, a unified Reinforcement Learning (RL) for Combinatorial Optimization (CO) library. We employ state-of-the-art software and best practices in implementation, such as modularity and configuration management, to be flexible, easily modifiable, and extensible by researchers. Thanks to our unified codebase, we benchmark baseline RL solvers with different evaluation schemes on zero-shot performance, generalization, and adaptability on diverse tasks. Notably, we find that some recent methods may fall behind their predecessors depending on the evaluation settings. We hope RL4CO will encourage the exploration of novel solutions to complex real-world tasks, allowing the community to compare with existing methods through a unified framework that decouples the science from software engineering. We open-source our library at https://anonymous.4open.science/r/rl4co-iclr.", "keywords": "Reinforcement Learning;Neural Combinatorial Optimization;Combinatorial Optimization;Library;Benchmark", "qa_pairs": [{"question": "What are the theoretical and practical reasons that policy optimization methods outperform value-based methods in neural combinatorial optimization (NCO)?", "answer": "Policy optimization methods outperform value-based methods in NCO due to several reasons: \n(1) Value-based methods require accurate reward prediction, which is challenging in CO problems like routing, where small decision variations can lead to significant reward differences. \n\n(2) The reward is sparse, as it is only known after the full solution is decoded, hindering value-based methods' accuracy. \n\n(3) Using policy gradients (PG) as REINFORCE methods has theoretical advantages over Actor-Critic (AC) methods in the paper's setting. Accurately estimating $R(x,a)$ is important to train the NCO solvers with high performance. According to the estimation methods of $R(x,a)$, AC and PG can be differentiated. The AC methods (AM-PPO, AM-Critic) typically rely on the learned critics $V_\\phi(x)$ to estimate $R(x,a)$ in the spirit of temporal-difference (TD) methods. TD methods can be sample-efficient and typically exhibit lower variance on estimating $R(x,a)$ compared to PG. However, $V_\\phi(x)$ is known to be a biased estimator of $R(x,a)$. On the other hand, the PG method estimates $R(x,a)$ via the Monte-Carlo method, which is known as an unbiased estimator of $R(x,a)$, but can exhibit higher variance on estimating $R(x,a)$ than AC methods. Typically, the paper's benchmarked algorithms, such as AM, POMO, and Sym-NCO, utilize proper baselines that can reduce the negative effect of larger variance on estimating $R(x,a)$.\n\n(4) Generating training samples in NCO is simpler than in other RL applications, as the environment rollout involves plugging the constructed solution into the CO problem. \n\nThus, PG methods with proper baselines(e.g., AM, POMO, Sym-NCO) tend to perform better than AC methods(e.g., AM-PPO, AM-Critic) in NCO settings, while benefiting from reduced variance (due to baselines) and being unbiased in estimating $R(x,a)$.", "category": "Method_Comparison"}, {"question": "How do the constructive and improvement methods in RL algorithms for CO differ, and why does the paper focus on constructive methods, given that there could be different ways to formulate the CO as an RL problem, not limited to the one outlined in Equation (1) - (3). For example, neighborhood search first constructs an initial feasible solution and optimizes the current solution at each step [1,2].\n\n[1] Learning to Perform Local Rewriting for Combinatorial Optimization. Xinyun Chen, Yuandong Tian. NeurIPS 2019.\n\n[2] Learning Large Neighborhood Search Policy for Integer Programming. Yaoxin Wu, Wen Song, Zhiguang Cao, Jie Zhang. NeurIPS 2021.", "answer": "RL algorithms for CO can be generally divided into 1) constructive and 2) improvement. Constructive methods usually generate the solution from scratch, which this paper focuses on. These methods are generally faster but the quality of solutions can be lower than that of improvement methods. In construction methods, inference techniques (such as greedy multistart, sampling, and augmentation) focus on selecting the best solution from several generated ones; search methods, like EAS, focus on efficiently improving the quality of generated solutions by updating the model at test time. On the other hand, improvement methods, as demonstrated by the papers [1,2], learn to improve existing solutions. Usually, they obtain a better solution, but are slower and may be less general (i.e., specialized to a class of problems such as routing). The reason why this paper focuses first on construction methods is that they are more generally applicable to various CO problems, such as routing and electronic design automation, while most learned improvement approaches are specific to a single class of problems, such as integer programming [2] or routing problems [3].\n\n[3] Hottung, André, and Kevin Tierney. \"Neural large neighborhood search for the capacitated vehicle routing problem.\" ECAI 2020 (2020).", "category": "Motivation_Analysis"}, {"question": "Does the proposed RL4CO framework cover other types of combinatorial optimization problems beyond TSP/VRP and graph-based CO, such as bin-packing, job scheduling, and mixed integer programming?", "answer": "RL4CO is designed to cover a wide variety of combinatorial optimization problems. It includes implementations for electronic design automation and scheduling problems like the Single Machine Total Weighted Tardiness Problem (SMTWTP) and the Flexible Flow Shop Problem (FFSP). Additionally, it is easily extensible, as demonstrated by recent extensions[1] to the max cover problem and facility location.\n\n[1]Anonymous. “Unsupervised combinatorial optimization under complex conditions: Principled objectives and incremental greedy derandomization”. Under submission at ICLR2024 (2023).", "category": "Method_Disambiguation"}, {"question": "Why were the experiments conducted with TSP/VRP sizes limited to 50 nodes instead of larger instances?", "answer": "In the sense that most papers in the NCO area pre-train models on 100 nodes. While pre-training on several hundred nodes and even more is possible, especially thanks to advancements the paper has included as FlashAttention, this would result in burdensome computational costs that would be prohibitive with the paper's limited resources - for instance, Sym-NCO took the authors around 2 weeks to train on a single problem of 100 nodes. It is believed that pre-training on small instances and generalizing to larger ones is an important topic in NCO (for instance, some recent supervised approaches can effectively scale to several thousands of nodes such as [1] - as such, it would be an interesting avenue of research to make RL scale better as well). Another approach is adaptation, as the paper did with AS and EAS (up to 1000 nodes in the paper's experiments from the pre-trained models on 50 nodes).\n\n[1]Luo, Fu, et al. \"Neural Combinatorial Optimization with Heavy Decoder: Toward Large Scale Generalization.\" NeurIPS2023 (2023).", "category": "Motivation_Analysis"}, {"question": "Does RL4CO exclusively address routing-related problems, or does it encompass a broader range of problems?", "answer": "RL4CO is not confined to routing problems only. It includes decap placement problems from electronic design automation (EDA) in Appendix B, and the authors have recently introduced library scheduling problems such as the Single Machine Total Weighted Tardiness Problem (SMTWTP) and the Flexible Flow Shop Problem (FFSP). Additionally, RL4CO is extensible to other problems, as demonstrated by a recent paper[1] that extended its implementation to the max cover problem and facility location.\n\n[1] Anonymous. “Unsupervised combinatorial optimization under complex conditions: Principled objectives and incremental greedy derandomization”. Under submission at ICLR (2024).", "category": "Method_Disambiguation"}, {"question": "What insights or novel contributions does the paper provide regarding the organization and application of autoregressive constructive methods for combinatorial optimization, and how does it address the redundancy and management challenges in existing implementations?", "answer": "In terms of software organization, the authors' fundamental insight is to divide and conquer the problems (in terms of data, environment, model, training) into more manageable subproblems. For instance, while several practical implementations in the NCO community separate different problems into separate schemes - one can think about different folders each with its own problem to solve but with mostly redundant code, which makes management and progressive development hard, especially for new researchers - the authors decided to rely on unified structures. In the case of data, this was e.g. thanks to TensorDict, that allows for batched operations on named tensors, greatly simplifying passing around instances. In the case of environment, thanks to TorchRL the paper created a blueprint for vectorized CO environments that can run on GPU, thus greatly alleviating the typical CPU simulation bottleneck typical of RL. As for the models, given common structures in the encoding-decoding architecture, the paper managed to modularize the policies as described in Figure 3.2, with the main idea being to abstract away problem-specific modules from common structures as encoder and decoder layers. For the training, the paper uses PyTorch Lightning to reorganize the flow from input to output (in training, from data to loss function to optimize the parameter). Many of these procedures are common among different RL algorithms (e.g. in A2C and REINFORCE with different baselines, this means modularizing the baselines). In terms of research, the authors have identified several pitfalls that can be encountered when training CO solvers, such as out-of-distribution generalization, that could be alleviated by introducing several inductive biases or exploring different directions as highlighted in the paragraph 'Future Directions in RL4CO' in Section 4.3.", "category": "Method_Disambiguation"}, {"question": "Why does the comparison of methods for routing problems (TSP, CVRP, etc.) not include non-autoregressive methods[1,2,3] that have achieved better performance in terms of faster inference time, lower optimal gap, and scalability to larger instances?\n\n[1] Sun Z, Yang Y. Difusco: Graph-based diffusion solvers for combinatorial optimization[J]. arXiv preprint arXiv:2302.08224, 2023.\n\n[2] Qiu R, Sun Z, Yang Y. Dimes: A differentiable meta solver for combinatorial optimization problems[J]. Advances in Neural Information Processing Systems, 2022, 35: 25531-25546.\n\n[3] Sun H, Goshvadi K, Nova A, et al. Revisiting Sampling for Combinatorial Optimization[J]. ICML, 2023", "answer": "Non-autoregressive methods, such as the supervised DIFUSCO [1], the RL-based DIMES [2], and new sampling techniques as iSCO [3], can scale well to thousands of nodes but often suffer from poor generalizability to more constrained and practical problems. These methods are typically limited to NP-complete problems, which are relatively easier than NP-hard problems like vehicle routing. DIFUSCO is restricted to NP-complete problems like TSP, DIMES's extension to more complex constraints is unclear, and iSCO's formulation may fail when finding feasible solutions becomes nontrivial. Autoregressive approaches, such as those found in the now ubiquitous large language models, are more general and can deal with a wider variety of problems, that include, in combinatorial optimization, tasks such as routing, electronic design automation, scheduling, and ultimately more complex practical problems.", "category": "Motivation_Analysis"}, {"question": "Is the scope of RL4CO limited to routing-related problems, as suggested by the title?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the proposed RL4CO framework include coverage for CO problems beyond TSP/VRP and graph-based CO, such as bin-packing, job scheduling, and mixed integer programming?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "iwd9sQWnXb", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting", "authors": ["Zhen Qin", "Rolf Jagerman", "Kai Hui", "Honglei Zhuang", "Junru Wu", "Le Yan", "Jiaming Shen", "Tianqi Liu", "Jialu Liu", "Donald Metzler", "Xuanhui Wang", "Michael Bendersky"], "abstract": "Ranking documents using Large Language Models (LLMs) by directly feeding the query and candidate documents into the prompt is an interesting and practical problem. However, there has been limited success so far, as researchers have found it difficult to outperform fine-tuned baseline rankers on benchmark datasets. We analyze pointwise and listwise ranking prompts used by existing methods and argue that off-the-shelf LLMs do not fully understand these challenging ranking formulations. In this paper, we propose to significantly reduce the burden on LLMs by using a new technique called Pairwise Ranking Prompting (PRP). Our results are the first in the literature to achieve state-of-the-art ranking performance on standard benchmarks using moderate-sized open-sourced LLMs. On TREC-DL2020, PRP based on the Flan-UL2 model with 20B parameters outperforms the previous best approach in the literature, which is based on the blackbox commercial GPT-4 that has 50x (estimated) model size, by over 5% at NDCG@1. On TREC-DL2019, PRP performs favorably with supervised models and is only inferior to the GPT-4 solution among LLM-based methods on the NDCG@5 and NDCG@10 metrics, while outperforming other LLM-based solutions, such as InstructGPT which has 175B parameters, by over 10% for all ranking metrics. By using the same prompt template on seven BEIR tasks, PRP beats supervised baselines and outperforms the blackbox commercial ChatGPT solution by 4.2\\% and pointwise LLM-based solutions by over 10\\% on average NDCG@10. Furthermore, we propose several variants of PRP to improve efficiency and show that it is possible to achieve competitive results even with linear complexity. We also discuss other benefits of PRP, such as supporting both generation and scoring LLM APIs, as well as being insensitive to input ordering.", "keywords": "text ranking;large language models", "qa_pairs": [{"question": "How does the novelty of the proposed method, particularly in terms of pair-wise prompting and sliding window algorithms, differ from existing works in the field of information retrieval and LLM-based searching, given that the pair-wise ranking prompting strategy has also been explored in existing works [1]. The sliding window strategy has also been explored in works about searching using LLMs [2]?\n\n1] Dai et al. Uncovering ChatGPT’s Capabilities in Recommender Systems.\n\n[2] Sun et al. Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking Agents.", "answer": "1. The paper's work is the first to show compelling ranking performance on the important text relevance ranking task with open-sourced, moderated sized LLMs, which is only possible with the paper's proposed methodology and has not been done in the literature. This is novelty, which should not be judged in terms of simplicity. \n\n2. Paper [1] is concurrent work, which was not presented until RecSys (Sep 2023). Paper [1] is substantially different from the text relevance ranking task, a critical problem in web search. The paper discussed in Related Work how ‘pairwise’ is a general paradigm used in different ways for LLMs and argued that novelty should not be judged by general paradigms such as pointwise / pairwise. Also, it is non-trivial to apply pairwise comparisons to practical ranking tasks, where the paper studies various strategies with different trade-offs. \n\n3. Sliding windows are also a general tool to show the proposed method can be made more practical. Pairwise prompting avoids production of unexpected outputs such as repetition and missing from listwise prompting and naturally mitigates prompt length limit and position bias in the prompt. Thus, though RankGPT and the paper's both leverage sliding windows, their properties are significantly different due to the computation unit - note that the RankGPT listwise prompting is completely unusable with moderate LLMs as discussed in Sec 2.2, even if sliding window is used.\n\n[1]'Uncovering ChatGPT’s Capabilities in Recommender Systems', https://arxiv.org/abs/2305.02182", "category": "Method_Comparison"}, {"question": "How does the absence of the assumption of a total order, particularly transitivity, affect the validity of the output and the total number of API calls in the context of HeapSort, as reported in Table 1?", "answer": "It is unrealistic to expect a (LLM-based) pairwise ranker to always make correct predictions, because without knowing the ground-truth relevance of each document it is bound to be noisy. What most pairwise rankers can do is to be as 'probably correct' as possible. When the ranker is not 100% correct, the ranking is not guaranteed to be 100% correct as well. However, an 'approximately correct' ranking is still valuable, and a 'probably correct' pairwise ranker with sliding window/heap sort can still deliver an 'approximately correct' ranking, as shown in Table 2 and Table 3 on various benchmark datasets. ", "category": "Experimental_Exposition"}]}
{"id": "6LNTSrJjBe", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models", "authors": ["Andy Zhou", "Kai Yan", "Michal Shlapentokh-Rothman", "Haohan Wang", "Yu-Xiong Wang"], "abstract": "While large language models (LLMs) have demonstrated impressive performance on a range of decision-making tasks, they rely on simple acting processes and fall short of broad deployment as autonomous agents. We introduce LATS (Language Agent Tree Search), a general framework that synergizes the capabilities of LLMs in planning, acting, and reasoning. Drawing inspiration from Monte Carlo tree search in model-based reinforcement learning, LATS employs LLMs as agents, value functions, and optimizers, repurposing their latent strengths for enhanced decision-making. What is crucial in this method is the use of an environment for external feedback, which offers a more deliberate and adaptive problem-solving mechanism that moves beyond the limitations of existing methods. Our experimental evaluation across diverse domains, such as programming, HotPotQA, and WebShop, demonstrates the superiority and versatility of LATS for both reasoning and acting. In particular, LATS achieves 94.4\\% for programming on HumanEval with GPT-4 and an average score of 75.9 for web browsing on WebShop, demonstrating the effectiveness and generality of our method.", "keywords": "large language models;agent;reasoning;decision-making", "qa_pairs": [{"question": "How does the performance of WebGUM[1], a finetuned language model agent, compare to LATS and other gradient-free techniques in terms of SR and average score?\n\n[1] https://arxiv.org/abs/2305.11854", "answer": "WebGUM achieves a higher SR (45.0) compared to LATS and other gradient-free techniques, but it uses a different base language model and involves task-specific fine-tuning, making it an unfair comparison. However, LATS outperforms WebGUM in terms of average score (75.9 vs 67.5) despite being gradient-free.", "category": "Method_Comparison"}, {"question": "What are the key novelties of LATS that distinguish it from previous work like RAP and Reflexion?", "answer": "Although previous work has investigated either search algorithms or self-reflection, the combination of both has not been explored. In contrast, LATS synergistically integrates them, and more importantly, introduces unique and advanced methods to leverage search algorithms and self-reflection, and makes non-trivial adaptation of tree-based search with external feedback. Based on these, the paper has made significant improvements across multiple testbeds, including 94.4 pass@1 accuracy with GPT-4 on HumanEval and performance comparable to gradient-based fine-tuning methods on WebShop. More concretely, LATS is distinguished from prior work through the following key novelties: \n1. Paradigm contributions. The paper is the first to explore the use of search algorithms with LLM agents for decision-making tasks. Previous approaches like ReAct and Reflexion cannot sample and select more than one state at every step, limiting exploration in decision-making environments; ToT and RAP do not explore decision-making tasks. It is non-trivial to shift from non-interactive tasks to interactive ones, because an updated node, prompt, and more importantly, search algorithm must be designed to incorporate external input – an achievement  the paper has successfully accomplished. The paper's interactive tasks such as programming is verified. \n\n2. The paper's LATS improves the use of MCTS. Compared to RAP, LATS also improves the use of MCTS. In RAP, states are scored with a value function based on self-consistency, while The paper directly prompts the LM to score the state with justification. The paper's design is more effective in decision-making, which enables the state to also incorporate an external observation that adds context for a more accurate score. Additionally, LATS samples states through environment interaction, eliminating the need to use the LM as a world model as in RAP. This reduces the risk of error and makes LATS a general LLM framework. \n\n3. The paper's LATS improves the use of self-reflection. Compared to the self-reflection in Reflexion, LATS leverages semantic feedback and failed trajectories in additional ways. These are also added to the context of LM for the value function, thereby refining subsequent state evaluations in addition to the base decision-making. \n\n4. The paper's adaptation of search algorithms to external feedback is not trivial, as simply adapting existing tree-based methods, such as ToT and RAP, to external feedback hurts the performance. The paper designs a version of ToT and RAP that can incorporate environment feedback in HotPotQA using the ReAct prompt, and find that the new version of ToT and RAP generally performs worse than the reasoning-only setting of HotPotQA (0.55 vs. 0.49 for ToT, and 0.6 vs. 0.54 for RAP). This result is because the information retrieval setting is harder than the multihop question-answering setting, and it highlights that the paper's efforts of adapting search algorithms to decision-making scenarios is non-trivial. ", "category": "Method_Comparison"}, {"question": "How does the token consumption of LATS compare to other baselines such as CoT-SC, RAP, and ToT, and what are the theoretical and empirical reasons for these differences?", "answer": "Theoretically, LATS consumes more tokens than CoT-SC but has the same token consumption as RAP and ToT. This is because LATS expands with multiple samples per step, leading to an overall token consumption that is $n$ times higher than CoT-SC/ReAct with $k$ attempts, where $n$ is the number of samples per expansion. However,  as the token cost of other tree-structured search methods (RAP and ToT) is also parameterized by the number of attempts $k$ and the number of nodes expanded $n$, it means that LATS, RAP, and ToT have the same sample complexity, i.e., same asymptotic token cost. The sample complexity is summarized in the table below. \n\nEmpirically, LATS expands fewer nodes upon success than RAP and ToT, which means that it is cheaper. To count token consumption more accurately, the average number of nodes expanded upon success on HotPotQA is compared for LATS, RAP, and ToT. It is found that LATS not only has better performance but also tends to solve questions with fewer nodes (66.65) sampled than RAP (70.60) or ToT (84.05) with the same sample complexity, indicating a more effective, thus cheaper, search. The result is also shown in the table below.\n\n| Method             | Performance on HotPotQA (↑) | Sample Complexity (↓) | Average # of Nodes (↓) |\n| :----------------- | :-------------------------- | :--------------------- | :--------------------- |\n| ReAct (best k = 250) | 0.42                        | O(k)                   |                       |\n| CoT-SC (n = 1, k = 250) | 0.40                        | O(k)                   |                      |\n| LATS (n = 1, k = 50)  | 0.48                        | O(k)                   |                       |\n| ToT - ReAct (n = 5, k = 50) | 0.49                       | O(kn)                  | 84.05                  |\n| RAP - ReAct (n = 5, k = 50) | 0.54                       | O(kn)                  | 70.60                  |\n| LATS (n = 5, k = 50)  | 0.61                       | O(kn)                  | 66.65                  |\n", "category": "Method_Comparison"}, {"question": "Hasve the paper incorporated a baseline with external environment feedback in their ablation study, and if so, what are the results and findings from this incorporation?", "answer": "The authors have incorporated baselines with external environment feedback. They designed versions of ToT and RAP that can incorporate environment feedback using the ReAct prompt, denoted as ToT (ReAct) and RAP (ReAct). It is important to note that adapting RAP to this setting is non-trivial, as RAP is designed for scenarios where the LM can be repurposed as a world model. To address this issue, the authors follow the strategy in the paper's proposed LATS and directly use environmental interactions during sampling, but keep the original value function of RAP. The result is shown below :\n\n|Method | HotPotQA|\n| ---|---|\n| ToT (ReAct) | 0.49 |\n|ToT | 0.55 |\n|RAP (ReAct) | 0.54 |\n|RAP | 0.60 |\n|LATS (DFS) | 0.53|\n| LATS (No self-reflection) | 0.56|\n| LATS | 0.61|\n| LATS (CoT + ReAct) | 0.71 |\n\nThe results show that: 1) The DFS version of LATS outperforms ToT (ReAct), indicating the importance of self-reflection; 2) The version of LATS without self-reflection shows improvement over RAP (ReAct), reflecting the importance of the improved value function in MCTS; 3) Baselines generally perform worse in the acting-based setting compared to the reasoning-only setting(0.55 vs. 0.49 for ToT and 0.6 vs. 0.54 for RAP), indicating the challenge of adapting search algorithms to decision-making scenarios.", "category": "Experimental_Exposition"}, {"question": "How does the scalability of LATS compare to other methods like RAP/ToT and ReAct, and what mechanisms allow it to handle large-scale problems?", "answer": "LATS is scalable at least on par with RAP/ToT while outperforming them. Additionally, the sample size $n$ for expansion provides a natural, linear trade-off between computational cost and performance, which can be decreased for large-scale tasks with limited resources. Specifically, when $n=1$, LATS is as scalable as ReAct for the same number of trajectories and still performs better.", "category": "Method_Comparison"}, {"question": "How does the evaluation of the value function in LATS address the concern that LLMs may not be good at self-critique, as highlighted in Huang et al. 2023[1]?\n\n[1]Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. arXiv:2310.01798, 2023.", "answer": "The evaluation of the value function in LATS addresses this concern by providing the LM with access to an external environment, such as test case feedback in programming and API feedback in HotPotQA. This external feedback ensures that the LM does not rely solely on itself for self-correction, which is an improvement over methods like ToT/RAP and aligns with the findings of Huang et al. 2023 that LLMs are not good at evaluating their own outputs without external feedback.", "category": "Method_Disambiguation"}, {"question": "How does the HotpotQA environment provide feedback during the MCTS search process, and does it involve using ground-truth answers?", "answer": "The HotpotQA environment provides feedback through the API, including whether the answer was correct. This feedback mechanism matches the setting used in Reflexion and ReAct. The use of correctness labels is consistent across all methods. Additionally, the HotpotQA environment is designed to evaluate information retrieval in decision-making settings, making it reasonable to provide feedback based on ground truth or task completion. Note that this might be different from the standard QA setting", "category": "Experimental_Setup"}, {"question": "How is it possible to back up to an earlier step in a real application without an environment model?", "answer": "The method is model-free, and the inspiration from model-based RL lies only in the use of MCTS. Backing up in LLM decision-making in most settings is simply prompting with previous states of the frozen LLM. States in decision-making with LLM come from three sources: task context / prompt (initial state, unchanged during the task), thoughts / reflection (generated by the LLM itself), and external APIs. In LLM benchmarks, the calls of external APIs do not affect each other (e.g., the searches in database are independent in HotpotQA, evaluations are independent in programming, and the xmls can be reset conveniently by navigating to the previous website in Webshop.) Thus, it is easy to back up in LLM decision-making without an environment model – simply store the prior history of text and modify the input accordingly when the search needs to return to an earlier step.", "category": "Method_Mechanics"}, {"question": "How do the distinct properties of emerging environments for LM agents, particularly regarding back-up capabilities, differentiate them from traditional RL environments, and why does this distinction make the arguments about model-free approaches and back-up actions less applicable in the context of LM agents?", "answer": "The newly emerging environments in which LM agents operate, which the paper focuses on, differ from traditional RL environments. These two types of environments do have distinct properties, especially regarding backup. In fact, explicitly leveraging such backup capability to propose a principled approach is a key contribution of the paper's work.\n\n**Q1. Without environment models, one must roll out with real environments where backup is not feasible.**\nThis is not true, at least for the environments the paper considers. The paper's experiments cover API-calling / tool-use, programming, and web-browsing. These are core benchmarks for LM agents precisely due to their practical value, and LATS can be directly deployed in real-world counterparts. In WebShop, the environment can be reset to a prior state by changing the XML of the website; and for programming and API-usage, the actions of the LM do not directly affect the environment, so backing up is a matter of changing the prompt.\n\n**Q2. Being able to backup is not the property of decision-makers, but the environment.**\nThis is true, as one could of course use a more 'traditional RL agent', e.g., an LSTM network with a small MLP as the decision head, as an actor for the task and backup. However, the types of environments that the paper considers and allow for backup are closely related to LMs. For example, HotPotQA and WebShop are both recent benchmarks that are considered as NLP tasks; they are largely ignored by the mainstream RL community, and it is only through large language models that the machine learning community has made major progress on solving such tasks. The only class of benchmarks where backup is indeed hard are RL environments like AlfWorld and Minecraft, where the agent's actions permanently change the environment. However, it is important to recognize that this scenario represents only a subset of the environments in which LM agents can be deployed. Thus, it is appropriate to say 'LM agents are not constrained by concerns in traditional RL', as the traditional RL community does not focus on the paper's task of interest.\n", "category": "Method_Comparison"}, {"question": "Why does the ability to back up in environments like API-calling, programming, and web-browsing not invalidate the use of a model-free MCTS algorithm for decision-making, given that backup is a property of the environment and not the agent?", "answer": "In the environments considered (API-calling, programming, and web-browsing), the ability to back up is not a limitation. For WebShop, the environment can be reset by modifying the website's XML, and for programming and API-usage, the LM's actions do not directly affect the environment, making backup a matter of changing the prompt. Thus, backup is a property explicitly leveraged in the work and does not invalidate the use of a model-free MCTS algorithm.", "category": "Method_Disambiguation"}, {"question": "What is the novelty of LATS compared to existing methods like MTCS from RAP[1], ToT[2], and Reflexion[3], given that it appears to be a combination of these methods?\n\n[1] https://arxiv.org/abs/2305.14992\n\n[2] https://arxiv.org/abs/2305.10601\n\n[3] https://arxiv.org/abs/2303.11366", "answer": "The simplicity or combinational nature of an approach should not be misinterpreted as a lack of novelty. Instead, simplicity should be regarded as a commendable strength. The famous perspective from Michael Black on novelty [1] supports this argument:\n\n> “The simplicity of an idea is often confused with a lack of novelty when exactly the opposite is often true. A common review critique is: the idea is very simple. It just changes one term in the loss and everything else is the same as prior work. However, if nobody thought to change that one term, then it is ipso facto novel. The inventive insight is to realize that a small change could have a big effect and to formulate the new loss.”\n\n> “The idea is obvious because the authors just combined two well known ideas. Obvious is the opposite of novelty. So, if an idea is obvious after you’ve heard it, reviewers quickly assume it isn’t novel. The novelty, however, must be evaluated before the idea existed. The inventive novelty was to have the idea in the first place. If it is easy to explain and obvious in hindsight, this in no way diminishes the creativity (and novelty) of the idea.”\n\nThe paper's work precisely matches this notion of novelty: **It is simple, but is the first to explore a timely problem that leverages search algorithms with LLM agents for decision-making tasks, and demonstrates substantial performance improvement including 86.9 pass@1 accuracy with GPT-3.5 on HumanEval, a 27.6% improvement over Reflexion and higher performance than base GPT-4**. The paper's method makes non-trivial efforts to unify successful past designs and adapt to environments with feedback. \n\nWhile LLM scoring was adopted in ToT, it is important to note that none of the prior works employed LLM scoring with self-reflection, making the paper's approach novel, effective, and noteworthy within the community.\n\nUsing ‘prompt-engineering’ should not be a ground for lack of novelty. For instance, ReAct is about the design of prompts based on existing work of chain-of-thought; however, such a simple idea becomes fundamental in the field of LLM decision-making. **The paper's core technical contribution is not based on prompting, but about adapting search algorithms and MCTS to LM agents**.", "category": "Method_Comparison"}, {"question": "Why is the evaluation of LATS biased towards decision-making tasks like HotPotQA and WebShop, and how does LATS compare to ToT[2] and RAP[1] on reasoning benchmarks?\n\n[1] https://arxiv.org/abs/2305.14992\n\n[2] https://arxiv.org/abs/2305.10601", "answer": "The evaluation is biased towards decision-making tasks because LATS is primarily a framework for LM agents, with a core focus on improving performance in decision-making environments while maintaining competitive performance on reasoning tasks. LATS includes results on a reasoning benchmark (HotPotQA without external API) in Table 1, where LATS-CoT outperforms ToT but performs comparably to RAP. This is the standard reasoning setting for HotPotQA as a question-answering benchmark without the external API. In reasoning-only settings, the advantages of LATS are reduced due to the removal of external feedback, which diminishes the efficacy of self-reflection and the LM value function. However, LATS aims to maintain reasoning performance while adapting to decision-making environments with external feedback.", "category": "Method_Comparison"}, {"question": "How does the self-refinement process in ToT [1] compare to the self-reflection in LATS as described in Table 1?\n\n[1] https://arxiv.org/abs/2305.10601", "answer": "The self-refinement in ToT's creative writing experiment is task-specific and based on the coherency of the generated passage, not part of the base ToT design, and cannot be directly applied to decision-making environments. In contrast, LATS's self-reflection is general, describing errors and suggestions applicable to any environment.", "category": "Method_Comparison"}, {"question": "Why are the results of ReAct in Table 5 (WebShop) lower than those reported in the paper?", "answer": "The ReAct results in the paper were obtained using PaLM, a proprietary language model that the paper do not have access to. In the paper’s experiments, the authors used GPT-3.5 because GPT-3.5 is available, widely used, and capable of in-context learning. While GPT-3.5 is similar in model scale to PaLM, there are some performance differences between the two models.", "category": "Experimental_Analysis"}, {"question": "What is the intention behind Figure 5, and does it represent a conceptual description of Tree Search?", "answer": "Figure 5 is intended to show a qualitative example where LATS can improve over ReAct. It is not a conceptual description of Tree Search. ", "category": "Concept_Understanding"}, {"question": "What is the tradeoff between computational cost and performance for LATS compared to other methods like CoT-SC, RAP, and ToT?", "answer": "1. Theoretically, LATS consumes more tokens than CoT-SC but has the same token consumption as RAP and ToT. The token consumption of LATS is higher relative to other simpler baselines (e.g., CoT-SC), as LATS expands with multiple samples every step; the overall token consumption is $n$ times higher than CoT-SC/ReAct with $k$ attempts, where $n$ is the number of samples per expansion. However, as the token cost of other tree-structured search methods (RAP and ToT) are also parameterized by the number of attempts $k$ and the number of nodes expanded $n$, it means that LATS, RAP, and ToT have the same sample complexity, i.e., same asymptotic token cost. The sample complexity is summarized in the table below\n\n2. Empirically, LATS expands fewer nodes upon success than RAP and ToT, which means that it is cheaper. To count the token consumption more accurately, the authors compare the average number of nodes expanded upon success on HotPotQA for LATS, RAP, and ToT. The result shows that LATS does not only have better performance, but also tends to solve questions with fewer nodes (66.65) sampled than RAP (70.60) or ToT (84.05) with the same sample complexity, indicating a more effective, thus cheaper, search. The result is also shown in the table below.\n\n3. Even LATS without expansion of multiple nodes (i.e., $n=1$) outperforms other methods with the same sample complexity (CoT-SC, best among multiple runs for ReAct). To give a fairer comparison between LATS and CoT-SC which does not expand nodes, the authors also test a version of LATS with $n=1, k=50$, and a version of CoT-SC and ReAct with $k=50*5=250$. ReAct and CoT-SC cannot reach the performance of LATS with lower $k$ and $n=1$, let alone LATS with the same sample complexity ($n=5, k=50$).\n\nBelow is the table mentioned above:\n\nMethod | Performance on HotPotQA (&uarr;) | Sample Complexity (&darr;) | Average \\# of Nodes (&darr;)\n---|---|---|---\nReAct (best k = 250) | 0.42 | $O(k)$  |\nCoT-SC (n = 1, k = 250) | 0.40 | $O(k)$ | \nLATS (n = 1, k = 50) | 0.48 | $O(k)$ |\nToT - ReAct (n = 5, k = 50) | 0.49 | $O(kn)$ | 84.05\nRAP - ReAct (n = 5, k = 50) | 0.54 | $O(kn)$ | 70.60\nLATS (n = 5, k = 50) | 0.61 | $O(kn)$ | 66.65", "category": "Method_Comparison"}, {"question": "Given that exact reasoning costs, such as the precise number of API calls and the quantity of tokens used, are provided for each baseline, how does the paper’s method compare to other methods?", "answer": "1) With the same sample complexity, the paper's method performs the best, 2) the paper's method with $n=1, k=50$ performs better than other methods with higher $k$, and 3) while LATS has the same sample complexity as ToT and RAP, the paper's method expands fewer nodes on average upon success, which indicates less token usage.", "category": "Method_Comparison"}, {"question": "What is the novelty of the proposed LATS method compared to existing LLM prompting techniques, particularly in terms of combining search algorithms and self-reflection?", "answer": "The novelty of the proposed LATS method lies in its synergistic integration of search algorithms and self-reflection, which has not been explored in previous work. LATS introduces unique and advanced methods to leverage these components and makes a non-trivial adaptation of tree-based search with external feedback.Based on these, the paper has made significant improvements across multiple testbeds, including 94.4 pass@1 accuracy with GPT-4 on HumanEval and performance comparable to gradient-based fine-tuning methods on WebShop. Additionally, LATS is the first to explore the use of search algorithms with LLM agents for decision-making tasks. Previous approaches like ReAct and Reflexion cannot sample and select more than one state at every step, limiting exploration in decision-making environments; ToT and RAP do not explore decision-making tasks. It is non-trivial to shift from non-interactive tasks to interactive ones, because an updated node, prompt, and more importantly, search algorithm must be designed to incorporate external input – an achievement the paper has successfully accomplished. The paper’s proposed method is verified to accommodate external input, even when applied to original non-interactive tasks such as programming.", "category": "Method_Comparison"}, {"question": "Why was WebShop chosen as the benchmark for evaluating LATS instead of ALFWorld or Minecraft?", "answer": "ALFWorld was not chosen because Reflexion already achieves a success rate of above 90% on ALFWorld, leaving little room for improvement. WebShop was selected because Reflexion only achieves a success rate of 35% on WebShop, which makes it a more complex and challenging benchmark while being widely adopted. Minecraft, while a significant environment for decision-making, is not a standard benchmark for evaluating prompting methods. It requires extensive engineering on the environment itself, which is not relevant to the general decision-making framework. For example, Plan4MC (Haoqi Yuan et al., 2023)[1] uses a hierarchical RL (PPO+DQN) method to find items in the world, which together only serves as a low-level controller for LLM. Without a widely adopted test suite for LLM alone, the performance of an LLM agent in Minecraft largely relies on the quality of the low-level controller, which is out of the paper's scope of proposing a general framework for LLM decision-making (and other tasks such as reasoning). \n\n[1]Haoqi Yuan et al. Plan4MC: Skill Reinforcement Learning and Planning for Open-World Minecraft Tasks. ArXiv: 2303.16563, 2023.", "category": "Motivation_Analysis"}]}
{"id": "BfMQIJ0nLc", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "MMBench: Is Your Multi-modal Model an All-around Player?", "authors": ["Yuan Liu", "Haodong Duan", "Yuanhan Zhang", "Bo Li", "Songyang Zhang", "Yike Yuan", "Wangbo Zhao", "Jiaqi Wang", "Conghui He", "Ziwei Liu", "Kai Chen", "Dahua Lin"], "abstract": "Large vision-language models have recently achieved remarkable progress, \nexhibiting great perception and reasoning abilities concerning visual information. \nHowever, how to effectively evaluate these large vision-language models remains a major obstacle, hindering future development in this domain. \nTraditional benchmarks like VQAv2 or COCO Caption provide quantitative performance measurements but suffer from a lack of fine-grained ability assessment and non-robust evaluation metrics. \nRecent subjective benchmarks, such as OwlEval, offer comprehensive evaluations of a model's abilities by incorporating human labor, but they are not scalable and display significant bias.\nIn response to these challenges, we propose MMBench, a new benchmark for assessing multi-modal capabilities of VLMs. \nMMBench methodically develops a comprehensive evaluation pipeline, primarily comprised of two key features: 1. MMBench is a meticulously curated dataset that surpasses existing similar benchmarks in terms of the number and the variety of evaluation questions and abilities; 2. MMBench introduces a rigorous CircularEval strategy and incorporates the use of ChatGPT to convert free-form predictions into pre-defined choices, thereby facilitating a fair and robust evaluation despite of VLMs' different instruction following capabilities. \nMMBench is a systematically-designed objective benchmark for robustly evaluating the various abilities of vision-language models. \nWe hope MMBench will assist the research community in better evaluating their models and encourage future advancements in this domain.", "keywords": "Vision-language Pre-training;Multimodality;Benchmark;Dataset", "qa_pairs": [{"question": "Why are most of the questions rejected by Bard and GPT-4v, and does this rejection indicate a lack of capability in these models?", "answer": "Most questions rejected by Bard and GPT-4v are due to policies set by their provider organizations, not because the models are incapable of solving them. For example, Bard rejects questions with images containing recognizable human beings, and GPT-4v rejects questions with images of celebrities. ", "category": "Experimental_Analysis"}, {"question": "What is the licensing approach for collected Internet images in MMBench, and how is the issue of watermarks, specifically from Zhihu, addressed?", "answer": "All images in MMBench are licensed for academic use. During collection, annotators document the original URL of each image, and unsuitable images are manually removed. Authors can request image removal, which will be promptly complied with. Regarding the Zhihu watermark, their copyright policy allows use for academic and non-profit purposes but prohibits commercial misuse.", "category": "Experimental_Setup"}, {"question": "Why do open-source visual-language models (VLMs) exhibit poor performance, particularly in reasoning tasks, compared to GPT-4v, and what potential improvements can be made to address this?", "answer": "1. In Table 1, GPT-4v significantly outperforms the leading open-source model (Qwen-VL-Chat) by over 14% overall accuracy on MMBench-Dev. The improvement is primarily evident in reasoning tasks: \n    1. In Attribute Reasoning questions, GPT-4v outperforms the second best opensource VLM IDEFICS-80B by over 12% Top-1 accuracy.\n    2. In Logic Reasoning questions, GPT-4v outperforms all opensource VLMs by over 35% Top-1 accuracy.\n    3. In Relation Reasoning questions, GPT-4v outperforms all opensource VLMs by over 15% Top-1 accuracy.\n    Meanwhile, the gap between opensource VLMs and GPT-4v is not that huge on perception tasks (3% - 10% Top-1 accuracy). \n\n2. Under apple-to-apple comparisons (only questions answered by the closed source VLMs count), the advantage of closed source VLMs is more significant. Generally, the perception performance of open-source VLMs is good and comparable with GPT-4v or Bard (under some perception tasks, open-source VLMs even significantly outperform Bard). However, the reasoning performance of open-source VLMs still lags far behind GPT-4v or Bard. The superior reasoning capability of GPT-4v or Bard is attributed to their potentially larger and more powerful language models.\n\n3. Most existing opensource VLMs simply map the visual concepts to the language embedding space, and finetune the model with VL instruction data (with a large proportion of questions perception related). To further improve the reasoning capabilities, the developers can: 1. Switch to more powerful language backbones; 2. Design better finetuning algorithms; 3. Build reasoning-related instruction datasets with better quality.", "category": "Experimental_Analysis"}, {"question": "How were the 20 ability dimensions in MMBench selected and what was the process used to determine them?", "answer": "The 20 ability dimensions in MMBench were selected by gathering a group of experts from the fields of computer vision, natural language processing, and machine learning to discuss the specific taxonomy of sub-abilities in perception and reasoning. This process resulted in the identification of 20 ability dimensions and an ability hierarchy ranging from L1 to L3. The selection process also included referring to the definitions of perception and reasoning in biology, as discussed in Section 2.1 of the main paper. The dimensions are not fixed and will be continuously updated to improve the ability taxonomy.", "category": "Method_Mechanics"}, {"question": "How is the question 'How many apples are there in the image?' categorized in terms of ability dimensions in MMBench, given that it requires both numerical reasoning and perception?", "answer": "The paper categorizes the ‘counting’ ability under ‘Object Localization’. In MMBench, the paper does not consider ‘counting’ as a more complex reasoning problem, since for most counting questions, there are only a few objects to be counted. If multiple ability dimensions are intertwined, the paper consistently categorizes the question under the more challenging ability dimension. This is because, in most scenarios, the more difficult ability dimension requires the model to possess the capability to solve simpler tasks. By doing so, the model can effectively utilize the task outputs as intermediate results to tackle more complex tasks. For instance, all questions in ‘Attribute Comparison’ rely on the capability of ‘Attribute Recognition’. Since ‘Attribute Comparison’ is considered a more complex task, the paper categorizes these questions as ‘Attribute Comparison’.", "category": "Method_Mechanics"}, {"question": "How does the performance of the models vary across different ability dimensions, as detailed in Tables 9 through 14?", "answer": "Tables 9 through 14 provide additional details on the results, showing that Shikra demonstrates exceptional performance in Action Recognition, while LLaMA-Adapter excels in Attribute comparison. These results indicate that factors such as data composition, instruction data quality, and instruction template significantly influence a model's performance across different ability dimensions.", "category": "Experimental_Exposition"}, {"question": "Does the MMBench benchmark suffer from the same shortcut and bias issues as traditional VQA datasets, and what measures were taken to assess the necessity of image data for answering questions?", "answer": "One shortcut might be that the model could provide accurate answers even without relying on image information. To assess the necessity of image data for questions in the paper's dataset, the authors employed some state-of-the-art LLMs to respond to queries in the MMBench-dev. During the evaluation, the authors design meta prompts to guide the LLMs try their best to make a reasonable guess, even if a question is not theoretically answerable given only the text part. As demonstrated in the table, the performance of the text-only GPT-4 model — prompted solely with question text — is notably suboptimal. This limitation is particularly evident in responding to queries that require 'perception' abilities. Such questions, often intrinsically linked to visual information, include examples like 'What mood does this image convey?' or 'Who is this person?' In contrast, GPT-4 exhibits a degree of proficiency in 'reasoning' tasks, with a notable strength in logical reasoning (LR). The model's performance in LR is comparable to that of other leading open-source models. The effectiveness in LR could be attributed to the model's ability to leverage common sense in answering questions. Notably, some LR data, sourced from ScienceQA, suggest that a portion of these tasks can be effectively addressed using the model's inherent common sense reasoning.\n\n| Model         | Overall | CP    | FP-S  | FP-C  | AR    | LR    | RR    |\n|---------------|---------|-------|-------|-------|-------|-------|-------|\n| GPT-4v        | 75.06   | 82.77 | 71.65 | 66.43 | 77.88 | 69.29 | 74.31 |\n| Qwen-VL-Chat  | 60.57   | 79.39 | 66.21 | 48.25 | 59.8  | 32.2  | 43.48 |\n| Shikra        | 59.36   | 76.35 | 57.68 | 58.74 | 57.29 | 26.27 | 58.26 |\n| Bard          | 58.16   | 58.42 | 57.6  | 39.6  | 68.27 | 53.4  | 71.64 |\n| IDEFICS-80B   | 54.81   | 64.53 | 58.36 | 44.76 | 65.83 | 23.73 | 46.09 |\n| GPT-4         | 13.23   | 6.41  | 7.84  | 6.99  | 32.16 | 27.96 | 4.34  |\n| GPT-3.5       | 11.17   | 2.7   | 9.9   | 4.9   | 22.11 | 25.42 | 10.43 |\n| Claude-2      | 10.74   | 5.74  | 6.14  | 2.8   | 27.14 | 22.88 | 4.35  |\n", "category": "Experimental_Exposition"}, {"question": "What additional bias analysis was performed for models other than GPT-4, and what were the key findings?", "answer": "Additional bias analysis was performed for GPT-4v, Qwen-VL-Chat, Shikra, and Bard across dimensions such as CP, FP-S, FP-C, AR, LR, and RR. The results highlight differences in performance and provide insights into their strengths and weaknesses. Additionally, questions correctly answered by GPT-4 were removed, and accuracies were recalculated. The findings show that this removal had limited impact on final accuracy and did not alter the conclusions.", "category": "Experimental_Exposition"}, {"question": "What is the process for collecting samples in MMBench, and what steps are taken to ensure the quality of the benchmark?", "answer": "Samples in MMBench are collected by trained college student volunteers who gather images and create relevant questions and answers. The volunteers are provided with potential sources for images and question formulation, as shown in Table 7. The collected VQA data undergoes a quality inspection pipeline, including manual review to ensure (1) the accuracy and coherence of questions without logical or factual errors and (2) the presence of a single correct answer pertinent to each question. Samples failing these criteria are filtered out.", "category": "Experimental_Setup"}, {"question": "What steps were taken to address the issue of questions in MMBench that can be answered without the corresponding image, and how did this affect the evaluation of models?", "answer": "The authors input all MMBench questions into GPT-4 to determine which could be answered without images, finding that ~13% of dev questions and ~10% of test questions (mainly from ScienceQA) could be answered correctly. These questions were removed, and the remaining questions were used to evaluate models. The performance and rankings of the models remained similar to those before the image filter,  and the rankings of the different models were basically consistent, indicating the credibility of the MMBench leaderboard. ", "category": "Experimental_Exposition"}, {"question": "What is CircularEval, and how does it reduce the inference cost associated with querying ChatGPT?", "answer": "CircularEval is a method that inputs a single question into a vision-language model N times, with the choices shifted each time. It assigns a success score only if the model correctly answers the question all N times, and a fail score if the model provides an incorrect answer even once. This eliminates the need to query ChatGPT for the remaining samples, significantly reducing inference cost. For example, while VanillaEval prompts ChatGPT 383 times on average, CircularEval prompts it 681 times, representing a 78% increase, which is significantly less than a 300% increase. In some cases, CircularEval requires fewer ChatGPT calls than VanillaEval, especially when the heuristic strategy fails to match the VanillaEval prediction but succeeds in matching a CircularEval prediction with a wrong option.", "category": "Method_Disambiguation"}, {"question": "What evidence supports the classification of MMBench as a VQA benchmark rather than a QA benchmark?", "answer": "The paper's analysis shows that only 10% of the questions can be answered without image aid, while the majority require accurate image comprehension for correct responses.", "category": "Experimental_Exposition"}, {"question": "What is the methodology for answer extraction using LLMs in the paper's evaluation process, and how does it differ from conventional NLP practices?", "answer": "The approach involves first using heuristic matching to align option labels(A, B, C, D) or content with the model's predictions. If this fails, ChatGPT is employed to semantically match the prediction to one of the available choices, outputting the corresponding label. This method is more complex than straightforward answer matching and is the first to integrate the strong semantic matching capability of cutting-edge LLMs into the evaluation process of multimodal understanding. It makes the evaluation procedure yield more accurate results and reduces the number of false negatives.", "category": "Method_Mechanics"}, {"question": "What are the reasons for claiming that the MME evaluation setting is non-rigorous?", "answer": "MME continues to utilize the traditional method of exact matching, which necessitates the model to produce precise 'yes' or 'no' responses. This approach, however, tends to generate a number of false negative samples. The metric adopted by MME (accuracy & accuracy +) enables random guessing to achieve a substantial amount of score (37.5%) and also makes it difficult to reveal the performance gap between VLMs. Besides, it appears unreasonable that certain models yield scores on the MME that are lower than those of random guesses (especially for perception tasks). For instance, VisualGLM-6B and MiniGPT-4 demonstrate poorer performance compared to the random baseline in MME, which doesn't sound reasonable. However, in MMBench-Dev, the two approaches significantly outperform the random baseline (which is <0.5% Top-1 accuracy).", "category": "Experimental_Analysis"}, {"question": "How does the efficiency of the benchmark using CircularEval and ChatGPT compare to other benchmarks in terms of speed and cost?", "answer": "Incorporating ChatGPT into the evaluation pipeline may slow down the process compared to benchmarks like VQAv2 and GQA that use exact matching. However, when weighing the tradeoff between robustness and efficiency, using ChatGPT for extraction and matching emerges as the optimal choice. This is because it eliminates the issue of generating a high number of false negative samples, and the time required for evaluation remains within acceptable limits. The following table presents the number of ChatGPT API calls required by each VLM under VanillaEval and CircularEval.\n\n| Model            | VanillaEval GPT Calls | CircularEval GPT Calls |\n| ---------------- | --------------------- | ---------------------- |\n| OpenFlamingov2   | 13                    | 7                      |\n| LLaVA-7B         | 993                   | 1908                   |\n| VisualGLM-6B     | 1051                  | 1990                   |\n| MMGPT-9B         | 143                   | 229                    |\n| InstructBLIP-13B | 69                    | 102                    |\n| mPLUG-Owl-7B     | 701                   | 1510                   |\n| Qwen-VL          | 245                   | 352                    |\n| MiniGPT-4-7B     | 656                   | 1026                   |\n| Idefics-80B      | 23                    | 36                     |\n| PandaGPT         | 1167                  | 1993                   |\n| MiniGPT-4-13B    | 524                   | 1120                   |\n| OpenFlamingo     | 68                    | 57                     |\n| Llama-adapter    | 960                   | 1624                   |\n| Otter-9B         | 0                     | 0                      |\n\nAmong all VLMs, PandaGPT consumes most ChatGPT API calls (~2000). Each answer matching prompt is 500 tokens length on average. Based on the current price of ChatGPT (GPT-3.5, which is 0.001 USD / 1k tokens), the price upper bound for each evaluation is only 1 USD (~0.3 USD for all models on average). In practice,  for most models, the evaluation can finish in half an hour, with one single OpenAI key and parallel API calling.", "category": "Method_Comparison"}, {"question": "How is ChatGPT used in the MMBench evaluation pipeline?", "answer": "ChatGPT is used in the MMBench evaluation pipeline to extract answers from predictions and match them to available choices, with accuracy calculated by predefined functions using the CircularEval strategy. ", "category": "Experimental_Setup"}, {"question": "What novel insights regarding the shortcomings of existing MLLMs or suggestions for improvements are highlighted in the MMBench leaderboard evaluation?", "answer": "1. The performance of coarse perception is significantly lower than that of fine-grained perception across all models presented in Table 3 of the main paper. This trend is understandable, given that fine-grained perception necessitates the model's ability to precisely identify and comprehend the various objects present in an image. Furthermore, it  is discovered that enabling trainability of parameters in the vision encoder positively influences the model's fine-grained perception capabilities. This is because it equips the model with the flexibility to modify its vision feature extraction ability during visual-language pre-training and instruction tuning.\n\n2. The performance of cross-instance fine-grained perception is much lower than that of single-instance fine-grained perception. The leaderboard of MMBench indicates that a majority of pre-training vision-language datasets primarily include descriptions or questions pertaining to individual objects, thereby neglecting to address cross-object relationships. Furthermore,  incorporating a referring expression comprehension task can effectively enhance the performance of the cross-instance fine-grained perception task.\n\n3. The performance on perception tasks significantly surpasses that on reasoning tasks. Reasoning is a more complex task compared to perception, as it demands the model to not only accurately identify various objects in an image, but also possess the ability to reason about these objects, including their function and social identity. Therefore, enhancing the complexity of the instruction data, rather than posing simple questions about the color and shape of objects, is anticipated to improve this ability.", "category": "Experimental_Analysis"}, {"question": "Have potential vulnerabilities or successful attacks on ChatGPT been identified that could affect its reliability within the MMBench evaluation pipeline?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did GPT-4v refuse to answer a portion of the questions in MMBench-Dev, affecting the accuracy calculation?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the MMBench dataset suffer from the same shortcut and bias issues as traditional VQA datasets, where models can answer questions without real understanding?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "DS5qRs0tQz", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection", "authors": ["Shilong Liu", "Zhaoyang Zeng", "Tianhe Ren", "Feng Li", "Hao Zhang", "Jie Yang", "Chunyuan Li", "Jianwei Yang", "Hang Su", "Jun Zhu", "Lei Zhang"], "abstract": "In this paper, we develop an open-set object detector, called Grounding DINO, by marrying Transformer-based detector DINO with grounded pre-training, which can detect arbitrary objects with human inputs such as category names or referring expressions. \nThe key solution of open-set object detection is introducing language to a closed-set detector for open-set concept generalization. \nTo effectively fuse language and vision modalities, we conceptually divide a closed-set detector into three phases and propose a tight fusion solution, which includes a feature enhancer, a language-guided query selection, and a cross-modality decoder for cross-modality fusion. \nWhile previous works mainly evaluate open-set object detection on novel categories, we propose to also perform evaluations on referring expression comprehension for objects specified with attributes. \nGrounding DINO performs remarkably well on all three settings, including benchmarks on COCO, LVIS, ODinW, and RefCOCO/+/g. \nGrounding DINO achieves a $52.5$ AP on the COCO detection zero-shot transfer benchmark, i.e.,  without any training data from COCO. It sets a new record on the ODinW zero-shot benchmark with a mean $26.1$ AP.", "keywords": "Object Detection;Visual Grounding;Open Vocabulary", "qa_pairs": [{"question": "How does increasing the number of queries beyond 900 affect the model's generalization performance, particularly for rare classes, and what are the experimental results and implications of this increase?", "answer": "The authors expanded the experiments to include 1200 and 1500 queries on both the COCO and LVIS datasets. The results of these experiments are presented in Table R.1. Two model variants, O365 and O365+GoldG, were trained, each initialized from respective checkpoints. To maintain consistency, the learnable queries (tgt_embed) were kept unchanged. The training protocols were slightly varied: the O365 model underwent three epochs with a learning rate reduction after the second epoch, while the O365+GoldG model had two epochs with a similar rate adjustment after the first epoch. All tests were conducted using 8xA100 GPUs. Time and resource constraints may have limited the ability to fully optimize the models, particularly for the LVIS dataset. Extended training periods post-learning rate adjustment could potentially enhance performance further. Analysis shows that models with 1200 and 1500 queries show a modest improvement in performance on LVIS’s rare classes compared to the 900-query model. This suggests an improved coverage of rare classes. However, this improvement is marginal and the overall performances of models with varying query numbers are quite similar. This is likely because the 900-query model already provides ample coverage of LVIS objects. Furthermore, increasing the number of queries introduces a training challenge due to exacerbated data imbalances, as models are trained on objects from sampled categories. This imbalance could negate the benefits of additional queries.\n\n| Model       | Query Num | COCO AP | AP-small | AP-medium | AP-large | LVIS AP | AP-r  | AP-c  | AP-f  |\n| :---------- | :-------- | :------ | :------- | :-------- | :------- | :------ | :---- | :---- | :---- |\n| O365+GoldG  | 900       | 48.2    | 34.2     | 51.2      | 62.1     | 21.8    | 10.4  | 16.2  | 28.7  |\n| O365+GoldG  | 1200      | 48.0    | 34.6     | 51.2      | 62.2     | 21.5    | 10.9  | 15.8  | 28.4  |\n| O365+GoldG  | 1500      | 48.1    | 34.7     | 51.2      | 62.3     | 21.8    | 11.0  | 16.2  | 28.7  |\n| O365        | 900       | 46.4    | 33.1     | 49.8      | 60.2     | 14.4    | 6.6   | 8.0   | 21.4  |\n| O365        | 1200      | 46.5    | 32.6     | 49.5      | 60.4     | 14.6    | 6.4   | 8.4   | 21.7  |\n| O365        | 1500      | 46.3    | 32.7     | 49.3      | 60.3     | 14.8    | 6.3   | 8.6   | 21.8  |\n\nTable R.1 Results of the extended experiments with increased query numbers.\n", "category": "Experimental_Exposition"}, {"question": "What is the contribution of individual modeling choices in Tables 6 and 7, and how does tight fusion compare to other techniques?", "answer": "Regarding Table 7, the paper's primary aim was to demonstrate an efficient approach for Grounding DINO training. As detailed in Section 4.5, the paper illustrates that freezing a pre-trained DINO and tuning additional parameters introduced in Grounding DINO can achieve faster model convergence and comparable performance to full-parameter pretrained checkpoints. Regarding Table 6, the paper's data shows that encoder fusion (tight fusion) significantly improves model performance on both COCO and LVIS datasets. The results from comparing model #1 with the baseline model #0 validate this observation. Other techniques, such as language-guided query selection, text cross-attention, and sub-sentence text prompt, also contribute positively to the LVIS performance, yielding significant gains of +3.0 AP, +1.8 AP, and +0.5 AP, respectively. Additionally, these methods enhance the COCO zero-shot performance, further underscoring their effectiveness. However,it is observed that language-guided query selection and sub-sentence text prompt had minimal impact on the COCO fine-tune performance. This outcome is reasonable, given that these methods do not alter model parameters or add computational burdens. Text cross-attention, while introducing fewer parameters than encoder fusion, showed less performance improvement compared to encoder fusion (+0.6 vs. +0.8). This finding suggests that fine-tuning performance is predominantly influenced by the model's parameters, indicating that scaling models is a promising direction for enhancing performance.", "category": "Experimental_Exposition"}, {"question": "Does increasing the model size of the BERT text encoder (e.g., from BERT-B to BERT-L) improve the representation of complex text input and alleviate the limitations of the REC dataset?", "answer": "Experiments were conducted using BERT-B (bert-base-uncased) and BERT-L (bert-large-uncased) on a combined dataset of RefCOCO, RefCOCO+, and RefCOCOg, with leaked data removed for fairness.  It's important to note that training on this combined dataset, as opposed to tuning each dataset separately, might result in slightly lower performance. For a fair comparison, the authors initialized the models with the O365+GoldG+Cap4M checkpoint, except for the BERT parameters. Limited by time and resources, the training duration was capped at 18 epochs. The results showed that Grounding DINO with BERT-B outperformed or matched BERT-L in most metrics (Table 1). This suggests that BERT-B, used during the pre-train stage, may contribute to better performance. No significant improvements were observed in the late stages of training, indicating that both models were nearing their optimal performance. The main limitation in enhancing REC performance appears to lie in the detection branch rather than the language processing module. A dedicated model for REC data might be more effective.\n\nTable 1\n\n| Model                  | refococo (val) | refococo (testA) | refococo (testB) | refococo+ (val) | refococo+ (testA) | refococo+ (testB) | refococog (val) | refococog (test) |\n|------------------------|----------------|------------------|------------------|-----------------|-------------------|-------------------|-----------------|------------------|\n| Grounding DINO T (BERT-B) | 87.4           | 91.6             | 84.2             | 78.6            | 86.5              | 73.4              | 81.6            | 83.3             |\n| Grounding DINO T (BERT-L) | 87.0           | 91.4             | 84.0             | 78.1            | 86.4              | 73.0              | 81.7            | 83.3             |\n", "category": "Experimental_Exposition"}, {"question": "How does the performance of Grounding DINO compare to GLIPv2 on ODinW, particularly in terms of $AP_{average}$ and $AP_{median}$, and what implications does this have for its effectiveness in true open-set scenarios?", "answer": "Grounding DINO and GLIPv2-T show similar $AP_{average}$ on ODinW, but Grounding DINO significantly outperforms GLIPv2-T in $AP_{median}$ (11.9 vs 8.9). This suggests that while GLIPv2 may exhibit larger performance variance across different datasets, Grounding DINO maintains a more consistent performance level. This consistency is particularly relevant in true open-set scenarios, where variance can greatly impact overall effectiveness. Additionally, it's important to note the complexity and size differences between the models. GLIPv2 incorporates advanced techniques like masked text training and cross-instance contrastive learning, making it more complex than the paper's Grounding DINO model. Moreover, the paper's model is more compact (172M parameters) compared to GLIPv2 (232M parameters).", "category": "Method_Comparison"}, {"question": "What were the results of the experiments that verified the poor performance of the query selection module on LVIS rare categories was due to an insufficient top-K value?", "answer": "The authors expanded the experiments to include 1200 and 1500 queries on both the COCO and LVIS datasets. The results of these experiments are presented in Table R.1. Two model variants, O365 and O365+GoldG, were trained, each initialized from respective checkpoints. To maintain consistency, the learnable queries (tgt_embed) were kept unchanged. The training protocols were slightly varied: the O365 model underwent three epochs with a learning rate reduction after the second epoch, while the O365+GoldG model had two epochs with a similar rate adjustment after the first epoch. All tests were conducted using 8xA100 GPUs. Time and resource constraints may have limited the ability to fully optimize the models, particularly for the LVIS dataset. Extended training periods post-learning rate adjustment could potentially enhance performance further. Analysis shows that models with 1200 and 1500 queries show a modest improvement in performance on LVIS’s rare classes compared to the 900-query model. This suggests an improved coverage of rare classes. However, this improvement is marginal and the overall performances of models with varying query numbers are quite similar. This is likely because the 900-query model already provides ample coverage of LVIS objects. Furthermore, increasing the number of queries introduces a training challenge due to exacerbated data imbalances, as models are trained on objects from sampled categories. This imbalance could negate the benefits of additional queries.\n\n| Model       | Query Num | COCO AP | AP-small | AP-medium | AP-large | LVIS AP | AP-r  | AP-c  | AP-f  |\n| :---------- | :-------- | :------ | :------- | :-------- | :------- | :------ | :---- | :---- | :---- |\n| O365+GoldG  | 900       | 48.2    | 34.2     | 51.2      | 62.1     | 21.8    | 10.4  | 16.2  | 28.7  |\n| O365+GoldG  | 1200      | 48.0    | 34.6     | 51.2      | 62.2     | 21.5    | 10.9  | 15.8  | 28.4  |\n| O365+GoldG  | 1500      | 48.1    | 34.7     | 51.2      | 62.3     | 21.8    | 11.0  | 16.2  | 28.7  |\n| O365        | 900       | 46.4    | 33.1     | 49.8      | 60.2     | 14.4    | 6.6   | 8.0   | 21.4  |\n| O365        | 1200      | 46.5    | 32.6     | 49.5      | 60.4     | 14.6    | 6.4   | 8.4   | 21.7  |\n| O365        | 1500      | 46.3    | 32.7     | 49.3      | 60.3     | 14.8    | 6.3   | 8.6   | 21.8  |\n\nTable R.1 Results of the extended experiments with increased query numbers.", "category": "Experimental_Exposition"}, {"question": "What are the key differences and similarities between the feature enhancer and cross-modality decoder in the proposed method compared to those in GLIP[1] and X-Decoder[2]?\n\n[1] Grounded Language-Image Pre-training, CVPR 2022 \n\n[2] Generalized Decoding for Pixel, Image and Language, CVPR 2023", "answer": "In summary, the paper’s feature enhancer is similar to GLIP but more aligned with the paper’s pure Transformer architectures. X-Decoder has no similar modules like the paper's feature enhancer. The paper’s cross-modality decoder, especially the text cross-attention, is unique to the designs in GLIP and X-Decoder. Below are detailed comparisons:\n - **Comparison with GLIP:** \n   - **Feature Enhancer:** Tthe paper's feature enhancer, though similar to GLIP's, is more aligned with the paper's pure Transformer architecture. While GLIP uses DyHead for enhancing visual features, the paper's method employs a Deformable Transformer for image encoder features, ensuring a more consistent architecture across different modalities.  \n    - **Cross-Modality Decoder:** Unlike GLIP, which uses the same head as ATSS/RetinaNet, the paper's DETR-like model uniquely incorporates a cross-modality decoder. This decoder leverages the Transformer's capability to attend to features from both text and image modalities, a distinction validated through the paper's ablation studies.\n\n- **Comparison with X-Decoder:** \n  - **Feature Fusion:** X-Decoder utilizes a standard image encoder and object decoder for different tasks, without integrating visual and text features during the encoding phase. In contrast, the paper's model employs a feature enhancer for fusing features from both modalities. \n  - **Queries in Decoder:** The paper's model's queries in the decoder aggregate features from both image and text, unlike X-Decoder's approach where the interactions are limited to queries and image features only. These comparisons highlight the unique aspects of the paper's method in the context of existing models. ", "category": "Method_Comparison"}, {"question": "How does the performance of Grounding DINO compare to DetCLIP-v2[1] on the ODinW and LVIS benchmarks, considering factors such as zero-shot performance, fine-tuning performance, and dataset size?\n\n[1] DetCLIPv2: Scalable Open-Vocabulary Object Detection Pre-training via Word-Region Alignment, CVPR 2023", "answer": "**ODinW Benchmark Performance**\nIt's important to note that **DetCLIP-v2 did not report zero-shot performance on the ODinW benchmark** in their paper, and their results could not be found on the ODinW leaderboard. Without access to their models, code, and training data, it is not feasible to independently evaluate their zero-shot performance. However, in terms of fine-tuning performance, **the paper's Grounding DINO Tiny model outperforms DetCLIP-v2 Tiny (Grounding DINO Tiny: 70.0 vs. DetCLIP-v2 Tiny: 68.0) with much less data (Grounding DINO Tiny 5.2M vs. DetCLIP-v2 Tiny: 16.2M data)**. For this comparison, the authors used the ODinW-13 metric, a subset of the standard ODinW-35 metric, to align with DetCLIP-v2's benchmarking approach. **LVIS Benchmark Performance** While DetCLIP-v2 demonstrates better zero-shot performance on the LVIS benchmark, they utilize a substantially larger dataset (~16.2M data) compared to Grounding DINO (~5.2M data). This difference in data volume is a significant factor in performance. To illustrate the potential of the paper's model under similar conditions, the authors conducted oracle experiments. The authors pseudo-labeled the ImageNet dataset using a pre-trained Detic model and filtered out LVIS-related images for training, creating a new pseudo-labeled dataset named IN22K-LVIS-1M. It contains about 1M pseudo-labeled images for training. Note that the paper's model under the setting may not be a real zero-shot setting, as the paper's pseudo labeler Detic is trained with LVIS data. The results from these experiments suggest that Grounding DINO can achieve promising LVIS performance, even without direct LVIS training data. A similarly distributed dataset can help the model to generalize well. \n\n| Model            | Pre-train Data      | LVIS Zeroshot AP(r/c/f)      | LVIS Finetune AP(r/c/f)        | \n|------------------|---------------------|-----------------------------|-------------------------------| \n| DetCLIPv2-T      | OG + CC15M          | 40.4 (36.0 / 41.7 / 40.0)   | 50.7 (44.3 / 52.4 / 50.3)     | \n| Grounding DINO T | OG+IN22K-LVIS-1M    | 40.6 (38.5 / 41.1 / 40.4)   | 54.5 (47.3 / 53.9 / 56.1)     | Table R.3. Oracle experiments on LVIS.", "category": "Method_Comparison"}]}
{"id": "UU9Icwbhin", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Retentive Network: A Successor to Transformer for Large Language Models", "authors": ["Yutao Sun", "Li Dong", "Shaohan Huang", "Shuming Ma", "Yuqing Xia", "Jilong Xue", "Jianyong Wang", "Furu Wei"], "abstract": "In this work, we propose Retentive Network (RetNet) as a foundation architecture for large language models, simultaneously achieving training parallelism, low-cost inference, and good performance. We theoretically derive the connection between recurrence and attention. Then we propose the retention mechanism for sequence modeling, which supports three computation paradigms, i.e., parallel, recurrent, and chunkwise recurrent. Specifically, the parallel representation allows for training parallelism. The recurrent representation enables low-cost $O(1)$ inference, which improves decoding throughput, latency, and GPU memory without sacrificing performance. The chunkwise recurrent representation facilitates efficient long-sequence modeling with linear complexity, where each chunk is encoded parallelly while recurrently summarizing the chunks. Experimental results on language modeling show that RetNet achieves favorable scaling results, parallel training, low-cost deployment, and efficient inference. The intriguing properties make RetNet a strong successor to Transformer for large language models.", "keywords": "Large Language Models;Inference Efficiency", "qa_pairs": [{"question": "Is 100B training tokens sufficient for evaluating the performance of a 7B model, given that Transformers are data-eager?", "answer": "Yes, 100B training tokens are sufficient for evaluating the performance of a 7B model. According to the Chinchilla scaling law, the convergence speed of a 7B model slows down at $FLOPs=10^{21}$. The flops for 100B tokens are more than $(7\\times 10^ 9) \\times (100 \\times 10^ 9) \\times 6 > 4\\times 10^{21}$, which is more than enough for model evaluation.", "category": "Method_Disambiguation"}, {"question": "What are the specific shortcomings of the RetNet approach, and on which tasks does it not perform well?", "answer": "RetNet's primary shortcoming is its lack of strict comparability under small sizes, such as 1.3B parameters. Additionally, it is designed for GPT settings and cannot be migrated to bidirectional transformers, impacting its performance on regular NLP tasks that follow pre-training performance.", "category": "Method_Disambiguation"}, {"question": "Why does Table 1 claim that the inference cost of RetNet is $O(1)$ when the recurrent state $s_n$ in Eq (1) appears to have a shape of $d\\times d$, which could be very large in practice, while in S4 or other SSMs, the recurrent hidden state has a shape of $h\\times d$ with a relatively small $h$?", "answer": "Eq. (1) is discussed under one-head settings. The claim in Table 1 is based on the multi-head settings used in Eq. (8), where the shape of $s_n$ is $h\\times d$, with $h$ being the query and key's head dimension. This ensures the inference cost remains $O(1)$, as demonstrated in the inference cost experiment in Figure 4.", "category": "Concept_Understanding"}, {"question": "Why is there no direct comparison of RetNet with other models in a large-scale setting?", "answer": "The conclusions from comparing RetNet with other architectures remain consistent even with larger-scale training. However, scaling up every method exceeded the paper's computation budget. Given that the Transformer is the current standard architecture, the authors allocated more resources to comparisons with Transformers, which were conducted in large-scale settings.", "category": "Motivation_Analysis"}, {"question": "What is the mathematical basis for the equivalency between the forms shown in Figure 2 (a) and Figure 2 (b)?", "answer": "The equivalency between the parallel and recurrent forms is discussed mathematically from Eq. (1) to Eq. (6), with Figure 2a corresponding to Eq. (5) and Figure 2b corresponding to Eq. (6).", "category": "Concept_Understanding"}, {"question": "What empirical and theoretical evidence supports the claim that retention is more capable than full attention in terms of utilizing numerical ranges and performance under various tasks?", "answer": "The paper evaluates the comparable capabilities of retention and full attention empirically. Retention allows negative values for 'retention scores,' whereas full attention only permits positive values for attention scores. Considering finite arithmetic precision, retention potentially utilizes the capacity more efficiently by not wasting the negative numerical ranges. Additionally, the claim that 'Full attention should be more capable than recurrent networks' is not trivial, as attention often underperforms compared to RNN or S4 in tasks like learning formal languages.", "category": "Experimental_Analysis"}, {"question": "Why does RetNet show greater effectiveness in the large model regime in Figure 3, and how does this align with prior work[1] suggesting that two model architectures should not cross over when scaling up in log scale?\n\n[1] https://arxiv.org/abs/1712.00409", "answer": "The comparison in Figure 3 involves two different architectures, and the curves do not cross over for the same architecture. However, when comparing different methods, it is possible for the curves to cross over due to differences in the learning curve parameters ($\\alpha, \\beta, \\gamma$). Specifically, if $f_1(m) = f_2(m)$ and $\\beta_{g1} > \\beta_{g2}$, the curves can cross over. This behavior is consistent with the mathematical formulation of the learning curves and does not contradict prior work.", "category": "Experimental_Analysis"}, {"question": "What are experimental results on generative tasks such as translation and question answering to demonstrate the generalization capability of the model architecture?", "answer": "The language modeling perplexity correlates well with different tasks. Moreover, the language models have unified all tasks under generation. The paper reports language modeling perplexity and various downstream task performances. The paper does not evaluate translation tasks because the model’s training corpus contains English data instead of multilingual data. The authors additionally evaluate one-shot performance on two open-ended question answering tasks with 7B models as follows. Notice that the recall metric is reported in the table, i.e., whether the answers are contained in the generated response.\n\n| Dataset | Transformer | RetNet |\n|---------|-------------|--------|\n| Squad   | 67.7        | 72.7   |\n| WebQS   | 36.4        | 40.4   |\n", "category": "Experimental_Exposition"}, {"question": "What are the empirical and theoretical reasons that RetNet could outperform full attention-based transformers in terms of quality?", "answer": "RetNet allows negative values for retention scores, which enables it to utilize the numerical capacity more efficiently compared to full attention, which only allows positive values for attention scores. This is particularly relevant when considering finite arithmetic precision in practice. Additionally, the claim that full attention is inherently more capable than recurrent networks is not trivial, as attention mechanisms often underperform compared to RNNs or S4 in tasks such as learning formal languages.", "category": "Experimental_Analysis"}, {"question": "Why do the model scaling curves in Figure 3 cross over, and how does this align with previous empirical findings[1]?\n\n[1]https://arxiv.org/abs/1712.00409", "answer": "The curves in Figure 3 cross over because they compare two different architectures. For the same architecture, the curves do not cross. When comparing different methods, if the learning curve is given by $f(m) = \\alpha m^{\\beta_g} + \\gamma$, there are three parameters: $\\alpha$, $\\beta$, and $\\gamma$. If $f_1(m) = f_2(m)$ at some point and $\\beta_{g1} > \\beta_{g2}$, it is possible for the two lines to cross over. This phenomenon does not contradict earlier papers.\n", "category": "Experimental_Analysis"}, {"question": "How does the paper justify the claim in Figure 1 regarding the performance of RWKV and H3 models compared to Transformers, given that these models have demonstrated comparable performance at the billion-scale level with parallel training and constant inference?", "answer": "The authors justify the claim by explaining that 'comparable performance' refers to similar results under the same settings, such as number of parameters and training corpus. They note that previous comparisons used Transformers with absolute position, while the compared methods benefited from relative position modeling. Additionally, the H3 paper's comparable results were in hybrid settings, combining H3 and Transformer layers, whereas the authors did not add any Transformer layers. They conducted controlled experiments with matched parameters and the same training corpus to compare different architectures and are confident in their claim, supported by the experiments in Table 4 which show a significant gap in performance.", "category": "Experimental_Analysis"}, {"question": "How does the description of RWKV in Table 1 accurately reflect its training parallelization and performance comparison with Transformers? According to RWKV and the description in [1], RWKV can indeed be computed in parallel. On the other hand, according to the RWKV paper, its performance is comparable to Transformers, so describing its performance as ✔ is  inaccurate.\n\n[1]Tri Dao, Daniel Y. Fu, Khaled Kamal Saab, Armin W. Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. CoRR, abs/2212.14052, 2022.", "answer": "The description in Table 1 defines 'training parallelization' from a sequential perspective, indicating whether the training implementation is sequentially parallelized. Although RWKV uses channel-wise parallelism, the paper's experiments show that its performance is not comparable to Transformers under the same conditions (same #parameters, same data, and with relative position modelings), making the statement fair.", "category": "Method_Disambiguation"}, {"question": "How does the implementation of GroupNorm in Equation 8 differ from NormAttention proposed in [1]\n\n[1]Zhen Qin, Xiaodong Han, Weixuan Sun, Dongxu Li, Lingpeng Kong, Nick Barnes, and Yiran Zhong. The devil in linear transformer. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7025–7041, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguistics.", "answer": "GroupNorm is significantly better than LayerNorm for multi-scale retention, as different heads have different statistic information. The implementation is different from NormAttention here.", "category": "Method_Comparison"}, {"question": "What is the distinction between ‘Linear Attention’ and ‘Linearized Attention’, and how does ‘approximating softmax’ relate to previous methods in [1] and [2]? Given that Linear Attention refers to the use of the Right-product trick to reduce complexity, it does not necessarily imply approximation. For example, methods [1], [2], and [3] do not involve an approximation approach like softmax.\n\n[1] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. In ICLR, 2022.\n\n[2] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 5156–5165. PMLR, 2020.\n\n[3] Zhen Qin, Xiaodong Han, Weixuan Sun, Dongxu Li, Lingpeng Kong, Nick Barnes, and Yiran Zhong. The devil in linear transformer. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7025–7041, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguistics.\n", "answer": "'Linear Attention' refers to the method described in the paper 'Transformers are RNNs: Fast autoregressive transformers with linear attention', and is distinct from 'Linearized Attention'. 'Approximating softmax' in previous methods means encoding a 'probability distribution' on different positions, which is applicable in [1] and [2].", "category": "Concept_Understanding"}, {"question": "Why does Table 2 not include a comparison with open-source models?", "answer": "The paper uses Transformers with RoPE relative positions as a strong baseline, which differs from LLaMA only in RMSNorm and SwiGLU. RMSNorm improves stability rather than performance, and both RMSNorm and SwiGLU are orthogonal to the work. The focus is on fair comparisons, ensuring rigorous and fair settings, rather than chasing benchmarks. Comparisons with different checkpoints are avoided due to variations in training data, preprocessing pipelines, architecture configurations, number of training tokens, and hyperparameters.", "category": "Motivation_Analysis"}, {"question": "How does the evaluation scope address the concern of being too limited, particularly regarding the types of tasks and languages included?", "answer": "The language modeling perplexity correlates well with performance on different tasks. Moreover, the language models have unified all tasks as generation. The paper reports language modeling perplexity and performance on various downstream tasks. The paper doesn’t evaluate translation because its training corpus contains English data instead of multilingual data. The authors additionally evaluate one-shot performance on two open-ended question answering tasks using 7B models as follows. Notice that the recall metric is reported in the table, i.e., whether the answers are contained in the generated response.\n\n| Dataset | Transformer | RetNet |\n|---------|-------------|--------|\n| Squad   | 67.7        | 72.7   |\n| WebQS   | 36.4        | 40.4   |\n", "category": "Experimental_Exposition"}, {"question": "Why are the evaluation datasets in Tables 4 and 5 inconsistent with Table 3?", "answer": "The evaluation is divided into two groups: large-scale comparison with Transformers and small-size comparison with other baselines. Perplexity is used for small-size models because it is a stable and predictable metric for language modeling, while end-task accuracy is reported for larger-size models. As indicated in previous work ([2304.15004]), perplexity is smoother and more robust than accuracy for small models.", "category": "Method_Disambiguation"}, {"question": "Was one-shot performance evaluated on two open-ended question answering tasks using 7B models?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the reported metrics for the evaluated tasks limited to recall, indicating whether answers are contained in the generated response?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is RWKV's performance described as comparable to Transformers in Table 1, according to the experimental findings?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does Table 1 accurately describe RWKV's capability for parallel computation according to its definition in the caption?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the evaluation include translation tasks?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "fYerSwf1Tb", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "HawkesVAE: Sequential Patient Event Synthesis for Clinical Trials", "authors": ["Chufan Gao", "Mandis Beigi", "Afrah Shafquat", "Jimeng Sun"], "abstract": "Generating sequential events data, such as adverse patient events, can provide valuable insights for clinical trial development, pharmaceutical research, patient modeling, and more. One approach to generate such data is by using generative AI models, which can synthesize data that resembles real-world data. However, in the domains such as clinical trials, patient data is especially limited. Data generation methods from literature such as LSTM, Probabilistic Auto-regressive, and Diffusion-based data generators struggle with this particular task off the shelf, as we show empirically. \nTo address this task, we propose HawkesVAE, a Variational Autoencoder (VAE) that models events using Hawkes Processes (HP). Hawkes Processes are specialized statistical models designed specifically for the task of event and time-gap prediction, and VAEs enable an end-to-end generative design. Additionally, traditional VAEs rely solely on embeddings to decode events, but in a data-limited setting, this approach can have issues fitting the data. Therefore, we experiment with different ways of allowing the decoder to access varying amounts of information from the input events. Our experiments show that HawkesVAE outperforms other methods in terms of fidelity and allows the generation of highly accurate event sequences in multiple real-world sequential event datasets with only a small amount of external information. Furthermore, our empirical experiments demonstrate that HawkesVAE generates data that allows for superior utility over existing baselines while maintaining privacy.", "keywords": "Synthetic Data Generation;Sequential Event Generation;VAE;Hawkes", "qa_pairs": [{"question": "In what context is it necessary to know or constrain the event types occurring within a sequence?", "answer": "In clinical trial settings, it is essential to ensure patient fidelity, meaning the generated patients must be significantly similar to the original patients for the data to be useful. Knowing which events occur in a patient to generate a similar patient would not be unreasonable. The 'Events Known' model was specifically designed to simulate only relevant events, ignoring irrelevant ones that could confuse the generator.", "category": "Method_Disambiguation"}, {"question": "How are the event time predictions for all known types combined to create a final multi-event sequence prediction, and why is the latent space of these distinct models summed?", "answer": "Since the event type is known, the specific encoder/decoder pair is indexed by the event index. The independent event time predictions for all known types are combined using their predicted event times to create a final multi-event sequence prediction. The latent space is summed because the VAE model requires a single embedding to sample a latent vector. The paper sums the patient embedding outputs of all expert event encoders and passes this joint embedding to the expert decoders. Each decoder is trained to generate the predicted time and type sequence of its respective expert event. Essentially, this approach is purely motivated by the paper’s limitation of having a fully generative model. The paper has to constrain the model to a constant-size embedding space, as it would be unfair to have varying-size latent spaces for each patient.", "category": "Motivation_Analysis"}, {"question": "How are the encoder and decoder functions specified in HawkesVAE, including the parameterization of $p_\\theta(|zS_z)$, $q_{\\phi}(z|S_z)$, and the use of $z$ in the log-likelihood $p_{\\theta}(S_z|z)$?", "answer": "The encoder model $E_{(\\mathcal{H}_i)} \\rightarrow \\hat{\\mu}, \\hat{\\sigma}$ takes in the original event types and times, and predicts the mean and standard deviation to sample $z \\sim Normal(\\hat{\\mu}, \\hat{\\sigma})$. This is used for $q_\\phi(\\cdot | S_z)$ in the loss function and follows the previous Transformer Hawkes Process implementation of a transformer encoder, with alternating sine and cosine waves to denote temporal information. The decoder model is more complicated. Hawkes Process is usually evaluated via one prediction step at a time (i.e., the input is the ground truth time-step and event type, and the task is to predict the next time step and event type). To meet the paper’s requirements, this approach is adapted into an autoregressive decoding scheme akin to the language modeling task. At training time, the input to the decoder $D(z, \\mathcal{H}_i) \\rightarrow (\\hat{t}_{i+1}, \\hat{k}_{i+1}, \\lambda)$ is hidden vector $z$ and a sequence of ground truth event types and times. It is tasked with predicting the next event type $\\hat{k}$, next event time $\\hat{t}$, and the $\\lambda$s necessary to compute $P_\\theta(\\cdot|S_z)$. At inference time, $z$ is the sole input to the decoder, from which the predicted event types and times are auto-regressively decoded. To predict next time and event tuple $(\\hat{t}_i, \\hat{k}_i)$, the input is the previously predicted times and events $\\{(\\hat{t}_1, \\hat{k}_1), \\dots, (\\hat{t}_{i-1}, \\hat{k}_{i-1}))\\}$. (each predicted time and event is repeatedly appended to the input).", "category": "Method_Mechanics"}, {"question": "What is the nature of the downstream classification task, and how is the death event handled in the dataset and model training?", "answer": "The downstream classification task is binary classification of death in the patient data. Death is not an explicit event in the dataset; the dataset contains events and timestamps of medications, procedures, and adverse events. The model is trained only on event prediction, and death is an external label available in the dataset. No regression is done on explicit time-to-death prediction.", "category": "Method_Disambiguation"}, {"question": "How is Hawkes-VAE implemented for event forecasting, and why does providing more information ('Events Known') result in less accurate event ordering?", "answer": "HawkesVAE is trained via next event prediction, similar to Transformer Hawkes, which also predicts the next time and event tuple $(t_i, k_i)$ using the true previous times and events $\\{(t_1, k_1), \\dots, (t_{i-1}, k_{i-1})\\}$ as input. The paper's main results for ML Efficiency are obtained through an autoregressive approach (where each predicted time and event is repeatedly appended to the input), since the authors do not have access to the true events for future steps. Providing more information in the 'Events Known' category surprisingly performs worse. This is because this approach essentially involves $num_types$ different Hawkes models, each individually modeling its respective event and event times, which are then combined to create a final multi-event sequence prediction. While its predictions are generally useful, the strict ordering of the events may not be accurately captured, which is why it performs worse in the strict forecasting sense. In contrast, the general HawkesVAE model is a multivariate Hawkes process that considers ALL events and times simultaneously, so its forecasting performance should be intuitively better.\n", "category": "Experimental_Analysis"}, {"question": "What are the specific events included in the datasets that are used to predict the external label of the death event?", "answer": "The events in the datasets include medications, procedures, and some adverse events like vomiting. These events and their occurrence times are used to predict if the subject experiences the death event, which is an external label.", "category": "Method_Disambiguation"}, {"question": "Does the model predict event types, and why does HawkesVAE require event lengths and event types at inference time? Are these inputs provided to baseline models as well?", "answer": "The model autoregressively predicts event types and times, with event lengths and event type information being optional inputs. The model without event type information is denoted as HawkesVAE (Multivariate), and the model with event type information is denoted as HawkesVAE (Events Known). The comparisons are fair because baseline models and HawkesVAE (Multivariate) are given the same amount of information.", "category": "Method_Disambiguation"}, {"question": "What are the primary points of novelty in the proposed method compared to existing approaches that combine VAEs with Hawkes processes or related methods like Transformer Hawkes or Cox-Hawkes?", "answer": "The primary points of novelty include: (1) Although the paper’s method applies VAE and Hawkes Processes together, its merit lies in its superior performance compared to synthetic tabular generation models, as well as LSTMs and PAR; (2) providing detailed user control over event types and times, which is valuable in applications like clinical trial patient generation[1,2] and not explicitly supported by other models; (3) testing on whole-sequence time and event generation, which is not addressed by Transformer Hawkes Models or Cox-Hawkes models; and (4) focusing on sequential event data generation, an underexplored field.\n\n[1] Beigi, M., Shafquat, A., Mezey, J., & Aptekar, J. (2022). Simulants: Synthetic Clinical Trial Data via Subject-Level Privacy-Preserving Synthesis. In AMIA Annual Symposium Proceedings (Vol. 2022, p. 231). American Medical Informatics Association.\n\n[2] https://www.medidata.com/wp-content/uploads/2023/10/FACT-SHEET-Simulants.pdf\n\n", "category": "Method_Comparison"}, {"question": "What are the formal mathematical details for the encoder and decoder models, including the equations for generating sequences and combining them in the encoder, as well as the autoregressive decoding scheme?", "answer": "The encoder model $E_{(\\mathcal{H}_i)} \\rightarrow \\hat{\\mu}, \\hat{\\sigma}$ takes in the original event types and times, and predicts the mean and standard deviation to sample $z \\sim Normal(\\hat{\\mu}, \\hat{\\sigma})$. This is used for $q_\\phi(\\cdot | S_z)$ in the loss function and follows the previous Transformer Hawkes Process implementation of a transformer encoder, with alternating sine and cosine waves to denote temporal information. The decoder model is more complicated. Hawkes Process is usually evaluated via one prediction step at a time (i.e., the input is the ground truth time-step and event type, and the task is to predict the next time step and event type). To meet the paper’s requirements, this approach is adapted into an autoregressive decoding scheme akin to the language modeling task. At training time, the input to the decoder $D(z, \\mathcal{H}_i) \\rightarrow (\\hat{t}_{i+1}, \\hat{k}_{i+1}, \\lambda)$ is hidden vector $z$ and a sequence of ground truth event types and times. It is tasked with predicting the next event type $\\hat{k}$, next event time $\\hat{t}$, and the $\\lambda$s necessary to compute $P_\\theta(\\cdot|S_z)$. At inference time, $z$ is the sole input to the decoder, from which the predicted event types and times are auto-regressively decoded. To predict next time and event tuple $(\\hat{t}_i, \\hat{k}_i)$, the input is the previously predicted times and events $\\{(\\hat{t}_1, \\hat{k}_1), \\dots, (\\hat{t}_{i-1}, \\hat{k}_{i-1}))\\}$. (each predicted time and event is repeatedly appended to the input).", "category": "Method_Mechanics"}, {"question": "Does the real-vs-synthetic classifier achieve a performance of 0.5 when applied to real data?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "IB1HqbA2Pn", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "LLaVA-Plus: Learning to Use Tools for Creating Multimodal Agents", "authors": ["Shilong Liu", "Hao Cheng", "Haotian Liu", "Hao Zhang", "Feng Li", "Tianhe Ren", "Xueyan Zou", "Jianwei Yang", "Hang Su", "Jun Zhu", "Lei Zhang", "Jianfeng Gao", "Chunyuan Li"], "abstract": "In this paper, we introduce LLaVA-Plus, an end-to-end training approach to systematically expanding the capabilities of large multimodal models (LMM), towards building general-purpose multimodal agents. It maintains a skill repository that contains a wide range of vision and vision-language pre-trained models as multimodal tools. Based on the user instruction and input image, LMM is trained to activate the appropriated tools when needed, grasping skills on the fly and aggregating the tool execution results to complete the real-world tasks in the wild. To facilitate the model capability on learning to use skills, we make the first attempt to build multimodal instruction-following data for tool use, covering skills in visual understanding, generation, external knowledge and their compositions. Empirical results show that LLaVA-Plus outperforms LLaVA in existing capabilities, and extends many new capabilities.\nCompared with large language model (LLM) based tool use methods, LLaVA-Plus is distinct in that the query image is considered throughout the entire interaction process, yielding higher multimodal tool use performance and enabling new scenarios.", "keywords": "Large Language Model;Large Multi-modal Model;Large Agent", "qa_pairs": [{"question": "Does the LLaVA-Plus model require retraining on an augmented dataset to master new tools or skills, and how does its flexibility compare to chain methods in terms of integrating and utilizing new tools?", "answer": " It's important to distinguish between 'skills' and 'tools' within the paper's model's framework. Skills are the model's capabilities, which can be achieved using various tools. When a new skill is required, it is true that the model needs to be retrained to effectively integrate and master this new skill. However, when it comes to replacing or upgrading tools, LLaVA-Plus exhibits greater flexibility. For example, if a more advanced version of a tool (like upgrading from CLIP-B to CLIP-L or switching from EasyOCR to Google OCR API) becomes available, the paper can seamlessly integrate this new tool into the model without the need for retraining. This approach ensures continuous improvement in the model's performance with minimal retraining requirements. Flexibility vs. Chain Methods: The paper's approach might not be as inherently flexible as chain methods in terms of tool use. However, the chain method comes with its own set of challenges, such as tool conflicts during inference. LLaVA-Plus is designed to avoid such conflicts, thereby ensuring more stable and reliable performance. Open-Source LMM Initiative: LLaVA-Plus is developed as an open-source LMM to further easily enhance the model capabilities. Tool chaining methods frequently require extensive context to describe tools, reducing overall model performance, especially when dealing with a large number of tools. The paper's framework is designed to bypass these issues, providing a more efficient and effective solution for integrating and utilizing a large tool set.", "category": "Method_Disambiguation"}, {"question": "Why is it necessary to augment a vision-language model, which can already directly perceive the visual world, with additional 'visual tools' and 'vision-language tools' such as OCR and Captioning models?", "answer": "There are primarily two strategies to improve a model’s capabilities: one is by enriching the training dataset, and the other is by integrating specialized external tools for specific tasks. While the former can be effective, it is often resource-intensive and costly. The paper’s focus has been on exploring the latter approach – utilizing expert tools – to strengthen the capabilities of LMM in a more efficient manner. The paper selects five tools to cover the essential image understanding abilities: RAM for image classification, Grounding DINO for object detection, Grounded-SAM for instance segmentation, BLIP2 for image caption, and EasyOCR for OCR. These tools are specifically designed for their respective tasks and may offer image representations that are complementary to LLaVA. By integrating these expert tools, the paper can significantly enhance the model’s performance in each of these key areas of image understanding. In the paper’s experiments, as shown in Table 5, it demonstrates that the incorporation of these tools not only improves the model’s performance in tasks like tagging and captioning but also adds new dimensions to its image understanding capabilities. The selection of these tools was driven by the objective to cover a wide range of essential image understanding abilities. Their integration into the vision-language model is not a redundancy, but rather a strategic enhancement to leverage the strengths of specialized tools, thereby enabling the model to perform a variety of tasks with greater accuracy and efficiency.", "category": "Motivation_Analysis"}, {"question": "What aspects of the paper demonstrate its novelty?", "answer": "The novelty of the paper lies in the conceptualization and effective execution of integrating Large Multimodal Models (LMM) with tools, which significantly enhances LMM capabilities. The authors are the first to explore the combination of LMM, tools, and training (LMM+tools+train), a path distinct from the widely used LLM+tools+chain approach. Additionally, the proposed visual instruction tuning approach and data format are innovative contributions that enhance the utility and clarity of LMM outputs.", "category": "Method_Disambiguation"}, {"question": "What does the term 'all tools + GPT4' in Table 5 refer to, and how is it used in the experiment?", "answer": "In Table 5, 'all tools + GPT4' refers to a setup where the outputs from all selected tools are integrated and then fed into GPT4. This process allows GPT4 to aggregate the diverse inputs and synthesize an answer. The approach is part of an oracle experiment aimed at demonstrating the potential effectiveness of tool use in enhancing the performance of a language model like GPT4 in processing and synthesizing multimodal information.", "category": "Concept_Understanding"}, {"question": "How does the performance of non-tool based multimodal models, such as the plain LLaVA, compare with tool-based approaches on the Llava-bench benchmark as shown in Tables 4 and 5?", "answer": "The plain LLaVA model, trained on visual instruction data without additional tools, serves as a non-tool based, multimodal baseline in Tables 4 and 5. Its performance is compared with tool-enhanced models to highlight the advantages of incorporating tools in multimodal modeling, demonstrating how tool-based approaches can enhance capabilities in processing complex, multimodal datasets.", "category": "Experimental_Analysis"}, {"question": "What are the comparative performance results of GPT4-v against LLaVA and LLaVA-Plus on the LLaVA-Bench?", "answer": "As shown in the table below, these results indicate that GPT4-v performs significantly better than both LLaVA and LLaVA-Plus. Even though LLaVA-Plus has been augmented with additional tools, this enhancement does not seem to bridge the performance gap.\n\n| Model | Conv | Detail | Reasoning | ALL |\n| --- | --- | --- | --- | --- |\n| LLaVA | 82.0 | 69.1 | 92.6 | 81.2 |\n| LLaVA-Plus | 81.6 | 74.5 | 95.7 | 83.9 |\n| GPT4V | 67.3  | 104.3 | 108.4 | 93.3 |", "category": "Experimental_Exposition"}, {"question": "What are the failure cases that demonstrate the limits of the agent's reasoning capabilities, and how do they highlight the model's limitations?", "answer": "Failure cases are presented in Figure 16 and Figure 17. In Figure 16, the model fails to comprehend the user’s intent, defaulting to a generic caption 'correspond object' for Grounding DINO instead of specifying the object likely to contain a cold beverage. In Figure 17, the model struggles with understanding spatial relationships between objects, showing inconsistent performance in accurately responding to spatial queries. These cases highlight limitations in the model's intent comprehension and spatial reasoning capabilities.", "category": "Experimental_Analysis"}, {"question": "Does adding a new tool to LLaVA-Plus require instruction tuning from scratch, or can the model be fine-tuned only for the new instructions?", "answer": "When fine-tuning LLaVA-Plus on a particular tool, it was observed that while the model’s performance in that specific area improves, it adversely affects its abilities in handling other tools. For instance, after fine-tuning solely on grounding data, the model's proficiency in tagging, captioning, and OCR significantly diminished. Skills represent the model's capabilities and may require a complete retraining process for effective integration of new skills. Conversely, the model is more adaptable when it comes to incorporating upgraded tools. For instance, replacing an existing tool with its advanced version (e.g., upgrading from CLIP-B to CLIP-L) can be done seamlessly without necessitating a full retraining of the model. Therefore, if the goal is to add a new skill, a comprehensive fine-tuning across all tools would be necessary to maintain the model's overall proficiency. On the other hand, upgrading tools can be achieved with minimal impact on the model’s existing capabilities.", "category": "Method_Disambiguation"}, {"question": "How does the model handle conflicting grounding information from different tools, such as when detection and segmentation tools disagree about the presence of a particular object?", "answer": "The model associates each skill with a specific tool, eliminating tool conflicts within the same skill. However, conflicts can arise between different skills due to ambiguities in human instructions. To address this, the data processing pipeline employs two strategies: (1) Diversifying Questions: The dataset includes diverse questions to reduce ambiguities and improve tool use accuracy. (2) Quality Control in Data Generation: Data is generated using GPT4V, which, while powerful, can produce hallucinations or undesirable data items. To ensure data quality, heuristic methods are applied to filter out low-quality data, such as removing apologetic responses, excluding irrelevant content, validating CLIP-Retrieval data, and ensuring accuracy in object detection and segmentation tasks. These measures minimize conflicts and ambiguities, ensuring reliable and accurate outputs.", "category": "Method_Mechanics"}, {"question": "How does the paper define and use the terms 'planning' and 'tools', and how does this align with the broader context of multimodal agents and representation learning?", "answer": "In the context of the paper, 'planning' refers to the process by which the model determines the most appropriate tool (APIs and systems) to use for a given task, aligning with terminologies used in projects like HuggingGPT and AutoGPT. The term 'tools' describes the APIs and systems the model interfaces with, consistent with works like ToolFormer and Visual ChatGPT. Regarding representation learning, the submission category was chosen to reflect the multimodal nature of the model and its applicability across various domains, though the authors are open to reconsidering the category if a more suitable one is suggested.", "category": "Concept_Understanding"}, {"question": "How does expanding to new skills affect LLaVA-Plus's performance in previously acquired skills?", "answer": "Expanding to new skills yields mixed results for LLaVA-Plus. The tool use rate for grounding improved with the addition of more tools, while there was a slight decrease in caption performance. This is attributed to the model's enhanced focus on tool-specific tasks with more diverse training data, which may compromise performance in tasks with more general or overlapping features, such as captioning. Detailed results are available in Table 13.", "category": "Experimental_Exposition"}, {"question": "Does LLaVA-Plus have the capability to perform error correction or validation to prevent errors from sub-models from propagating downstream?", "answer": "Yes, LLaVA-Plus can theoretically perform error correction due to its Large Multimodal Model (LMM) nature, which allows it to process and interpret image inputs directly. The visual information provided by these inputs can play a crucial role in enabling the model to identify and rectify errors or inconsistencies in the outputs of sub-models. A practical instance of this error correction ability is demonstrated in Table 12, focusing on the Grounding DINO tool. Observations indicate that the original Grounding DINO model occasionally produces false positive predictions. However, LLaVA-Plus has shown the capability to filter out these inaccurate predictions effectively. This results in enhanced accuracy and reliability, as the model leverages its multimodal understanding to validate and correct the tool outputs.", "category": "Experimental_Analysis"}, {"question": "What is the approximate amount of data curation required for each skill in the model, and how does this quantity impact the scalability of the process?", "answer": "Approximately 5000 samples are generally sufficient to effectively train the model in invoking the necessary tools for a given skill. This quantity balances comprehensiveness and manageability, ensuring accurate tool utilization while keeping the curation process feasible.", "category": "Experimental_Analysis"}, {"question": "What is the objective function of the auto-regressive model, and how is it designed to facilitate learning within the assistant and feedback loop framework?", "answer": "The paper's approach involves constructing data in a text format suitable for training an auto-regressive model. This model is designed to learn how to accurately answer user questions.\n*Data Representation and Training Approach*:\nEach interaction is represented as a single sequence in the paper’s dataset, as exemplified in Table 1. This sequence encapsulates the entire dialogue flow, including both the human queries and the assistant's responses. The format of the data, as illustrated in Equation (2), is structured as follows:\n\n$Human : I_q < n> X_q<STOP> Assistant : X_{skill\\underline{}use}<STOP>$\n\n$Human : X_{skill\\underline{}result}<STOP> Assistant : X_{answer}<STOP>$\n\nIn this representation, $I_q$ and $X_q$ denote the initial query and its context, $X_{skill\\underline{}use}$ represents the assistant's tool usage, $X_{skill\\underline{}result}$ is the outcome of the tool use, and $X_{answer}$ is the assistant's final response.\n*Objective Function and Loss Calculation*:\nThe loss is specifically computed on the $X_{skill\\underline{}use}<STOP>$ and $X_{answer}<STOP>$ tokens. This approach ensures that the model is trained to predict these critical tokens accurately, which are essential for effective tool use and generating accurate responses. By focusing the loss calculation on these parts of the sequence, the objective function aligns the model's learning with the key task of providing relevant and accurate tool-based assistance in response to user queries. The objective function is tailored to facilitate the learning of an effective dialogue flow, emphasizing the correct usage of tools and the generation of appropriate responses, thereby enabling the auto-regressive model to function effectively within the assistant and feedback loop framework.\n", "category": "Method_Mechanics"}, {"question": "How does LLaVA-Plus differ from LLaVA and ToolFormer in terms of functionality and capabilities?", "answer": "*Comparison with LLaVA:*\n\n- LLaVA primarily supports conversational interactions based on images. However, its output is limited to text.\n- LLaVA-Plus extends the capabilities of LLaVA to include more diverse functionalities such as segmentation, editing, retrieval, and generation. This expansion is achieved by integrating external tools, thereby enhancing LLaVA's original text-only output to support a broader range of multimodal applications.\n\n*Comparison with ToolFormer:*\n\n- ToolFormer, in contrast, is designed primarily for text-based applications. It supports a range of text-oriented tools like question answering, Wikipedia search, calculators, calendars, and machine translation.\n\n- Unlike LLaVA and LLaVA-Plus, ToolFormer is a pure language model and does not inherently support multimodal functionalities. Even with the integration of image-related tools, ToolFormer lacks the capability to process image inputs directly. This limitation means that ToolFormer cannot independently verify or interact with tool results that are based on visual content.\n\nWhile LLaVA-Plus shares some foundational aspects with LLaVA and ToolFormer, it distinguishes itself through its enhanced multimodal capabilities and its ability to integrate and leverage external tools for a wider range of applications, including those involving direct image processing and interaction.", "category": "Method_Comparison"}, {"question": "Was the paper submitted under the category 'representation learning for computer vision, audio, language, and other modalities' based on the multimodal nature of the model and its applicability across various domains?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper use the term 'planning' to refer to achieving a specific goal through a sequence of actions as commonly understood in traditional planning approaches?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the term 'tools' used in the paper to describe robotic planning and object-affordance exploitation?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "JVeM7uwDwK", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Revealing the Illusion of Joint Multimodal Understanding in VideoQA Models", "authors": ["Ishaan Singh Rawal", "Shantanu Jaiswal", "Basura Fernando", "Cheston Tan"], "abstract": "While VideoQA Transformer models demonstrate competitive performance on standard benchmarks, the reasons behind their success are not fully understood. Do these models jointly capture and leverage the rich multimodal structures and dynamics from video and text? Or are they merely exploiting shortcuts to achieve high scores? Hence, we design $\\textit{QUAG}$ (QUadrant AveraGe), a lightweight and non-parametric probe, to critically analyze multimodal representations. QUAG facilitates combined dataset-model study by systematic ablation of model's coupled multimodal understanding during inference. Surprisingly, it demonstrates that the models manage to maintain high performance even under multimodal impairment. We extend QUAG to design ''QUAG-attention'', a simplistic and less-expressive replacement of self-attention. We find that the models with QUAG-attention achieve similar performance with significantly less mulops without any finetuning. These findings indicate that the current VideoQA benchmarks and metrics do not penalize models that find shortcuts and discount joint multimodal understanding. Motivated by this, we propose the $\\textit{CLAVI}$ (Counterfactual in LAnguage and VIdeo), a diagnostic dataset for coupled multimodal understanding in VideoQA. CLAVI consists of temporal questions and videos that are augmented to curate balanced counterfactuals in language and video domains. We evaluate models on CLAVI and find that all models achieve high performance on multimodal shortcut instances, but most of them have very poor performance on the counterfactual instances that necessitate joint multimodal understanding. Overall, with the multimodal representation analysis using QUAG and diagnostic analysis using CLAVI, we show that many VideoQA models are incapable of learning multimodal representations and that their success on standard datasets is an illusion of joint multimodal understanding.", "keywords": "video question answering;multimodal representation learning;interpretability;vision-language model", "qa_pairs": [{"question": "What is the motivation behind selecting self-attention based fusion transformers instead of cross-attention in this study?", "answer": "Self-attention was chosen because it is present in almost all fusion architectures and both modalities are interleaved in the attention operation. Additionally, popular interpretability methods like VL-InterpreT[1], attention rollout, and attention flow[2] are based on self-attention architectures. Cross-attention, on the other hand, segregates (TV, VT) modality interactions into distinct attention units, making it a trivial case for analysis.\n\n[1]Aflalo, E et al., \"VL-interpret: An interactive visualization tool for interpreting vision-language transformers\", CVPR 2022.\n\n[2]Samira A. et al.,. Quantifying Attention Flow in Transformers. ACL 2020", "category": "Motivation_Analysis"}, {"question": "What is the novelty of QUAG compared to existing techniques that use averaging in self-attention based fusion transformers?", "answer": "The novelty of QUAG lies in its interpretability insight. While previous works have used averaging for either token pruning (for example, TESTA [1]) or multimodal visualization studies [2] (for example, VL-Interpret), the paper's (i) intention, (ii) method, and insights are substantially different and insightful from previous works:\n**Intention:**\nShortcuts are (i) spurious features in a dataset (ii) that are learned by the model [3]. QUAG analyzes the dependence of a model on specific types of modality interaction (for example, cross-modal interaction) through ablations. Therefore, if a model achieves high accuracy with QUAG, it is relying on shortcuts for its high accuracy. Since QUAG impairs token mixing without any fine-tuning, it does not augment the learned weights (representations) of the model. This facilitates analyzing the representations learned by a model on a dataset.\n**Method and insights:**\nMultimodal visualization studies like [2] average individual attention sub-matrices to classify heads as uni/cross-modal to visualize feature importance. \n1. The paper's first contribution in QUAG is that consistent row-wise averaging of the sub-matrices across all heads and layers has an effect of impairment (that is, short-circuiting). Replacing an entire submatrix with an average value of the submatrix (and not row-wise average), as in [2], cannot faithfully short-circuit specific modality. While visualization methods can provide some interesting qualitative results, QUAG is the only method that provides quantitative dependence on modality interactions through its ablation.\n2. The paper's second contribution is QUAG-attention, which calculates attention on already-averaged keys to reduce the size of the attention matrix with (i) hardly any drop in performance and (ii) not relying on 'already computed attention scores in self-attention based fusion transformers'. Hence, (i) without any fine-tuning and optimization for performance, and (ii) leveraging the biased representations learned by the model using first-principle analysis of self-attention, QUAG prunes the size of the attention matrix by as little as 45% and as much as 90% (assuming $L_\\mathcal{V} = L_\\mathcal{T} = 10$) with minimal performance drop.\n\n[1] Ren, S. et al., \"TESTA: Temporal-Spatial Token Aggregation for Long-form Video-Language Understanding\", EMNLP 2023\n[2] Aflalo, E et al., \"VL-interpret: An interactive visualization tool for interpreting vision-language transformers\", CVPR 2022.\n[3] Murali, N. et al., \"Shortcut Learning Through the Lens of Early Training Dynamics\", Workshop on Spurious Correlations, Invariance, and Stability, ICML 2023", "category": "Method_Comparison"}, {"question": "What is the motivation for introducing CLAVI?", "answer": "The paper's motivation to introduce CLAVI was that 'many observations may still be related to dataset bias rather than the model bias' (Section 3, titled 'Does multimodal sub-optimality stem from dataset biases?'). CLAVI warrants joint multimodal understanding and penalizes shortcut learning.", "category": "Motivation_Analysis"}, {"question": "How does the dataset design address potential issues related to temporal ordering and boundary cues in video segments, given that the dataset is created by changing the order of video segments?", "answer": "The dataset design addresses potential issues related to temporal ordering and boundary cues through the following strategies:\n1.  **Temporal Separation and Duration:** While many models operate on sparsely sampled frames (e.g., 3 frames for the All-in-one model) and perform well on standard benchmarks, this dataset further reduces reliance on boundary cues by ensuring instances have a minimum separation of 4 seconds between action classes. Additionally, each action occurs for a sufficiently long duration (average length: 7.7 seconds), mitigating issues related to temporal ordering ambiguity.\n2.  **Action Class Independence:** The dataset is curated so that each action class within a pair involves different objects and verbs (e.g., \"turning on light\" and \"holding clothes\"). This composition ensures a degree of independence between the classes, making their temporal order less critical for distinguishing them.\n3.  **Question Type and Balance:** Only questions of existence, before/after, and beginning/end types—which do not necessarily require precise boundary information—are included. Furthermore, the questions are balanced with the video instances to prevent boundary information from serving as a shortcut for answering.\nBy incorporating these elements—(1) sufficient separation between actions, (2) independence between action classes, and (3) balanced instances paired with simple questions—the paper ensures the dataset is capable of diagnosing joint multimodal understanding in VideoQA models.\n", "category": "Method_Mechanics"}, {"question": "What is the explaination about the correctness of the indices used in the QUAG operator's formulation and clarify the meaning of $[s_1 ... s_m]$ and terms like $\\mathcal{T}\\mathcal{T}$, $\\mathcal{T}\\mathcal{V}$ in the context of the equation on page 3?", "answer": "Yes, the formulation is correct. $[R(Z,W)]_{ij}$ refers to the element $(i,j)$ in the row-wise transformed matrix, $R(Z,W)$. The definition is applied $i \times j$ times. If $(i,j)$ does not lie in the submatrix W, $[R(Z,W)]_{ij}$ remains $[Z]_{ij}$. If $(i,j)$ lies in the submatrix, row-wise averaging is performed.That is, for the given row $i$, the mean of all the values in that row, indexed by k, is taken. Since we are averaging only across the columns in the sub-matrix, iteration is only performed over columns. Note that this can be efficiently implemented as a batched-matrix operation in pytorch. $[s_1 ... s_m]$ refers to a list of quadrants, and $\\mathcal{T}\\mathcal{T}$, $\\mathcal{T}\\mathcal{V}$, etc., correspond to $A_{\\mathcal{T}\\mathcal{T}}$, $A_{\\mathcal{T}\\mathcal{V}}$, etc. ", "category": "Concept_Understanding"}, {"question": "Was the usage of CLS tokens for fusing token representations investigated, given that averaging is used for this purpose in the study?", "answer": "The paper did not aim to compare fusion techniques. Instead, its focus was on using averaging of submatrices to disrupt specific modality interactions (QUAG). QUAG can be used with any fusion method built on top of self-attention, including CLS-based methods. For example, the JustAsk model uses CLS tokens to infer the final answer, and applying QUAG-attention to JustAsk demonstrated even after reducing the size of the attention matrix by 90% (that is retaining only the average text key and average video key) and saving 68% mulops, there is hardly any drop in the accuracy.", "category": "Method_Disambiguation"}, {"question": "Is the evaluation of QUAG and QUAG-attention on only two models sufficient to conclude their effectiveness in analyzing the reliance of models on biased representations?", "answer": "QUAG and QUAG-attention are methods to effectively introduce the reliance of the model on its biased representations. The paper did not make any claims about efficiency. Shortcuts are (i) the spurious features in a dataset (ii) that are learnt by the model [1]. QUAG analyzes the dependence of a model on specific types of modality interaction (for example, cross-modal interaction) through ablations. Therefore, if a model manages to achieve high accuracy with QUAG, it is relying on shortcuts for its high accuracy. Hence, QUAG enables a combined dataset-model analysis. The paper chose 4 standard datasets (ActivityNet-QA, MSRVTT-QA, NeXT-QA, ATP-Hard) and 2 popular standard baseline models in VideoQA (JustAsk - ICCV’21 Oral; FrozenBiLM – NeurIPS’22). Therefore, 2 models and 4 datasets, that is 8 instances, are considered as representative to expose that the performance of many dataset-model combinations do not rely on multimodal understanding. Furthermore, the paper proves the mathematical correctness of QUAG , and extends it to QUAG-attention that leverages the biased representations through skinny attention matrices, without significant performance drop; reasserting the correctness and efficacy of QUAG. The paper has now also included the results for All-in-one+ (CVPR’23) on ActivityNet-QA and MSRVTT-QA datasets . It contains 1/8 parameters than that of FrozenBiLM. The paper could not finetune it on NeXT-QA because of paucity of time. The paper’s findings are consistent with the authors’ claims on improved multimodal understanding. However, since the model just uses 3 randomly sampled frames, it might be insufficient to have high consistent accuracy on CLAVI.\n\n|      | A-QA | M-QA |\n|-------------|------|------|\n| Acc         | 41.9 | 43.1 |\n| vid_only    | 14.2 | 4.2  |\n| text_only   | 23.5 | 20.8 |\n| SC: unimodal| 11.4 | 3.8  |\n| SC: crossmodal | 20.6 | 27.6 |\n| SC: video   | 19.2 | 12.7 |\n| SC: text    | 5.6  | 7.3  |\n\n[1] Murali, N. et al., \"Shortcut Learning Through the Lens of Early Training Dynamics\", Workshop on Spurious Correlations, Invariance, and Stability, ICML 2023", "category": "Experimental_Analysis"}, {"question": "Does the CLAVI dataset ensure that the models are free from shortcuts, or does it only penalize shortcuts over joint multimodal understanding?", "answer": "The CLAVI dataset does not ensure that models are free from all biases or shortcuts. Instead, it is designed to penalize shortcuts over joint multimodal understanding through consistent accuracy metrics. Low consistent accuracy on CLAVI confirms the lack of joint multimodal understanding, as observed in 3 out of the 4 tested models (i.e., Singularity-Temporal, All-in-one+, JustAsk). High CLAVI results are necessary but not sufficient to conclude joint multimodal understanding. This is further validated by QUAG results, which show that FrozenBiLM’s higher consistent performance on CLAVI is indeed due to joint multimodal understanding. Specifically, the consistent accuracy of FrozenBiLM on QUAG drops significantly under multimodal impairment (Section 3.2, Paragraph 3).", "category": "Method_Disambiguation"}, {"question": "How do the proposed probes (QUAG and QUAG-Attention) compare in effectiveness to evaluating model generalization ability through zero-shot settings for diagnosing models?", "answer": "While the generalization abilities can be used for model diagnosis (say, through zero-shot evaluation), the additional testing data might also contain the biases [1] that the model learnt during pretraining (hence, they are shortcuts). Since the accuracy metrics don’t typically penalize the shortcut learning, these methods cannot be directly used for diagnosis. Consider the zero-shot results from CLAVI below. The balanced accuracy is close to random chance, 50%. This is because the models either always predict the answer as 'yes' (JustAsk) or 'no' (FrozenBiLM, Singularity-T) (ans frequency bias). The consistent accuracy on the counterfactual subset confidently unveils the illusion of joint multimodal understanding in the model. Furthermore, QUAG and QUAG-Attention, unlike generalization methods (i) do not rely on any external dataset, and (ii) pinpoints the exact failure modes in the models.\n\n|        | JustAsk | FrozenBiLM | Singularity-T |\n| :----------- | :------: | :---------: | :-----------: |\n| Balanced Acc |   49.7  |      50     |       50       |\n| $CAcc_{V}-counter$     |    0.2  |      0.1    |       0        |\n| $CAcc_{V}-counter$    |    0.8  |      0.2    |       0        |\n\n\n[1]Guo, Y. \"Loss re-scaling VQA: Revisiting the language prior problem from a class-imbalance view.\" IEEE Transactions on Image Processing (2021).", "category": "Method_Comparison"}, {"question": "What are the mathematical definitions of video-consistency and text-consistency?", "answer": "Video-consistency is defined as {(F(V, Q) = $A_{V,Q}$) AND (F(V’, Q) = $A_{V’,Q}$)}, where V is a video, Q is a question with answer $A_{V,Q}$. V’ is the counterfactual video, and F is the VideoQA model that maps a video and question pair to an answer. Text-consistency is defined as {(F(V, Q) = $A_{V,Q}$) AND (F(V, Q’) = $A_{V,Q’}$)}, where Q’ is the counterfactual question.\n", "category": "Concept_Understanding"}, {"question": "Is it reasonable to replace attention blocks progressively in the proposed model to identify where performance drops, given that different blocks (e.g., early-fusion vs. late-fusion of modalities) may behave differently?", "answer": "Probes like QUAG can enhance the understanding of deep models. In fact, other works, such as [1], also use similar formulations of bi-modal attention matrices by using gradient-weighted attention maps for visualization. The paper's results from QUAG are consistent with those from QUAG-attention and CLAVI. Yes, it is quite possible that the blocks can behave differently. Note that progressive-QUAG won’t ablate the modality interactions completely because the downstream blocks would mix the tokens within the modality (through the residual input), but it can provide insights about the relative importance of upstream fusion blocks. The paper provides the results for progressive-QUAG on FrozenBiLM on the NeXT-QA and ActivityNet-QA datasets (Appendix A.2.8). Progressive-QUAG further underscores the importance of the first half of the fusion layer in the model's performance. Overall, the results are consistent and supplement the paper's findings.\n\nReferences: [1] Chefer, H. et al. 'Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers.' ICCV 2021.", "category": "Method_Disambiguation"}, {"question": "Why are the short-circuiting results in Table 1 not conclusive for the JustAsk model, and what does this imply about its utilization of multimodal dataset features?", "answer": "The original dataset-centric belief was that shortcuts are spurious features present in a dataset. Rather, in the paper's work, it presents a unified dataset-model analysis of shortcuts as the spurious features present in a dataset that are learned by the model. The short-circuiting operations that cause the drop in dataset-model combinations is summarized below (Uni is unimodal and cross is crossmodal SC):\n| |ActivityNetQA  |MSRVTT-QA  |NeXT-QA  |ATP-H | \n|-|-|-|-|-| \n|JustAsk   |\\- |\\-  |\\-   |\\-   | \n|FrozenBiLM|Text, Uni, **Cross**|Text, Uni, **Cross**|Text, Uni|Text, Uni| \n\nThe idea is that if a specific token mixing operation is important, short-circuiting it should decrease the performance. As shown above, none of the models are learning and leveraging the core features from video modality interactions. JustAsk model does not exploit the rich multimodal structure of the datasets to answer those datasets. FrozenBiLM, however, consistently relies on unimodal and text modality interactions. However, for ActivityNet-QA and MSRVTT-QA (and not NeXT-QA), it also relies on crossmodal interactions. This means that FrozenBiLM does not learn rich crossmodal representations through NeXT-QA and ATP-Hard, but does learn it for ActivityNet-QA and MSRVTT-QA.", "category": "Experimental_Analysis"}, {"question": "Why is fine-tuning needed for the proposed benchmark, and how does the approach address concerns about temporal discontinuities in permuted video segments?", "answer": "Temporal discontinuity is not a problem for the paper's setting because:\n1.  Most of the models operate on sparsely sampled frames (for example, 3 frames for the All-in-one model) but still manage to achieve high performance on standard benchmarks. However, to further reduce the dependence on boundary, the paper only considers the instances in which the action classes are separated by at least 4 seconds and each action occurs for sufficiently long duration (average length: 7.7 seconds).\n2.  While curating the dataset, the paper ensures that each action class in a pair is composed of different objects and verbs (for example, turning on light and holding some clothes), so that swapping the order of events makes them loosely independent of each other.\n3.  The paper only considers questions of existence, before/after, beginning/end types that do not necessarily require boundary information. \n\nAlso, the balanced questions and videos ensure that the boundary information does not provide shortcut information.\nWith sufficient (1) separation between the actions, (2) independence between the action classes, (3) balanced instances with simple questions, the paper ensures that the dataset is capable of diagnosing joint multimodal understanding in VideoQA models.\n", "category": "Method_Mechanics"}, {"question": "Why are the positive and negative QAs not balanced in the CLAVI dataset as mentioned on Page 8?", "answer": "The positive and negative QAs are not balanced because CLAVI exhaustively covers all possible permutations of counterfactuals, including negative control questions whose answers are always 'no', as explained in Section 3.1, Paragraph 3, Page 7. That is, for a given temporal question Q and video V, and their counterfactuals Q’ and V’, CLAVI consists of all – (Q,V), (Q,V’), (Q’,V) and (Q’,V’) instances. ", "category": "Concept_Understanding"}, {"question": "How does CLAVI ensure that there are no ambiguities in video-question pairs, especially in cases where the altered video might lead to conflicting interpretations, such as the example in the example shown with Fig 2, 'holding on clothes' and 'turning on a light'; when the altered video starts, the light is already on, so saying that the light was turned on at the beginning is not completely wrong?", "answer": "CLAVI is curated to include all necessary information in the video to answer the questions unambiguously. The interpretation of the question *'Did turning on a light happen before holding on to clothes?'* is based on the sequence of events in the video, where *'turning on a light'* occurs before *'holding on to clothes'*. This interpretation is consistent across training and testing phases, and finetuning ensures the model understands the question unambiguously.", "category": "Experimental_Analysis"}, {"question": "How is ambiguity in negative control questions, such as the example in Table 2, addressed in the dataset?", "answer": "Ambiguity in negative control questions is addressed by randomly sampling an action class that is not composed of any of the (i) verbs or (ii) objects that occur in the untrimmed video (original video). This scheme is consistently employed to curate the entire dataset, ensuring that all questions are designed with the information required to answer them completely contained within the video. This approach is consistent with many popular datasets. For example, in Activitynet-QA, 'is the boy in white squatting before diving?' the answer is 'no' (question id: v_Bs3OMhhUlY4_3).", "category": "Method_Mechanics"}, {"question": "What is the rationale for selecting JustAsk and FrozenBiLM as the baseline models, given that they have been outperformed by newer models?", "answer": "JustAsk and FrozenBiLM were selected because they are (i) standard baseline models in VideoQA, (ii) compute-friendly to experiment with, (iii) many VideoQA models lack open-sourced code and checkpoints, and (iv) they still perform well on the datasets considered and on newer benchmarks[2]. Additionally, they are used to demonstrate the issue of the illusion of joint multimodal understanding in VideoQA models, and the findings are verified using CLAVI. For instance, Results for All-in-one+ (CVPR ‘23) on ActivityNet-QA and MSRVTT-QA datasets are also included in Appendix A.2.8, though it contains 1/8 the parameters of FrozenBiLM and was not fine-tuned on NeXT-QA due to time constraints. The paper’s findings are consistent with the authors’ claims on improved multimodal understanding. However, since the model just uses 3 randomly sampled frames, it might be insufficient to achieve high consistent accuracy on CLAVI.\n\n| Metric             | A-QA | M-QA |\n| :----------------- | :---: | :---: |\n| Acc                | 41.9  | 43.1  |\n| vid_only           | 14.2  | 4.2   |\n| text_only          | 23.5  | 20.8  |\n| SC: unimodal       | 11.4  | 3.8   |\n| SC: crossmodal     | 20.6  | 27.6  |\n| SC: video          | 19.2  | 12.7  |\n| SC: text           | 5.6   | 7.3   |\n\n[1]Xiao, J. et al, \"Can I Trust Your Answer? Visually Grounded Video Question Answering.\" arXiv preprint arXiv:2309.01327 (2023).\n\n[2]Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Ge Yuying, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified video-language pretraining. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,2023a.\n", "category": "Motivation_Analysis"}, {"question": "Did the paper conduct ablation studies to isolate the influence of the data itself rather than the methodology, particularly to determine whether video or text inherently poses greater learning challenges?", "answer": "The authors analyzed the difference between attention scores learned in video-consistent and text-consistent pairs in CLAVI, as detailed in the representation sensitivity analysis in A.3.7. The results indicate that FrozenBiLM learns more distinct representations for counterfactual videos than for counterfactual text (Table 16, Figure 8). ", "category": "Experimental_Exposition"}, {"question": "Are there ambiguities in the video-question pairs when videos are altered, such as in the case where the light is already on at the start of the altered video?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the paper claim that Short-circuit and QUAG-attention are efficient methods based on the evaluation of two models?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the positive and negative QAs balanced in the dataset due to the exhaustive coverage of counterfactual permutations?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was All-in-one+ finetuned on NeXT-QA due to time constraints?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the claim that high CLAVI results alone are sufficient to conclude a model has joint multimodal understanding?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the mathematical correctness of QUAG proven in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the performance on the CLAVI dataset guarantee that a model is free from all shortcuts?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the maximum input sequence length for the multimodal fusion module denoted as $l$?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "DzxaRFVsgC", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest", "authors": ["Shilong Zhang", "Peize Sun", "Shoufa Chen", "Minn Xiao", "Wenqi Shao", "Wenwei Zhang", "Yu Liu", "Kai Chen", "Ping Luo"], "abstract": "Visual instruction tuning large language model(LLM) on image-text pairs has achieved general-purpose vision-language abilities. However, the lack of region-text pairs limits their advancements to fine-grained multimodal understanding. In this paper, we propose spatial instruction tuning, which introduces the reference to the region-of-interest(RoI) in the instruction. Before sending to LLM, the reference is replaced by RoI features and interleaved with language embeddings as a sequence. Our model GPT4RoI, trained on 7 region-text pair datasets, brings an unprecedented interactive and conversational experience compared to previous image-level models. (1) Interaction beyond language: Users can interact with our model by both language and drawing bounding boxes to flexibly adjust the referring granularity. (2) Versatile multimodal abilities: A variety of attribute information within each RoI can be mined by GPT4RoI, e.g., color, shape, material, action, etc. Furthermore, it can reason about multiple RoIs based on common sense. On the Visual Commonsense Reasoning (VCR) dataset, GPT4RoI achieves a remarkable accuracy of 81.6\\%, surpassing all existing models by a significant margin (the second place is 75.6\\%) and almost reaching human-level performance of 85.0\\%. \nThe code, dataset, and demo can be found at https://github.com/Anonymous-Researcher1/GPT4RoI.", "keywords": "vision language model;large language model;instruction tuning;visual commonsense reasoning", "qa_pairs": [{"question": "Which specific component of the GPT4RoI model design contributes to positional awareness, and how does it function within the model?", "answer": "The core design of GPT4RoI incorporates a region feature extractor, which replaces the ROI reference in the instruction with the extracted region feature. The implementation of the region feature extractor can vary, such as using deformable attention or RoI Align operation, but in this study, RoI Align was chosen due to its simplicity and familiarity. The region feature extractor is fundamental in object detection and instance segmentation, and its details and parameters have been thoroughly ablated. The following are the origin and significance of each sub - module: \n1. The Scale Shuffle module is an algorithm that the authors have chosen from the multi - level feature fusing modules [1, 2, 3]. These modules are typically utilized to provide robust feature representation for regions of varying scales. They play a crucial role as an essential component in enhancing the performance of all instance recognition tasks. \n2. RoI Align is an operation that is employed to extract region features by utilizing the coordinates of a bounding box. It serves as the fundamental module in R - CNN detectors preceding DETR - style ones. \n3.When absolute position information is required in a task and to address the problem of translation invariance in CNNs, a commonly employed technique, as described in [4, 5, 6], is to augment feature maps with feature coordinates.\n4. When aiming to extract more detailed information within a region, the authors usually increase the RoI Align resolution, such as from (7x7) in object detection [7] to (14x14) in instance segmentation [8]. \n\n[1] Scale-Equalizing Pyramid Convolution for Object Detection\n\n[2] Unifying Object Detection Heads with Attentions\n\n[3] Dense Distinct Query for End-to-End Object Detection\n\n[4] An intriguing failing of convolutional neural networks and the coordconv solution\n\n[5] SOLO: Segmenting Objects by Locations\n\n[6] SOLOv2: Dynamic and Fast Instance Segmentation\n\n[7] Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks\n\n[8] Mask R-CNN for Object Detection and Segmentation", "category": "Method_Mechanics"}, {"question": "Why does the model produce only short descriptions, and is this a limitation of the dataset or the model itself?", "answer": "The model produces only short descriptions because all open-source data includes brief text descriptions, which is a common issue for current large vision language models. Approaches like LLaVA[1] and MiniGPT4[2] use ChatGPT/GPT4 to refine these short captions, but this has limited significance as the additional descriptions inferred by LLMs often lack accuracy or informative value. Achieving accurate and meaningful long context requires human participation in data annotation to establish large datasets, which is beyond the scope of this paper.\n\n[1] Visual Instruction Tuning\n\n[2] MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models", "category": "Experimental_Analysis"}, {"question": "What are the key challenges and distinctions of methods that use textual coordinates as grounding tokens compared to incorporating detection functions into LLMs?", "answer": "The key distinction lies in whether to incorporate the detection function into the LLM. For the method that uses textual coordinates as the grounding token, it has to solve the following challenges: 1. Aligning a large number of position tokens with their corresponding positions in the image by training on a large set of datasets. However, this is actually a simple rule that can be naturally implemented with the operation in detection architectures. 2. Modeling geometric properties can be challenging. For example, if the ground truth box is $<x_1 = 0, y_1 = 0, x_2 = 5, y_2 = 5>$, a predicted box of $<x_1 = 1, y_1 = 1, x_2 = 4, y_2 = 4>$ would be considered a better result than $<x_1 = 1, y_1 = 1, x_2 = 8, y_2 = 8>$ because it has a higher overlap with the ground truth. However, incorporating this geometric property into the next token prediction task using cross - entropy loss can be challenging. On the other hand, utilizing traditional loss functions such as L1 or IoU loss can naturally handle this geometric constraint. 3. Dense to Sparse [1][2] is a crucial design for detection performance, but embedding such an idea into the sequential form of LLM is challenging. Another approach is to use an external detector to find the potential region of interest, whereas LLM only focuses on analyzing the corresponding region of interest. This is the motivation of GPT4RoI. It requires much less data and allows for quick adaptation to specific domain problems with the corresponding detector. However, the drawback is that the framework may appear less elegant and it assumes input contains all regions of interest that need to be analyzed. Both approaches have their advantages and disadvantages, and academic research in both directions is thriving. It is still too early to make a definitive judgment on which approach has absolute superiority at this time. \n\n\n[1] Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks\n\n[2]DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection", "category": "Method_Comparison"}, {"question": "What is the novelty and significance of expanding from image-level to region-level instruction tuning in GPT4RoI, given that some other papers also explore the region-level large language models [1] but lack the performance comparison?\n\n[1] ChatSpot: Bootstrapping Multimodal LLMs via Precise Referring Instruction Tuning.", "answer": "Expanding from image - level to region - level instruction tuning is a fundamental progression in multimodal large language models. The emergence of ChatGPT (LLM) is well - recognized as a new era of deep learning, just like AlexNet in 2012. At that time, how to expand AlexNet, which was used for image classification tasks, to object - level recognition tasks was a big challenge for the whole community, until the proposal of R - CNN in 2014. Although R - CNN is quite a simple design, no one would say that the transition from AlexNet to R - CNN is just a natural progression.\n\nSimilarly, the proposed GPT4RoI is motivated to expand the LLM’s visual instruction tuning from image - level to region - level and serves a fundamental role in the related area. A large portion of academic researchers pursue sophisticated designs and even agree that complexity is equal to innovation, but this is definitely wrong. Whether considering Occam’s razor or practical applications, the basic rule is: the simpler, the better. GPT4RoI incorporates region references in instruction and develops the pioneering model for general - purpose region understanding. It can be believed that it is novel and provides a fresh perspective for the region - understanding field.", "category": "Method_Disambiguation"}, {"question": "While this paper utilized more datasets for training, is the improvement in results relatively marginal, as shown in Table 5?", "answer": "The baseline 12-in-1 in Table 5 uses 12 pre-training datasets, which is more data than the paper's method. Table 5 (visual 7W) is only one benchmark for reasoning ability. However, the VCR dataset (Table 6) is more popular and challenging, and the paper's method achieves a 6-point improvement over the previous state-of-the-art on this dataset with significantly less training data. This demonstrates the effectiveness of the paper's method.", "category": "Experimental_Analysis"}, {"question": "What is the magnitude comparison of the data samples used in 12-in-1 and GPT4RoI?", "answer": "12-in-1 uses 4.4 million data samples, while GPT4RoI uses 0.43 million data samples. The paper uses only a subset of the data that is used by 12-in-1, which is ten times smaller (0.43 million versus 4.4 million).", "category": "Method_Comparison"}, {"question": "What is the rationale for choosing specific levels of visual and textual features to fuse, and is the improved performance of spatial instruction tuning due to the proposed method or the amount of training data used?", "answer": "The specific levels of visual and textual features were chosen based on their ability to capture complementary information, as detailed in Section 3.1 of the paper. The improved performance of spatial instruction tuning is attributed to the proposed method rather than the amount of training data. Statistics on the training data used for the baselines, included in the tables, confirm that the performance improvement is primarily due to the proposed method.", "category": "Motivation_Analysis"}, {"question": "Does the model exhibit a slight preference for textual features in certain scenarios?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were additional experiments conducted on datasets not used for pre-training to address concerns of overfitting?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does GPT4RoI use a dataset that is ten times smaller in terms of data samples compared to the 12-in-1 dataset?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "cijO0f8u35", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Scaling Relationship on Learning Mathematical Reasoning with Large Language Models", "authors": ["Zheng Yuan", "Hongyi Yuan", "Chengpeng Li", "Guanting Dong", "Keming Lu", "Chuanqi Tan", "Chang Zhou", "Jingren Zhou"], "abstract": "Mathematical reasoning is a challenging task for large language models (LLMs), while the scaling relationship of it with respect to LLM capacity is under-explored.\nIn this paper, we investigate how the pre-training loss, supervised data amount, and augmented data amount influence the reasoning performances of a supervised LLM.\nWe find that pre-training loss is a better indicator of the model's performance than the model's parameter count.\nWe apply supervised fine-tuning (SFT) with different amounts of supervised data and empirically find a log-linear relation between data amount and model performance, and we find better models improve less with enlarged supervised datasets.\nTo augment more data samples for improving model performances without any human effort, we propose to apply Rejection sampling Fine-Tuning (RFT).\nRFT uses supervised models to generate and collect correct reasoning paths as augmented fine-tuning datasets.\nWe find with augmented samples containing more distinct reasoning paths, RFT improves mathematical reasoning performance more for LLMs.\nWe also find RFT brings more improvement for less performant LLMs.\nFurthermore, we combine rejection samples from multiple models which push LLaMA-7B to an accuracy of 49.3\\% on GSM8K which outperforms the supervised fine-tuning (SFT) accuracy of 35.9\\% significantly.", "keywords": "Mathematical Reasoning;Scaling Relationship;Large Language Model", "qa_pairs": [{"question": "How does the paper address the limitation of evaluating only one dataset to ensure generalizability to other multi-hop reasoning problems?", "answer": "The paper with additional experiments uses a random sample of the Pile test corpus to align the pre-train losses of LLaMA, LLaMA2, and Pythia series. The authors conduct SFT and RFT experiments on the GSM8K benchmark with the Pythia series and SFT experiments on the MATH benchmark with LLaMA and LLaMA2. The detailed results are listed in the appendix H . Key findings include: (a) pre-training losses remain negatively linearly correlated to SFT performance including Pythia series models, (b) RFT significantly improves the performance of Pythia series models (Pythia-410M SFT 5.6 vs RFT 18.9, Pythia-2.8B SFT 18.8 vs RFT 34.6), and (c) LLaMA series models continue to improve performance linearly with double fine-tuning data on the MATH benchmark.", "category": "Experimental_Analysis"}, {"question": "What is the computational overhead of the proposed approach compared to pre-training, and how does it compare to using GPT4 API for data augmentation?", "answer": "The RFT inference and training FLOPs have a significantly lower computational requirement ($1\\times 10^{-4}$) compared to pre - training, making it an acceptable option. Using the GPT4 API for data augmentation neglects the computational cost of inferencing with GPT4. The paper's approach relies on enhancing models like LLaMA without the need for GPT4, unlike other math SFT papers such as WizardMATH, Meta - Math, and Mammoth, which require the GPT API for data augmentation. ", "category": "Method_Comparison"}, {"question": "Can the findings of this paper be generalized to other domains beyond mathematical reasoning tasks, and do similar scaling relationships and performance improvements hold in other large language model (LLM) applications?", "answer": "The authors have conducted experiments on human alignment by fine-tuning LLaMA-7/13/33B models on the ShareGPT dataset with varying amounts(1, 1/4, 1/16, 1/64) of data and measuring alignment performance using MT-Bench. For LLaMA - 7B models, MT - Bench scores are 5.88, 5.85, 5.61, 5.11. For LLaMA - 13B models, MT - Bench scores are 6.13, 6.03, 5.66, 5.24. For LLaMA - 33B models, MT - Bench scores are 6.63, 6.66, 6.17, 5.99. Results show that model performance generally improves with increased data amount, suggesting that similar scaling relationships and performance improvements may hold in other LLM applications.", "category": "Experimental_Analysis"}, {"question": "In the RFT experiments, were there instances where the model performance showed insignificant improvement or deterioration, particularly with larger pre-trained models, and what might be the reasons for this?", "answer": "In the RFT experiments, insignificant improvement was observed with larger pre-trained models like LLaMA-65B (+0.4) and LLaMA2-70B (+1.6). The main reason is that these larger models have already acquired substantial reasoning ability during pre-training, leaving limited space for further improvement.", "category": "Experimental_Analysis"}, {"question": "What is the practical utility of the fine-tuning scaling law and augmented fine-tuning scaling law for LLM practitioners, given that the importance of model size, pre-training loss, and data amount is already well-known?", "answer": "To improve a specific ability of LLMs, the paper proposes three directions: (1) improving the base model via pre - training; (2) obtaining more human - written fine - tuning data; (3) augmenting more model - generated fine - tuning data. The fine - tuning scaling law and augmented fine - tuning scaling law help LLM practitioners decide whether to generate human - written fine - tuning data or model - generated data. When the amount of SFT data for a task (e.g., gsm8k) is limited, practitioners should prioritize generating human - written data. When the amount of SFT data is large or the fine - tuning scaling law is similar to the augmented scaling law, practitioners should focus on generating data through LLM augmentation.", "category": "Method_Disambiguation"}, {"question": "Do the conclusions regarding the relationship between pre-training loss and performance, as well as the impact of supervised data size, hold true when tested on other math world datasets beyond GSM8K?", "answer": "The paper with additional experiments uses a random sample of the Pile test corpus to align the pre-train losses of LLaMA, LLaMA2, and Pythia series. The authors conduct SFT and RFT experiments on the GSM8K benchmark with the Pythia series and SFT experiments on the MATH benchmark with LLaMA and LLaMA2. The detailed results are listed in the appendix H . Key findings include: (a) pre-training losses remain negatively linearly correlated to SFT performance including Pythia series models, (b) RFT significantly improves the performance of Pythia series models (Pythia-410M SFT 5.6 vs RFT 18.9, Pythia-2.8B SFT 18.8 vs RFT 34.6), and (c) LLaMA series models continue to improve performance linearly with double fine-tuning data on the MATH benchmark.", "category": "Experimental_Analysis"}, {"question": "Why are the RFT (k=100) experiments not conducted for models larger than 33B in the scaling figures?", "answer": "The RFT (k=100) experiments are not conducted for models larger than 33B due to two main reasons: 1) Sampling 100 times with 65B/70B models is computationally expensive, requiring 10^19 FLOPs, and 2) LLaMA-33B SFT models have shown a strong tendency for over-fitting(or memorizing) GSM8K problems, producing fewer diverse reasoning paths compared to LLaMA-7/13B, making them unsuitable for RFT.", "category": "Motivation_Analysis"}, {"question": "Why is the pretraining loss used on the x-axis in Figure 1 instead of the fine-tuning loss, given the variability in pretraining datasets and tokenizers?", "answer": "The x - axis is based on pretraining loss because fine - tuning losses are strongly influenced by the number of epochs and can be arbitrary in model performance after the first epoch. Fine - tuning loss is a metric of model performance, while pre - training loss is a performance indicator. To ensure comparability, the authors calculated pretraining losses for LLaMA, LLaMA2, and Pythia using the same pretraining dataset (The Pile test set), and the results are shown in Appendix H, confirming that the findings remain valid.", "category": "Motivation_Analysis"}, {"question": "Did the study investigate the effect of different pretraining losses on models of the same size and training tokens?", "answer": "In Appendix H, the authors conducted experiments with Pythia - v2 models, which use the same pretraining data and optimization steps. Significant performance improvements on GSM8K with Pythia - v2 and RFT were observed. Additionally, the authors calculated pre - train losses for LLaMA, LLaMA2, and Pythia using The Pile test set and found that the paper’s findings still hold.", "category": "Experimental_Exposition"}, {"question": "Were the training losses of LLaMA, LLaMA2, and Pythia recalculated using the same pretraining dataset, The Pile test, to address comparability concerns?", "answer": "True", "category": "Claim_Verification"}, {"question": "Has the paper demonstrated similar results using models other than LLaMA by conducting additional experiments with Pythia on GSM8K and LLaMA on MATH?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the evaluation of the paper limited to only one dataset, without considering other multi-hop reasoning problems?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were experiments conducted on datasets other than GSM8K to validate the conclusions about the relationship between pre-training loss and performance?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do LLaMA series models still improve performance linearly when doubling the fine - tuning data for the MATH benchmark?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the LLaMA series models show linear improvement in performance on the MATH benchmark when doubling the fine-tuning data?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the improvement in performance with decreasing pretraining loss demonstrated using models with identical size and training tokens?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the experiments in the appendix H include testing on the GSM8K benchmark with the Pythia series models?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the study include experiments on models of the same size and pretraining data to analyze the impact of different pretraining losses on tuned and ICL performance?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the pre-train losses of LLaMA, LLaMA2, and Pythia calculated using the same pretraining dataset for comparative analysis?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the distinct reasoning paths the most important factor for augmentation dates?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "lifLHzadgr", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Cumulative Reasoning with Large Language Models", "authors": ["Yifan Zhang", "Jingqin Yang", "Yang Yuan", "Andrew C Yao"], "abstract": "While language models are powerful and versatile, they often fail to address highly complex problems. This is because solving complex problems requires deliberate thinking, which has been only minimally guided during training. In this paper, we propose a new method called Cumulative Reasoning (CR), which employs language models in a cumulative and iterative manner to emulate human thought processes. By decomposing tasks into smaller components, CR streamlines the problem-solving process, rendering it both more manageable and effective. For logical inference tasks, CR consistently outperforms existing methods with an improvement up to 9.3%, and achieves an accuracy of 98.04% on the curated FOLIO wiki dataset. In the context of the Game of 24, CR achieves an accuracy of 98%, which signifies a substantial enhancement of 24% over the previous state-of-the-art method. Finally, on the MATH dataset, we establish new state-of-the-art results with 58.0% overall accuracy, surpassing the previous best approach by a margin of 4.2%, and achieving 43% relative improvement on the hardest level 5 problems (22.4% to 32.1%).", "keywords": "cumulative reasoning;large language models;reasoning", "qa_pairs": [{"question": "How is the background on logic in Section 2 relevant for tasks that are not explicitly based on logic?", "answer": "The background on logic in Section 2 serves as a conceptual framework for the reasoning strategies employed by the CR method, even for tasks not explicitly grounded in logic, such as the Game of 24 or the MATH dataset. It helps contextualize the approach and its potential applications in a broader spectrum of reasoning tasks, including those not strictly defined in FOL, such as the AutoTNLI dataset, which can only be represented in Higher-order Logic, and the MATH dataset, which can only be represented in Categorical Logic and Topos Theory.", "category": "Method_Mechanics"}, {"question": "How does the time and compute efficiency of the proposed method compare to the baselines, particularly in terms of the number of GPT4 prompts used per state visit?", "answer": "The total number of prompts used in the proposed method (CR) is controlled to not exceed a certain multiple of the baselines, keeping the total invocations within a manageable limit comparable to other methods. As for the time aspect, it’s challenging to provide an exact estimate due to dependencies on the response rate of the OpenAI API. This response rate can vary based on user activity at the time of execution. Considerable variance in the server response speeds was observed during the experiments. However, the paper ensured that efficiency was within reasonable bounds, making CR a viable alternative to existing methods in practical scenarios.", "category": "Method_Comparison"}, {"question": "Is it accurate that CR requires significantly more computational resources (API calls) compared to CoT - SC? For instance, in the case of the Game of 24, CoT - SC employs 16 samples and conducts a majority vote among them. This necessitates 16 API calls to the model. In contrast, CR maintains ‘n = 4,’ indicating that at each step, the proposer makes four API calls, the verifier makes one call, and while the reporter makes another call to determine whether the step constitutes the final answer. Consequently, a total of six calls are made at each step. With a maximum of 50 iterations, the maximum API calls can reach as high as 300. Should CR be evaluated against ‘k = 300’ for a fair comparison?", "answer": "Yes, CR requires slightly more computational resources than other methods, such as CoT - SC. For example, CR with PHP needs 12 API calls, while complex - cot (w/ PHP) uses 7.77 API calls. However, CR explores fewer states compared to CoT/CoT - SC/ToT, meaning it spends more time on each state. Two important aspects in this context are highlighted:\n- **Algorithmic Nature and Resource Utilization**: It is difficult to adjust each algorithm so that all of them use a similar number of API calls, because each algorithm inherently varies in its approach to problem - solving. CoT, for instance, employs a straightforward linear thinking process, naturally limiting the time spent per state. On the other hand, ToT involves a breadth - first search, where the extent of exploration is adjustable. CR's slightly higher resource use is a byproduct of its design philosophy, spending more resources on each state.\n- **Exploring LLMs' Capabilities**: The current era is about pushing the boundaries of what LLMs can achieve. The focus of the paper with CR is to test the limits of LLMs in solving complex problems, potentially to a human - equivalent level. This exploration, even with a higher resource footprint, is vital for understanding the full potential of LLMs in tackling challenging tasks.\n", "category": "Method_Comparison"}, {"question": "Why was a comprehensive comparison with other prior works such as 'Faithful Reasoning Using Large Language Models' (Creswell and Shanahan), which follows a similar methodology of select, infer, and halt; 'Maieutic Prompting' (Jung et al.), where they generate and prune; 'Training Verifiers' (Cobbe et al), which involves generating multiple solutions and ranking them; 'LM vs LM' (Cohen et al), in which one large language model can identify consistencies in another, not included in the study, and how does the proposed method compare to CoT, CoT-SC, and ToT?", "answer": "A comprehensive comparison with the mentioned works was not included primarily because most of these methods are not open - sourced(except Maieutic prompting), making a fair comparison challenging.  Additionally, the experimental results from the DetermLR paper support the efficacy of the proposed method. Works like DetermLR (https://arxiv.org/pdf/2310.18659.pdf) have replicated and extended CR, showing that it consistently outperforms CoT, CoT - SC, and ToT with fewer visited states on datasets like LogiQA, ProofWriter, and LogicalDeduction. \n**Detailed Comparative Analysis**: \n - LogiQA: CR outperforms ToT in accuracy (45.25% vs. 43.02%) while maintaining a comparable number of visited states (17 vs. 19.87). \n - ProofWriter: CR surpasses ToT (71.67% vs. 70.33%) with fewer visited states (16.76 vs. 24.57). \n - FOLIO: CR matches closely with ToT in accuracy (69.1%) but is more efficient in the number of states visited (15.87 vs. 19.12). \n - Logical Deduction (LD): CR consistently outperforms ToT in accuracy (78.33% vs. 76.83%) with fewer visited states (16.76 vs. 24.57). ", "category": "Method_Comparison"}, {"question": "What is the average number of model calls required for CR with and without PhP in the MATH experiment, and how does it compare to the number of calls for complex-cot with and without PhP?", "answer": "In the MATH experiment, CR without PhP requires 4 API calls due to a simplified linear process without an explicit verifier. CR with PhP requires 12 API calls, as PhP adds an extra call per round to check for answer consistency. In comparison, complex-cot without PhP needs 2 API calls, split between intermediate reasoning and final answer generation(this is also why CR requires 4 API calls), while complex-cot with PhP averages 7.77 calls.", "category": "Experimental_Exposition"}, {"question": "How does the proposed method (CR) for the Game of 24 dataset differ from a naive brute force search, and what modifications were made to the model to achieve higher accuracy and fewer visited states compared to baselines like ToT?", "answer": "The proposed method (CR) differs from a naive brute force search by exploiting the preferences of the LLM based on the current situation, aiming to find the correct answer by visiting as few states as possible and avoiding unnecessary mistakes. This process is non-trivial and distinguishes CR from a pure brute force algorithm. The main modification to the model involves setting the current explored state as a 'premise', while keeping the rest of the model unchanged. This approach results in higher accuracy and fewer visited states compared to baselines like ToT.", "category": "Method_Comparison"}, {"question": "What are the contributions of the proposed method, specifically the self-correctness ability of the LLM, the grounding to First-order-logic (FOL), and the wider search among candidates and hypotheses, as demonstrated by the ablation study?", "answer": "The ablation study on FOLIO - wiki demonstrates that the Verifier (self - correctness ability) and the random choice of premises in the Proposer (wider search among candidates and hypotheses) are both effective. The results show that if the Verifier is removed but the random choice of premises skill in the Proposer is retained, the quality of propositions put forward by the Proposer without verification will obviously decrease significantly. When both methods are applied together, they synergistically enhance the reasoning capability of GPT - 3.5 - Turbo on FOLIO - wiki, with a combined effect greater than the sum of their individual improvements. The results are shown in the table below:\n\n| Method | CoT | CR | CR(w/o Verifier) | CR(w/o random choice of premises) | CR(w/o Verifier and random choice of premises) |\n| --- | --- | --- | --- | --- | --- |\n| **Acc.** | 64.61 | 73.03(+8.42) | 64.23(-0.38) | 68.73(+4.12) | 67.23(+2.62) |\n\n", "category": "Experimental_Exposition"}, {"question": "What does the parameter 'n' represent in the proposed approach, and how does it differ from the number of beams in beam search? Additionally, does the approach still provide performance gains over baselines like CoT and SC when 'n=1'?", "answer": "In FOLIO tasks, 'n' represents the number of verified generated propositions. When 'n=1' and equipped with verifiers, CR still exhibits gains over baselines such as CoT and SC, indicating that the benefits of CR extend beyond merely a wide beam search approach. It is not the beam search numbers, but the number of accumulative intermediate results, that is, the beam search's results are disjointed, but the paper's thinking context is consecutive, which makes the reasoning process of CR different from CoT-SC and ToT. In AutoTNLI, where the paper doesn't implement verifiers, 'n' refers to the number of accumulated generated propositions, and when n grows, the accuracy consistently improves. For more ablations on logical inference datasets, please see Figure 3 (The impact of the number of generated determinate premises) in the paper DetermLR [1], which also shows consistent improvement with more accumulated propositions.\n\n[1]https://arxiv.org/pdf/2310.18659.pdf", "category": "Method_Comparison"}, {"question": "Does the CR method require more computational resources in terms of API calls compared to CoT-SC, as evidenced by the comparison in the Game of 24 example?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the primary focus of using CR to explore the limits of LLMs in solving complex problems, even if it results in higher resource usage?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper include an accuracy versus cost tradeoff graph to evaluate the computational expense of CR versus other methods?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is it possible to adjust each algorithm to use a similar number of API calls due to their inherent differences in problem-solving approaches?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "0A5o6dCKeK", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "NExT-GPT: Any-to-Any Multimodal LLM", "authors": ["Shengqiong Wu", "Hao Fei", "Leigang Qu", "Wei Ji", "Tat-Seng Chua"], "abstract": "While recently Multimodal Large Language Models (MM-LLMs) have made exciting strides, they mostly fall prey to the limitation of only input-side multimodal understanding, without the ability to produce content in multiple modalities. As we humans always perceive the world and communicate with people through various modalities, developing any-to-any MM-LLMs capable of accepting and delivering content in any modality becomes essential to human-level AI. To fill the gap, we present an end-to-end general-purpose any-to-any MM-LLM system, NExT-GPT. We connect an LLM with multimodal adaptors and different diffusion decoders, enabling NExT-GPT to perceive inputs and generate outputs in arbitrary combinations of text, images, videos, and audio. By leveraging the existing well-trained highly-performing encoders and decoders, NExT-GPT is tuned with only a small amount of parameter (1%) of certain projection layers, which not only benefits low-cost training but also facilitates convenient expansion to more potential modalities. Moreover, we introduce a modality-switching instruction tuning (MosIT) and manually curate a high-quality dataset for MosIT, based on which NExT-GPT is empowered with complex cross-modal semantic understanding and content generation. Overall, our research showcases the promising possibility of building a unified AI agent capable of modeling universal modalities, paving the way for more human-like AI research in the community.", "keywords": "Large Language Model;Diffusion Model", "qa_pairs": [{"question": "How does the proposed alignment learning technique address the challenges of balancing performance across different modalities in an any-to-any modality paradigm?", "answer": "(1) The alignment learning techniques proposed in the paper are lightweight and cost - effective, which facilitates rapid iteration of computational - centric model training.\n(2) The authors have indeed considered the challenges introduced by the ‘any - to - any modality’ paradigm to some extent, including the performance balance issue. For example, utilizing a unified encoder (i.e., ImageBind) to perceive non - language modalities, instead of separate encoders, helps mitigate imbalances between models. Additionally, the paper has designed three independent output projection layers to align the LLM with downstream generation systems, alleviating the interference between modalities and maintaining a balance.\n(3) In contrast to balancing the performance across different modalities, it is emphasized that maximizing the performance of individual existing encoders and decoders used in the paper’s system can be given higher priority, achieving fully optimal performance of each modality.", "category": "Method_Mechanics"}, {"question": "Does the observed worse performance on benchmarking datasets when introducing additional modalities suggest that the proposed alignment technique is ineffective in enhancing model performance?", "answer": "The paper does not claim that introducing other modalities enhances the performance of specific modalities. The focus is on constructing a unified agent (NExT - GPT) to mimic the complementary nature of different modalities in real - world scenarios. The proposed alignment tuning is designed to project different modalities into a shared semantic space, achieving such a unified agent where different modalities are complementary, rather than enhancing individual modality performance. The performance mutual enhancement of multimodal systems is mostly observed and expected in tasks involving multimodal inputs where different modalities refer to shared semantic meaning, which however is not the primary focus of the paper’s system. In addition, regarding the concern about ‘leads to worse performance on benchmarking datasets’, experimental results already indicate that the paper’s system outperforms the baselines in tasks ranging from text - to - video to audio - to - text, image - to - text, and video - to - text generation. For the remaining evaluation tasks, the paper’s model still secured performances on par with the best - performing baselines, as shown in the tables. Most SoTA baseline systems are deliberately created for addressing those particular benchmarks and tasks which are fine - tuned on the datasets to achieve a very high score, while our systems focus more on the overall general - purpose any - to - any capabilities. The paper doesn’t do much task - specific optimization and fine - tuning of its system as the baselines do, which actually leaves ample room for further huge improvement of benchmarks.", "category": "Experimental_Analysis"}, {"question": "What measures are implemented in Next-GPT to address ethical concerns related to the content generated by the model and the dataset content?", "answer": "Next - GPT addresses ethical concerns through several measures. For the model, it leverages safety mechanisms from existing well - established models to ensure content quality and emphasizes user awareness by advising adherence to usage guidelines (e.g., interpreting the generated content with caution) and by limiting applications to academic research. For the dataset, meticulous privacy protection measures are taken, such as removing or obfuscating personally identifiable information. Bias mitigation is prioritized during dataset collection to ensure representativeness and to avoid favoring or disadvantaging any specific group or perspective. It is important to note that the paper possesses the necessary licenses to collect data. For data distribution to researchers and individuals, adherence to the stipulated license is mandatory.", "category": "Method_Mechanics"}, {"question": "Why was the NExT-GPT model evaluated on individual datasets instead of a dedicated benchmark for the any-to-any task format, and how does this approach ensure a fair assessment of the model's effectiveness?", "answer": "The paper’s system introduces a novel any - to - any task format, for which currently no dedicated benchmark is designed for evaluation. Under such a circumstance, when expecting to quantify the effectiveness of the paper’s model, the only viable approach is to conduct evaluations on existing benchmark datasets of traditional individual cross - model or multimodal tasks (e.g., text - to - image, video - to - text), comparing with baselines. It is worth noting that the paper strictly adheres to the experimental settings used in the baselines of each dataset to ensure a fair comparison, including data splitting and fine - tuning/zero - shot setups. While lacking a specific any - to - any multimodal benchmark (evaluation set), this can be achieved by utilizing the MosIT dataset proposed in this paper, e.g., splitting it into training and testing sets, to assess the performance of any - to - any models. The paper’s model achieved the best performance on the majority of tasks compared to baselines, including Audio - to - text generation, Image - to - text generation, Video - to - text generation, and Text - conditioned audio editing, and also secured performances comparable to SoTA baselines in other tasks such as Text - to - image generation, Text - to - audio generation, and Text - conditioned image/video editing. Notably, even in a zero - shot scenario, particularly in the Text - to - video generation task, the paper’s model surpasses the SoTA baseline. It is very important to emphasize a factual point that, as the paper’s goal is to build a general - purpose MM - LLM, the paper runs on the downstream in - domain datasets to exhibit that the paper’s system indeed can perform well on existing tasks. Thus the paper doesn’t do much task - specific optimization of the paper’s system the way the baselines do for these in - domain datasets, which actually leaves ample room for further huge improvement to those benchmarks.", "category": "Motivation_Analysis"}, {"question": "What would be the expected performance impact if the model was trained on a mixture of the proposed MosIT dataset and benchmark datasets, and then evaluated on the benchmarks?", "answer": "NExT - GPT roughly considers four stages of the training/learning process:\n\nStage - 1: Encoding - size Alignment Learning. The input projection layer is trained by adopting an ‘X - to - Text’ generation task on Webvid - 2M, CC3M, and AudioCaps dataset.\n\nStage - 2: Decoding - side Alignment Learning. The three output projection layers are optimized on Webvid - 2M, CC3M, and AudioCaps.\n\nStage - 3: End - to - end Instruction - Tuning. The paper optimizes the whole NExT - GPT using instruction - following datasets, including Text+X → Text Dataset, Text → Text+X Dataset, and MosIT dataset. This step is to enhance the ability of LLM to understand user input instructions.\n\nStage - 4: In - domain Fine - tuning on Task - specific Training Dataset. NExT - GPT is fine - tuned on the specific in - domain dataset of a specific task, after which the system will be evaluated on the testing set.\n\nIt is worth noting that Stage - 3, involving instructive tuning with MosIT data, is designed to enhance the model’s proficiency in performing complex, multi - turn, modality - switching interactions. However, in Stage - 4, when evaluating on traditional benchmarks, modality alignment without intricate instruction understanding is essentially required. Without instructive tuning with MosIT, the expected impact on performance during traditional benchmark evaluation could be limited or minimal.\n\nWith the above discussion, for the results when the model is trained on a mixture of the proposed MosIT dataset and benchmark datasets, as the objectives of the benchmark datasets and MosIT dataset are quite distinct, mixing them probably introduces the risk of interference of feature patterns, where the information learned from in - domain task - specific dataset might be disrupted by instructive tuning data. Hence, theoretically and practically, it is likely that such a mixture of training data may lead to a degradation in performance.", "category": "Experimental_Exposition"}, {"question": "How does the proposed multimodal language model perform on recent benchmarks such as MME, MMBench, and SEEDBench, particularly in terms of text generation?", "answer": "The proposed model has been evaluated on MME, MMBench, and SEEDBench, with results presented in Table 12, and 13. The model mostly achieves better text generation performance than the baseline MM-LLMs. However, it is important to note that these benchmarks are limited to text generation, whereas the proposed model extends its capabilities to any-to-any generation, including image, video, and audio. The model's broader generation capabilities set it apart from the benchmarks mentioned. ", "category": "Experimental_Exposition"}, {"question": "How does NExT-GPT compare with other MLLMs, such as InstructBLIP, LLaVA, mPLUG-Owl for text generation, and DreamLLM and EMU for conditioned image generation, in terms of performance on NoCaps, Flicker 30K, and COCO datasets?", "answer": "NExT-GPT was compared with other MLLMs on NoCaps, Flicker 30K, and COCO datasets under a fair zero-shot setting, and its conditioned image generation results were compared with DreamLLM and EMU. The CIDEr score was used for text generation, and FID was used for image generation. The results show that NExT-GPT achieves competitive or better performance in text-to-image and image-to-text generation tasks compared to other MLLMs. Additionally, NExT-GPT's broader capability in any-to-any generation, including video and audio, distinguishes it from the benchmarks mentioned. \n\n### Table 1: Zero-shot evaluation of image-to-text generation with CIDEr score (the higher the better). Results marked with * are copied from the original paper.\n\n| Model | NoCaps | Flickr 30K | COCO |\n| ---- | ---- | ---- | ---- |\n| InstructBLIP(7B) | 123.1* | 82.4* | 102.2* |\n| LLaVA(7B) | 120.7 | 82.7 | - |\n| mPLUG-Owl(7B) | 117.0 | 80.3 | 119.3 |\n| EMU(13B) | - | - | 117.7* |\n| DreamLLM(7B) | - | - | 115.4* |\n| NExT-GPT(7B) | 123.6 | 84.5 | 124.9 |\n\n### Table 2: Zero-shot evaluation of Text-to-Image Generation (COCO) with FID (the lower the better) score. Models pre-trained with LION (2B) data are denoted with *.\n\n| Model | FID |\n| ---- | ---- |\n| EMU*(13B) | 11.66 |\n| DreamLLM*(7B) | 8.46 |\n| NExT-GPT(7B) | 13.85 |\n| NExT-GPT*(7B) | 8.62 |\n", "category": "Method_Comparison"}, {"question": "Why does the paper propose aligning semantic tokens with outputs from text encoders of diffusion models instead of directly using the diffusion model's text encoder to encode textual captions, or having the LLM output captions in a fixed format for diffusion model generation?", "answer": "Using LLM to output textual captions and feed them to the follow - up diffusion models for generation is one type of prior existing methods for achieving the goal of (nearly) any - to - any MM - LLM systems. In these methods, LLMs are straightforwardly combined with external tools by connecting the LLMs with other modules using explicit textual instructions (e.g., captions or descriptions of content to generate) as the messenger. They essentially have a pipeline architecture. Representative models include Visual - ChatGPT and HuggingGPT.\n\nHowever, there are at least two limitations in such an approach. Firstly, the information transfer between different modules is entirely based on discrete texts produced by the LLM. In this cascade process, noise is inevitably introduced and errors are propagated. This is a problem rooted in the core of different modalities; that is, there is an inherent gap between language and other modalities that cannot be eliminated. Utilizing textual captions as intermediate representations runs the risk of overlooking modality - specific features when expressing non - linguistic modalities solely through language. In other words, relying on textual captions as an intermediate representation may create a bottleneck, especially when handling complex or implicit input instructions. For example, when prompting the pipeline system to generate a picture of a dog, the LLM might produce a simple caption like ‘a dog running in a park’. However, the authors expect the model to generate images with much more detailed and intricate scenes, not just including the ‘dog’ and a ‘park’, but also vivid background details like ‘a playground with green grass, with many people walking by’, and the dog itself ‘holding a frisbee in its mouth’. All these can highlight the potential limitations of simply using textual captions as an intermediate step. Secondly, the entire system only leverages existing pre - trained tools for inference, limiting its capabilities of content understanding and multimodal generation, especially in interpreting intricate and implicit user instructions. By using the continual representations of special signal tokens as the messenger between different modules, the paper proposes an end - to - end unified MM - LLM system in the true sense. The full system can be updated throughout, which thus helps it have a much stronger capability in better understanding complex user instructions and generating content more accurately and with high quality.", "category": "Motivation_Analysis"}, {"question": "What visualizations have been in the paper to illustrate the differences between the pipeline method and the proposed NExT-GPT, especially for complex instructions, and how do these examples demonstrate the challenges faced by Stable Diffusion with pure captions versus LLM token embeddings?", "answer": "Appendix H. Case Study on Pipeline - style vs. End - to - end Unification has several examples of comparisons between these systems as in Figure 6, 7 and 8. Figure 6 presents the case of image generation from a simple input user instruction, while Figure 7 and Figure 8 present two cases of image generation from comparatively complex input user instructions. For the simple instruction case, all generated image content from both pipeline - style and end - to - end (the paper’s) systems seems correct and coincides with the input prompt. However, when handling the complex instructions, as seen in Figure 7 and Figure 8, the generated image content can be wrong and biased against the user intention. The problems are rooted in the core of different modalities, i.e., there are inherent gaps between language and visual modalities that cannot be eliminated. Here are two representative attributes: the numeration of vision (cf. Figure 6) and the visual - spatial relational semantics (cf. Figure 7), which could be hard to (or even cannot) be expressed by the intermediate captions conveniently. Utilizing textual captions as intermediate representations runs the risk of overlooking these modality - specific features when expressing non - linguistic (e.g., visual) modalities solely through language. By the way, with the intermediate captions produced from the pipeline - style systems in Figure 7 and 8, the Stable Diffusion model just has difficulty in accurately understanding the vision numeration and visual - spatial relation and generating correct answers. That is, these are the problems inherent in the Stable Diffusion model itself, and it is tricky for Stable Diffusion alone to overcome these problems. Most recent work tries to solve this issue by integrating the vision - specific features into the Stable Diffusion [1,2] via additional feature engineering. But, in the paper’s NExT - GPT with an end - to - end solution, the implicit modality signal token embeddings that carry rich modality - specific features of non - linguistic modalities will be naturally encoded and passed to the downstream modules (e.g., Stable Diffusion), without any further external effort.\n\n[1] LayoutLLM-T2I: Eliciting Layout Guidance from LLM for Text-to-Image Generation\n\n[2] LayoutGPT: Compositional Visual Planning and Generation with Large Language Models", "category": "Experimental_Analysis"}, {"question": "Does the model rely on multiple pre-trained models for encoding different modalities, and how does the quality of pre-training impact its performance?", "answer": "The model does not rely on multiple pre-trained models for encoding different modalities. Instead, it uses a unified solution, ImageBind, which encodes and projects different types of information. ImageBind is a state-of-the-art model capable of understanding six modalities, eliminating the need for managing multiple heterogeneous modal encoders. The model architecture also allows for easy and cost-effective extension to powerful pre-trained models, enhancing the understanding capabilities of different modalities.", "category": "Method_Disambiguation"}, {"question": "What are the sizes of tunable parameters for each modality in the model?", "answer": "The parameter size for the input alignment module is 4M. For the output alignment modules, the sizes are 31M for image, 31M for audio, and 32M for video modalities.\n\n|Modality | Input Projection Param.|Output Projection Param. |Total |\n| :-----: | :----: |  :-----: | :----: | \n|Text | - | - | - |\n| Image | 4M | 31M | 35M|\n| Audio| 4M | 31M | 35M|\n| Video| 4M | 32M | 36M|", "category": "Experimental_Setup"}, {"question": "Why does the proposed NExT-GPT system not outperform all baseline models on all benchmarks, and how does this relate to the efficacy of the alignment learning strategies?", "answer": "The NExT - GPT system outperforms the best - performing baselines on the majority of tasks and performs on par with state - of - the - art (SoTA) baselines on the remaining tasks. The SoTA baselines are specifically fine - tuned for particular benchmarks and tasks, whereas NExT - GPT focuses on general - purpose any - to - any capabilities without doing much task - specific optimizations and fine - tuning of the paper’s system as the baselines do. This design choice leaves room for further improvements on benchmarks to achieve SoTA performance.", "category": "Method_Comparison"}, {"question": "Is the application of Next-GPT strictly limited to academic research and not for commercial purposes?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does Next-GPT utilize safety mechanisms from existing well-established models to ensure the quality of generated content?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "XD0PHQ5ry4", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "SELF: Language-Driven Self-Evolution for Large Language Model", "authors": ["Jianqiao Lu", "Wanjun Zhong", "Wenyong Huang", "Yufei Wang", "Fei Mi", "Baojun Wang", "Weichao Wang", "Lifeng Shang", "Qun Liu"], "abstract": "Large Language Models (LLMs) have showcased remarkable versatility across diverse domains. However, the pathway toward autonomous model development, a cornerstone for achieving human-level learning and advancing autonomous AI, remains largely uncharted. Drawing inspiration from the human capability for self-driven learning, characterized by introspection and continuous refinement, we introduce an innovative approach, termed ``SELF\" (Self-Evolution with Language Feedback). This methodology empowers LLMs to undergo continual self-evolution, thereby augmenting their inherent capabilities. Furthermore, SELF employs language-based feedback as a versatile and comprehensive evaluative tool, pinpointing areas for response refinement and bolstering the stability of self-evolutionary training. Through this approach, we aim to illuminate the prospects of autonomous AI advancement, drawing parallels with the human aptitude for learning and adaptation. \nInitiating with meta-skill learning, SELF acquires foundational meta-skills with a focus on self-feedback and self-refinement. These meta-skills are critical, guiding the model's subsequent self-evolution through a cycle of perpetual training with self-curated data, thereby enhancing its intrinsic abilities. Given unlabeled instructions, SELF equips the model with the capability to autonomously generate and interactively refine responses. This synthesized training data is subsequently filtered and utilized for iterative fine-tuning, enhancing the model's capabilities. Experimental results on representative benchmarks substantiate that SELF can progressively advance its inherent abilities without the requirement of human intervention, thereby indicating a viable pathway for autonomous model evolution. \nAdditionally, SELF can employ online self-refinement strategy to produce responses of superior quality.\nIn essence, the SELF framework signifies a progressive step towards autonomous LLM development, transforming the LLM from a mere passive recipient of information into an active participant in its own evolution.", "keywords": "Large Language Model;Alignment;Self-Evolution;Self-Refinement;Reasoning", "qa_pairs": [{"question": "What criteria and methodology are used to determine the number of self-evolution rounds and to decide when to end the self-evolution process?", "answer": "Table 1 in the paper illustrates that iteratively enhancing model capabilities with increased self - evolution training corpus progressively improves performance. However, it is presumed that this improvement may exhibit diminishing returns over time. To determine the optimal point for halting the self - evolution, the authors employ a validation set to monitor performance variance. When the model performance converges, indicating minimal gains from additional training, it is appropriate to stop the self - evolution process. Additionally, the upper bound of self - evolution effectiveness is influenced by several factors. Additional experiments indicate that stronger baseline models lead to more significant improvements in self - evolution performance, thereby raising their learning upper limits. The result is as follows:\n\n| Model                              | Initial GSM8k Result | Meta-Skill Learning(direction generation ->self-refinement) | Self-Evolution(direction generation ->self-refinement)|\n|------------------------------------|----------------------|---------------------|----------------|\n| VicunaV1.5 (finetuned from Llama2) | 18.5                 | 27.19 -> 29.27      | 32.63 -> 34.20 |\n| Vicuna(shown in paper)             | 16.43                | 25.39 -> 28.28      | 29.64 -> 31.31 |\n\nThis indicates that starting with a more robust model can enhance the efficacy of the self-evolution process.", "category": "Method_Mechanics"}, {"question": "Why were three rounds of self-evolution chosen, and what results are expected for rounds 4, 5, and 6?", "answer": "Three rounds of self-evolution were chosen to balance computational cost and performance improvement, as experiments showed diminishing returns after three rounds. Due to resource constraints, rounds 4, 5, and 6 were not conducted, but based on the observed trend, additional rounds are hypothesized to yield limited improvements. This aligns with the validation-based stopping criterion, where self-evolution is halted upon performance convergence.", "category": "Experimental_Exposition"}, {"question": "Does the SELF framework significantly increase training and inference time, and how does it compare to traditional supervised fine-tuning methods?", "answer": "The SELF framework consists of two primary steps:\n\n1) Creating self-evolution training data. In contrast to regular supervised fine-tuning, which often requires extensive manual data annotation—a laborious and time-consuming process—SELF automates the generation of its training data. This automation significantly reduces the effort and time typically associated with data annotation.\n\n2) Fine-tuning the model with this newly generated data. It's important to note that the actual training phase in SELF is similar to the Supervised Fine-Tuning (SFT) process. Therefore, SELF does not introduce additional computational costs in the training phase.\n\nWhen considering the overall process, SELF effectively reduces both the time and monetary costs associated with data collection, making it a more efficient alternative to traditional supervised training methods.", "category": "Method_Comparison"}, {"question": "What experiments were conducted to evaluate the robustness of SELF to noise in self-generated data and the effectiveness of meta-skills in filtering bad data?", "answer": "The authors conduct the following experiments to illustrate the robustness of SELF to noise in self-generated data and the effectiveness of meta-skills in filtering:\n\n| Model | Accuracy of Training Data(%) | Accuracy on Test set(%) |\n|--------------|----------|----------|\n| Unfiltered    | 29.89      | 26.90 |\n| Meta-Skill Filtered         | **44.10**                           |  **27.67** |\n\nThe second column means the accuracy of self-generated training data. The third column means the test performance of the model finetuned on such data.\n\nExperimental analysis:\n\n(1) Unfiltered: This strategy utilizes self-refinement to refine the self-generated data without filtering.  \n\n(2) Meta - Skills Filtered: After applying self - refinement, self - feedback is utilized to filter the data. There is a significant improvement in the accuracy of training data (44.10%). The performance of the fine - tuned model also improves (27.67%). It is observed that the performance improvement is not as significant as the accuracy improvement of self - generated data. The hypothesis is that the size of the self - generated data shrinks after applying filtering (from 4k to 1.8k).\n\nBy solely using self-refinement, the performance of the finetuned models still improves. This shows the robustness of SELF to noise in self-generated data. The accuracy of the self-refinement data significantly improves with the help of self-feedback meta-skill and results in better fine-tuned model performance. This shows that self-feedback has a strong ability to filter bad data.", "category": "Experimental_Exposition"}, {"question": "Would combining all data into a single round yield performance comparable to multiple rounds?", "answer": "\n| Training Method                 | Direct Generation (\\%) | Self-Refinement (%) | \n|---------------------------------|------------------------|-----------------|\n| SELF (Single Round)                     | 28.40                 | 30.55          | \n|  SELF (Iterative)         | 29.64                 | 31.31         | \n\n\nThe iterative training method shows higher performance scores (29.64% for direct generation and 31.31% for self-refinement) compared to a single round (28.40% for direct generation and 30.55% for self-refinement). The iterative approach benefits from improved LLMs in later rounds, producing higher-quality training data and, consequently, enhanced test performance.", "category": "Experimental_Exposition"}, {"question": "What is the relationship between \\(D_{QA}\\) and \\(D_{meta}\\)?\n", "answer": "(1) The QA data in Table 2, denoted as $D_{QA}$, consists of pairs of prompts and pseudo answers provided by GPT4. It does not include ground-truth labels.\n\n(2) The paper combines both $D_{QA}$ and $D_{meta}$. Given that the data structure in $D_{meta}$ diverges from conventional direct question answering formats, the paper also employs a dataset composed of pairs of questions and refined answers, denoted as $D_{QA}$, during meta-skill training.", "category": "Concept_Understanding"}, {"question": "How does the inclusion of the meta-skill learning corpus ($D_{meta}$) affect reasoning performance with and without the use of self-refinement during inference, as shown in the second row of Table 5?", "answer": "In the caption of Table 5, 'The right arrow indicates the performance improvement by Self-Refinement' indicates that the table displays results both **without** and **with** the application of self-refinement. The second row shows the result '25.39 → 28.28'. Here, the 25.39% reflects the **direct generation** performance, which is an improvement without self-refinement. The subsequent increase (+2.89%) to 28.28% represents the additional gain brought by self-refinement. It is evident that the inclusion of the meta-skill learning corpus ($D_{meta}$) leads to enhancements in both direct generation and self-refinement outcomes. Providing language feedback to the mistake of the model can improve the performance. ", "category": "Experimental_Analysis"}, {"question": "How does the performance of Vicuna fine-tuned on the GSM8K training set compare to the SELF framework, and does it surpass SELF?", "answer": "In addressing the comparison between the SELF framework and supervised fine - tuning using the GSM8K training set, the following results are presented:\n\n| Training Method                 | Direct Generation(\\%) | Self-Refinement(%) | \n|---------------------------------|------------------------|-----------------|\n| Vicuna + $D_{QA}$(7.5k)      |  28.05      | - |\n| Vicuna + meta-skill learning (7.5k)      | 31.23         | 32.98 |\n| Vicuna + meta-skill learning (7.5k)  + first round  Self-Evolution  |   35.43                | 36.22          | \n| Vicuna + meta-skill learning (7.5k)  + first and second round  Self-Evolution  |   **37.87**               | **38.12**         | \n|||||\n| Vicuna  + GSM8K training data (human-annotated, 7.5k) | 35.70 | - | \n||\n\n**Supervised Fine-Tuning vs SELF**:  The Vicuna model, when fine-tuned on the GSM8K 7.5k training set achieves an accuracy of 35.70%, which is lower than of SELF  (37.87%).\n\nIt is important to note that the 29.64% in the paper originated from a 3.5k meta-skill learning corpus. To ensure a fair comparison with the Supervised Fine-tuned model, new experiments were conducted using an expanded 7.5k meta-skill data as shown above.\n\nSpecifically, using 7.5k unlabeled training prompts to construct the meta-skill learning corpus, The baseline model Vicuna + $D_{QA}$ achieves 28.05%. After meta-skill learning, the initial result is 31.23%, which improves to 32.98% after self-refinement. The performance further increases in subsequent self-evolution rounds, with the second round reaching 37.87% to 38.12%, surpassing the supervised fine-tuning result(35.7%).", "category": "Method_Comparison"}, {"question": "Is self-refinement a necessary component for SELF, and what evidence supports this claim?", "answer": "Performance degradation occurs only in models lacking meta-skill learning in Table 1 of the paper (Vicuna, and Vicuna with $D_{QA}$). In contrast, models equipped with meta-skill learning consistently exhibit performance improvement via self-refinement. Table 1 demonstrates that SELF, when applied with self-refinement, achieves an accuracy of 31.31%, surpassing the 29.87% obtained using self-consistency. Furthermore, while self-consistency is limited to problems with unique answers, self-refinement can be applied to a more general domain as evidenced in Section 4.2.3(General Test).\n\nThe authors also conducted an ablation study to compare the effect of self-refinement and self-consistency on self-evolution training data quality and their subsequent impact on test performance:\n\n| Model | Acc. of Training Data(\\%) | Acc. on Test set(\\%) | \n|--------------|----------|----------|\n|  Self-Consistency Filtered (5x Majority)   | 28.27            | 26.77 | \n| Self-Refinement Revised     | 29.89                 | 26.90 | \n| Meta-Skills Filtered         | **44.10**                |  **27.67** |\n\nIn the above table, 'Acc. of Training Data' refers to the accuracy of self-generated training data post-filtering/refinement, while 'Acc. on Test Set' indicates the model's test performance after fine-tuning such training data.\n\nAs shown in the table, self-refinement produces higher quality training data compared with self-consistency, and results in better fine-tuned model performance. The final row demonstrates that further filtering the self-refinement data with self-feedback can improve both training data accuracy and test performance.\n\nThe study clearly shows that self-refinement is a **necessary component** of SELF. Coupled with self-feedback, self-refinement establishes a robust foundation for the self-evolution training process. Moreover, self-refinement can improve the model's performance during inference.", "category": "Experimental_Analysis"}, {"question": "How does the self-consistency approach of Wang et al. (2022a)[1] work, and how is it related to the proposed self-refinement method?\n\n[1]Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022a.", "answer": "The self-consistency approach of Wang et al. (2022a)[1] works by sampling a diverse set of reasoning paths and selecting the most consistent answer (the paper sampled 5 times in its experiments). Self-refinement enables models to evaluate and improve their outputs. Self-consistency operates without additional fine-tuning, while self-refinement applies to broader domains, not just those with unique correct answers, as demonstrated in Section 4.2.3 (General Test). Self-refinement can complement self-consistency, and integrating these two strategies leads to better performance.", "category": "Method_Comparison"}, {"question": "What are the key differences between the SELF method and reinforcement learning (RL) methods without human feedback, and how do their experimental results compare on the same datasets?", "answer": "SELF differs from RL methods by using informative natural language feedback instead of information-sparse scalar rewards. An RL-based experiment was conducted using the same SFT model (Vicuna + $D_{QA}$) for both SELF and RLHF, with the reward model trained on pair-wise comparison data(refined response is assumed better than original response) derived from the meta-skill learning corpus. The comparison data was provided by GPT-4 instead of humans, which can be seen as a form of reinforcement learning without human feedback. Results show that RL achieved a 25.55% accuracy on GSM8K, lower than SELF's 27.67%. The RL reward model often failed to identify correct responses leading to limited performance improvements, with 76% of incorrect answers receiving higher scalar rewards than correct answers on the GSM8K test set. In contrast, SELF’s natural language feedback provided more accurate evaluations, with only 28% of incorrect answers being judged as correct.\n\n| Method | GSM8K_test(%) |\n| ---- | ---- |\n| SFT | 24.49 |\n| RL | 25.55 |\n| SELF | 27.67 |\n", "category": "Method_Comparison"}, {"question": "What evidence supports the claim that large language models (LLMs) are good at self-feedback, and how does meta-skill learning influence their self-feedback accuracy?", "answer": "The claim that LLMs are capable of self-feedback is supported by studies such as [1], [2], and [3], which suggest that LLMs can engage in meaningful self-improvement. Meta-skill learning significantly improves self-feedback accuracy, increasing it to 70% compared to 33% without meta-skill learning, as models without it tend to self-feedback all responses as correct. Self-feedback ability may depend on factors like the model's intrinsic ability, prompt design, and problem domain.\n\n[1] William Saunders, Catherine Yeh, Jeff Wu, Steven Bills, Long Ouyang, Jonathan Ward, and Jan Leike. Self-critiquing models for assisting human evaluators. arXiv preprint arXiv:2206.05802, 2022.\n\n[2] Noah Shinn, Federico Cassano, Beck Labash, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning, 2023.\n\n[3] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. arXiv preprint arXiv:2303.17651, 2023.", "category": "Method_Disambiguation"}, {"question": "What does Figure 3 represent, including the meaning of the color coding and the methodology used to obtain the results?", "answer": "Figure 3 shows the comparison test result in the general domain. As noted in Section 4.2.3, the paper follows the evaluation procedures outlined in [Xu et al., 2023][1], which address the order bias issues identified in the evaluation methods proposed by [Chiang et al., 2023][2]. Specifically, different models (Vicuna, Vicuna fine - tuned on $D_{QA}$, apply SELF to Vicuna fine - tuned on $D_{QA}$, referred to as Vicuna, Vicuna + $D_{QA}$, and Vicuna + $D_{QA}$ + SELF) are tested on the Vicuna test set and EvolInstruct test set. Then, GPT - 4 is employed to assign an assessment score for each test response of different models. \nIn Figure 3, the color coding is as follows: Blue indicates test cases where the model under evaluation performs better than the baseline model (Vicuna), as judged by GPT - 4. Yellow indicates test cases of equal performance. Pink indicates test cases where the model under evaluation performs worse than the baseline model. \nFrom this figure, several key findings emerge: (1) After being fine - tuned on $D_{QA}$, the model demonstrates improved performance compared to the original Vicuna model, as evaluated by GPT - 4. (2) The application of the SELF framework to Vicuna + $D_{QA}$ enhances performance on both test sets. (3) Further improvements are observed when Vicuna + $D_{QA}$ + SELF applies self - refinement during inference. These results demonstrate that SELF is effective in the general domain.\n\n[1]Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. Wizardlm: Empowering large language models to follow complex instructions. arXiv preprint arXiv:2304.12244, 2023.\n\n[2]Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. URL https: //lmsys.org/blog/2023-03-30-vicuna/.\n", "category": "Concept_Understanding"}, {"question": "What is the difference between self-curated data and self-filtered data?", "answer": "Self-curated data refers to data after applying self-refinement to self-generated responses. Self-filtered data undergoes an additional self-feedback step to the self-refinement response, filtering out data judged as incorrect by self-feedback. ", "category": "Concept_Understanding"}, {"question": "Does the QA data used during the meta-skill learning stage include ground-truth labels?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "ttMwEuEPeB", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "3D-GPT: Procedural 3D Modeling with Large Language Models", "authors": ["Chunyi Sun", "Junlin Han", "Weijian Deng", "Xinlong Wang", "Zishan Qin", "Stephen Gould"], "abstract": "In the pursuit of efficient automated content creation, procedural generation, leveraging modifiable parameters and rule-based systems, emerges as a promising approach. Nonetheless, it could be a demanding endeavor, given its intricate nature necessitating a deep understanding of rules, algorithms, and parameters. To reduce workload, we introduce 3D-GPT, a framework utilizing large language models(LLMs) for instruction-driven 3D modeling. 3D-GPT positions LLMs as proficient problem solvers, dissecting the procedural 3D modeling tasks into accessible segments and appointing the apt agent for each task. 3D-GPT integrates three core agents: the task dispatch agent, the conceptualization agent, and the modeling agent. They collaboratively achieve two objectives. First, it enhances concise initial scene descriptions, evolving them into detailed forms while dynamically adapting the text based on subsequent instructions. Second, it integrates procedural generation, extracting parameter values from enriched text to effortlessly interface with 3D software for asset creation. Our empirical investigations confirm that 3D-GPT not only interprets and executes instructions, delivering reliable results but also collaborates effectively with human designers. Furthermore, it seamlessly integrates with Blender, unlocking expanded manipulation possibilities. \nOur work highlights the potential of LLMs in 3D modeling, offering a basic framework for future advancements in scene generation and animation.", "keywords": "intelligent agent;large language models;3D modeling;Procedural 3D Modeling", "qa_pairs": [{"question": "What specific language model was used in the experiments, and how is the function set F defined and its size determined? Additionally, how does the size of F affect scene quality, how many examples are provided per function, and is there any relationship between these examples and the user-supplied prompt L? Finally, what impact does zero-shot or few-shot learning have on performance?", "answer": "GPT-3.5 was used for all experiments. The function set F includes all functions and subfunctions from the specified script, excluding the 'creatures' class. The size of F directly impacts scene quality. Infinigen's scene generation employs a fixed function set. Removing a function, such as 'flower_generation,' eliminates the ability to create corresponding elements (e.g., flowers) in the scene. Similarly, omitting a sub-function like 'control_flower_petal' restricts control over specific aspects (e.g., flower petal movement), impacting scene controllability and diversity. The prompt L is user-supplied. he paper provides one example per function. These examples might occasionally resemble some user inputs L, but they are not directly related. Regarding zero - shot/few - shot learning, no notable difference in performance was noted between one - shot, two - shot, and three - shot prompts, though introducing more examples tends to reduce parameter diversity.", "category": "Method_Mechanics"}, {"question": "What are the impacts of the absence of D, C, I, and E components on the method's performance, as revealed by the ablation studies?", "answer": "The absence of function documentation (D) impairs the method's ability to interpret function input parameters, reducing the CLIP Score and increasing the Failure Rate. Without code (C), the system grasps basic function and parameter knowledge but experiences a decline in CLIP Score and an increased Failure Rate. The lack of information (I) hinders the conceptualization agent's capacity to enrich text with vital details, leading to a significant drop in CLIP Score. Excluding examples (E) markedly diminishes the CLIP Score and heightens the Failure Rate.", "category": "Experimental_Exposition"}, {"question": "How do the complexity of the scene and the number of parameters influence the outcomes of the proposed method?", "answer": "The complexity of the scene influences outcomes based on the initial text input: complex objects beyond the procedural algorithm's scope limit the quality of generated content, while detailed initial text descriptions enhance the conceptualization agent's understanding, yielding results more aligned with the input. The number of parameters is critical, as removing functions or sub-functions (e.g., 'flower_generation' or 'control_flower_petal') limits modeling abilities, impacting controllability and diversity. The significance of specific functions varies; for instance, removing 'control_flower_petal()' may not significantly affect the CLIP score, but eliminating 'terrain_generation()' substantially reduces it,  as terrain modeling becomes impossible.", "category": "Method_Mechanics"}, {"question": "Can the described workflow be applied to any arbitrary procedural models, such as houses, cars, or airplanes, given the same amount of information (D, C, I, and E), rather than being limited to scenes in InfiniGen?", "answer": "Yes, the system functions in a zero-shot manner, with each function processed independently by the conceptualization and modeling agents. As long as the necessary information (code C, example E, and function document D) is supplied, the method can adapt to and understand new functions, as demonstrated in the paper on page 5.", "category": "Experimental_Analysis"}, {"question": "Do the selected functions establish dependencies that allow for the correct generation of 'flowers on the trees', ensuring that trees are created first and flower positions are based on tree branch positions?", "answer": "Yes, the selected functions, such as generate_trees(), autonomously design tree branches, calculate surfaces, and accurately place leaves and flowers, ensuring correct dependencies and sparing LLMs from determining precise 3D leaf/flower locations.", "category": "Method_Disambiguation"}, {"question": "How does the efficiency of the sequential editing task in Figure 4 compare for professional Blender users (e.g., game developers) in achieving a desired outcome, relative to manually tweaking parameters?", "answer": "For Blender experts in the scenario depicted in Figure 4, transforming flat terrain into mountainous landscapes would require manually rebuilding terrain and repositioning each element, which is time-intensive. For those familiar with Infinigen libraries, the task might be straightforward. However, for each new procedural library, the experts are required to understand every new function, which also poses a significant challenge for professional Blender users to quickly grasp and utilize every new procedural library and its functions. The paper's approach aims to simplify and expedite this process.", "category": "Method_Comparison"}, {"question": "Can 3D-GPT handle modeling tasks for objects and scenes beyond 3D plants and forests, such as humans and streets, and generalize well to more complex scenarios?", "answer": "Yes, 3D-GPT, leveraging the procedural algorithm library Infinigen, can handle various 3D modeling tasks in diverse contexts, including objects and scenes beyond plants and forests, without requiring fine-tuning. It operates in a zero-shot manner, independently processing each modeling function by the conceptualization and modeling agents, allowing it to adapt to new functions with the necessary information. ", "category": "Experimental_Analysis"}, {"question": "Why is ICLR considered a suitable venue for this paper, given its graphics focus and the suggestion to submit to SIGGRAPH or ToG?", "answer": "ICLR is considered suitable for this paper because it aligns with the broad scope of machine learning topics, particularly in the domain of large language models (LLMs) and AI agents, which are of keen interest to the ICLR community. The paper introduces 3D-GPT, a novel approach to the text-to-3D problem, addressing significant challenges in 3D scene creation such as generating expansive scenes, animating 3D models, maintaining coherence, and creating intricate details. These contributions are relevant to ICLR as they involve machine learning-based solutions.", "category": "Method_Disambiguation"}, {"question": "What distinguishes the novelty and complexity of the paper's work on combining LLMs with procedural 3D modeling tools compared to existing methods that use ChatGPT/GPT-4 for generating parameters for images[1] or 3D objects[2]?\n\n[1] Bubeck, Sébastien, et al. \"Sparks of artificial general intelligence: Early experiments with gpt-4.\" arXiv preprint arXiv:2303.12712 (2023). \n\n[2] https://twitter.com/gd3kr/status/1638149299925307392", "answer": "It is essential to differentiate between image generation [1] and 3D modeling, as these areas pose distinct challenges. Moreover, the 3D examples provided by [2] do not reflect the complexities the paper addresses. Such a way to generate Blender Python code lacks a comprehensive understanding of API functions and their alignment with natural language. The generated Python code would have high error rates. However, it lacks mechanisms for verifying code executability and correctness. Furthermore, directly generating Python codes does not effectively manage dependencies between 3D contents. For instance, when generating trees, the spatial relationship between the leaf and tree branches is not explicitly modeled. As a result, this method of applying LLMs to Blender code could produce inadequate 3D models, especially in the context of complex large-scale scenes. Additionally, the paper's research demonstrates that a simple combination of LLMs with the procedural generation library InfiniGen leads to subpar results. While the idea of combining LLMs with procedural 3D modeling may now appear straightforward in hindsight, creating an effective workflow is far from simple. The paper's solution involves three distinct agents: the Task Dispatch Agent interprets and chooses the right API calls, the Conceptualization Agent enhances the scene description with adequate details, and the Modeling Agent comprehends how textual descriptions translate into specific API parameters. This multi-agent approach ensures effective and detailed 3D model generation from textual inputs. The results presented in Section 4.2 (Table 1, Figure 6) and in Section 6.2 (Table 2 and Table 3) of the paper clearly illustrate the significant roles played by each component within the paper’s framework in enhancing the final output.", "category": "Method_Comparison"}, {"question": "How does the proposed method for text-guided 3D scene generation compare to the score-distillation-based method DreamFusion[1] in terms of generating single objects and large-scale natural scenes?\n\n[1] DreamFusion: Text-to-3D using 2D Diffusion. In ICLR 2023.", "answer": "The proposed method has been compared with DreamFusion in the paper (Section 6.3, Figures 7 and 8). While DreamFusion generates reasonable 3D for single objects, it struggles with large-scale natural scenes. In contrast, the proposed method is effective in both scenarios, showcasing its advantages in handling more complex 3D modeling tasks. Additionally, the proposed framework supports sequential scene editing without requiring re-training. Furthermore, it's worth noting that previous text-to-3D methods encountered challenges when generating large-scale natural scenes. These observations emphasize the strengths and advantages of the paper's approach in handling more complex 3D modeling tasks.\n", "category": "Method_Comparison"}, {"question": "Do the selected functions establish dependencies that ensure correct positioning of flowers based on tree branches in the 3D model generation process?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does 3D-GPT require fine-tuning to adapt to new 3D modeling tasks in different contexts?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is 3D-GPT limited to generating 3D plants and forests, or can it also handle other objects and scenes such as humans and streets?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "u6jbcaCHqO", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models", "authors": ["Xiaoxuan Wang", "Ziniu Hu", "Pan Lu", "Yanqiao Zhu", "Jieyu Zhang", "Satyen Subramaniam", "Arjun R Loomba", "Shichang Zhang", "Yizhou Sun", "Wei Wang"], "abstract": "Recent advances in large language models (LLMs) have demonstrated notable progress on many mathematical benchmarks. However, most of these benchmarks only feature problems grounded in junior and senior high school subjects, contain only multiple-choice questions, and are confined to a limited scope of elementary arithmetic operations. To address these issues, this paper introduces an expansive benchmark suite SciBench that aims to systematically examine the reasoning capabilities required for complex scientific problem solving. SciBench contains two carefully curated datasets: an open set featuring a range of collegiate-level scientific problems drawn from mathematics, chemistry, and physics textbooks, and a closed set comprising problems from undergraduate-level exams in computer science and mathematics. Based on the two datasets, we conduct an in-depth benchmark study of two representative LLMs with various prompting strategies. The results reveal that current LLMs fall short of delivering satisfactory performance, with an overall score of merely 35.80%. Furthermore, through a detailed user study, we categorize the errors made by LLMs into ten problem-solving abilities. Our analysis indicates that no single prompting strategy significantly outperforms others and some strategies that demonstrate improvements in certain problem-solving skills result in declines in other skills. We envision that SciBench will catalyze further developments in the reasoning abilities of LLMs, thereby ultimately contributing to scientific research and discovery.", "keywords": "scientific computing;problem solving;large language models", "qa_pairs": [{"question": "What is the distinction between the 'Evaluation' and 'Analysis' sections in Table 1, and how does the 'Auto' column in the 'Analysis' section relate to the automatic evaluation process used in the MATH dataset?", "answer": "In Table 1, the 'Evaluation' section evaluates the performance of LLMs under different settings, while the 'Analysis' section focuses on identifying specific problem-solving abilities where each setting falls short. The 'Auto' column in the 'Analysis' section indicates the use of an automated process to analyze the results produced by LLMs, categorizing errors into specific lacking abilities. This automated process aligns with the automatic evaluation used in the MATH dataset, which constrains outputs in the \\boxed{...}environment.", "category": "Concept_Understanding"}, {"question": "Why did the paper not initially explore the combination of external tools and zero-shot learning, given that Table 4 indicates few-shot learning sometimes has lower accuracy than zero-shot learning?", "answer": "The authors had concerns about the ability of LLMs to adapt to programming languages, particularly Wolfram Language, in a zero-shot setting. They conducted an experiment using gpt-3.5-turbo in a zero-shot Python setting, and present the results below.\n\n| atkins | cal   | chemmc | class | diff  | fund  | matter | quan  | stat  | thermo | avg |\n|--------|-------|--------|-------|-------|-------|--------|-------|-------|--------|---------|\n | 11.21  | 19.04 | 30.77  | 0.0   | 14.00 | 9.59  | 2.04   | 11.76 | 42.67 | 4.48   | 14.75   |", "category": "Experimental_Analysis"}, {"question": "What is the cause of the trade-off in skills observed in few-shot prompting, and can it be mitigated by changing the example prompts?", "answer": "The trade-off in skills in few-shot prompting is not due to the specific content of the example prompts. The paper's experiments, which involved multiple rounds of random selection of prompt examples, showed that the performance of the LLM in the few-shot setting is comparable to the zero-shot setting. The underlying issue is likely that the number of examples is insufficient to cover all necessary abilities, leading to distraction for the model.", "category": "Experimental_Analysis"}, {"question": "What are the key differences between SciBench and BIG-bench[1]?\n\n[1] https://arxiv.org/abs/2206.046159", "answer": "BIG-bench is different from SciBench because it mainly targets specific tasks like predicting element names or deducing how physical mechanisms work, with some of its tasks as high-school level with multiple-choice questions. SciBench, on the other hand, uses harder college-level free-response problems.", "category": "Method_Comparison"}, {"question": "Given that “Mathematical Capabilities of ChatGPT” (https://arxiv.org/abs/2301.13867) and “NaturalProofs” (https://arxiv.org/abs/2104.01112) for math already do some of the things SciBench proposed to achieve. For example, both mentioned papers achieve the enabling of assessing advanced problem - solving ability (see page 4). The first reference paper also achieves the inclusion of college - level problems and inaccessibility in text formats. So, how does the SciBench dataset differentiate itself from existing datasets in terms of the types of problems and required background knowledge?", "answer": "The SciBench dataset includes problems from various disciplines such as chemistry, physics, and computer science, requiring specific in-domain background knowledge. In contrast, the mentioned datasets(Mathematical Capabilities of ChatGPT and NaturalProofs) focus solely on mathematics, relying primarily on reasoning ability. There are fewer corpora and evaluation benchmarks for other scientific subjects compared to mathematics.", "category": "Method_Comparison"}, {"question": "What are the key differences between SciBench and other datasets mentioned, and how does Table 1 accurately represent the dataset, experiments, and analysis?", "answer": "The key difference of SciBench lies in its use of college-level questions requiring complex reasoning and computational steps, such as differentiation, and its preference for free-response answers over multiple-choice formats. It employs various configurations of Language Learning Models (LLMs) like Chain of Thought and external tools for evaluation, along with an automated error analysis pipeline. Table 1 accurately represents the dataset, experiments, and analysis: the columns 'Level,' 'computation,' 'solution,' and 'type' refer to the dataset (Section 3); 'zero-shot,' 'few-shot,' 'CoT,' and 'tool' refer to experiments (Section 4); and 'Human' and 'auto' refer to analysis (Section 5). The evaluation part indicates whether other benchmark works have been tested using these configurations, and the analysis part indicates whether manual or automatic analysis was used. An in-text comparison is provided in Appendix D, Figure S15.", "category": "Method_Comparison"}, {"question": "How was GPT-4 utilized in the process of distilling errors into essential skills, and what measures were taken to ensure the accuracy and reliability of its contributions?", "answer": "Two human annotators participate in the process. Decisions on the final abilities are determined by annotators, aided by assistants. By examining the errors, these two annotators develop ten abilities and then employ a Language Learning Model (LLM) as a third evaluator to suggest additional abilities. They then compare and refine their findings based on this input. Ultimately, the final outcomes are determined by the annotators. To ensure accuracy, after LLM annotates the error reasons, a human - check was conducted on 151 sampled examples across all settings to verify the annotations’ correctness. This human - AI cooperation aimed to reduce human post - analysis costs while ensuring the reliability of GPT - 4’s decisions.", "category": "Experimental_Setup"}, {"question": "Why was GPT-4 with code interpreter (ADA) not included in the evaluation, and how does the newly released GPT4-Turbo compare to the original GPT4 in terms of performance?", "answer": "The API for code interpreters for GPT models, including GPT-4 with code interpreter (ADA), is currently not accessible. However, a new experiment was conducted using the newly released GPT4-Turbo API. The results show that GPT4-Turbo performs slightly lower than the original GPT4 version, as indicated by the average scores (28.3 for GPT4-Turbo vs. 35.8 for GPT4).\n\n|           | atkins | cal  | chemmc | class | diff | fund | matter | quan | stat | thermo | avg |\n|-----------|--------|------|--------|-------|------|------|--------|------|------|--------|---------|\n| gpt4-turbo| 28.0   | 23.8 | 20.5   | 21.3  | 36.0 | 31.5 | 16.3   | 35.3 | 53.3 | 9.0    | 28.3    |\n| gpt4      | 21.1   | 42.9 | 41.0   | 17.0  | 34.0 | 38.4 | 28.6   | 38.2 | 69.3 | 29.9   | 35.8    |\n", "category": "Experimental_Exposition"}, {"question": "How does the proposed method compare with state-of-the-art methods such as hypothesis search and refinement, and the use of code interpreters in the context of general scientific problem-solving tasks?", "answer": "The experimental results under ‘CoT+Py’ and ‘CoT+Wol’ settings demonstrate the integration of LLMs with external programmatic tools. Many extensive search methods are not easily extendable to general scientific problem-solving tasks. The authors chose the advanced self-consistency method [1] using GPT-3.5-Turbo, presented the results, and showed them as follows:\n\n| atkins | chemmc | quan | matter | fund | class | thermo | diff | stat | calculus | avg   |\n|--------|--------|------|--------|------|-------|--------|------|------|----------|-------|\n| 15.6   | 23.8   | 11.4 | 11.8   | 15.8 | 2.0   | 4.5    | 8.0  | 31.2 | 26.0     | 15.6  |\n\n1] Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., ... & Zhou, D. (2022). Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.\n\n", "category": "Method_Comparison"}, {"question": "How does the dataset ensure that the questions used are not already part of the training data for the models, given that textbooks are available online and easily converted into text?", "answer": "The dataset ensures that the questions are not already part of the training data by manually extracting and aligning questions and answers from textbooks, where problems and solutions are located in separate sections (e.g., problems on Page 100 and solutions on Page 695). This layout makes it challenging for Large Language Models (LLMs) to correlate them without specialized extraction methods, and human annotators were employed to perform this non-trivial task.", "category": "Method_Mechanics"}, {"question": "How is the complexity and difficulty of the SCIBENCH dataset quantified, given its claim of requiring 'multiple steps of reasoning'?", "answer": "SciBench is a college-level problem set covering various subjects, making difficulty subjective. Due to the wide scope, assembling consistent annotators is challenging. To indicate problem complexity, the dataset shows the number of sentences in problems with step-by-step solutions: 18.75% have <5 sentences, 26.79% have 5-9 sentences, 45.53% have 10-19 sentences, and 8.93% have 20+ sentences.", "category": "Experimental_Setup"}, {"question": "Is the evaluation method, which considers responses with a relative error of 0.05 as correct and disregards cases where the numerical calculation is wrong despite a correct idea or formula, suitable for assessing scientific problem-solving ability?", "answer": "Problems with wrong numerical calculations are considered incorrect because calculation ability is considered one of the problem-solving abilities. Models should have the calculation ability to answer problems correctly. Given the overall difficulty of the paper’s tested problems, it is highly unlikely that the model derives the correct answer using incorrect formulas. Unlike multiple-choice questions, where random guessing has a high chance of yielding the correct answer, the paper’s evaluation method is more stringent as it involves a wide range of rational numbers, making it highly improbable to guess the correct answer randomly. If the model arrives at the correct answer, it is very likely due to a correct solution process, given the complexity involved in reaching these answers. The paper’s answers, which can be as specific as -0.0987 or as large as 1036592783, are difficult to guess accurately. Implementing a process verifier could enhance the quality of evaluation. However, employing human annotators for this task could significantly increase project costs. The paper’s evaluation criteria are consistent with most existing benchmarks, like GSM8K[1], where the assessment is based solely on the numerical equivalence of the final answer. Further error analysis for inaccurately answered questions is also included in Section 5.\n\n[1] Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., ... & Schulman, J. (2021). Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.", "category": "Method_Disambiguation"}, {"question": "What is the rationale behind using the final answer as the evaluation metric for the proposed dataset, given that the difference between arithmetic for elementary school students and scientific problem-solving for college students involves more than just the number of reasoning steps, such as background knowledge and advanced calculations?", "answer": "The difference between arithmetic for elementary school students and scientific problem-solving for college students is not in the number of reasoning steps. In addition, it involves background knowledge and more advanced calculations, including integration and differentiation. The paper’s reason for the current evaluation metric is that if the model can provide the correct final answer, it is highly likely that it has utilized accurate background knowledge and performed the calculations correctly. Compared to other advanced dataset benchmarks like AGIEval[1], which uses a multiple-choice format, the paper’s evaluation metric more effectively reveals the accuracy of the problem-solving ability of a model by using the free-response format. The paper has shown an illustrative comparison between AGIEval and SciBench in Appendix Figure S15, where using a multiple-choice format might allow LLMs to guess the correct answer from the options but the free-response format will much lower the possibility. Incorporating other evaluating metrics might be helpful. In Section 5, the paper showed an analysis using LLMs to analyze the solution process made by the original LLM and identifying the corresponding lacking abilities, which further analyzes the problem-solving abilities of LLMs. \n\n[1] Zhong, W., Cui, R., Guo, Y., Liang, Y., Lu, S., Wang, Y., ... & Duan, N. (2023). Agieval: A human-centric benchmark for evaluating foundation models. arXiv preprint arXiv:2304.06364.\n\n\n", "category": "Motivation_Analysis"}, {"question": "Are cases where the numerical calculation is correct but the underlying idea or formula is wrong considered correct in the evaluation setup?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the error rates of zero - shot setting 15.2% for both causality ability and logical decomposition ability?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the mentioned datasets (Mathematical Capabilities of ChatGPT and NaturalProofs) cover problems from multiple scientific disciplines beyond mathematics?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was a process verifier implemented to enhance the quality of evaluation due to concerns about the evaluation method?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the new dataset include questions on which the models have already been trained, considering the separation of problems and solutions in the textbooks?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are any of the questions in the new datasets directly searchable online or present in the latest LLMs?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are parsing errors, in addition to syntax errors, checked by human annotators during the verification of LaTeX documents?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "lgvOSEMEQS", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Lightweight Unsupervised Federated Learning with Pretrained Vision Language Model", "authors": ["Hao Yan", "Yuhong Guo"], "abstract": "Federated learning aims to tackle the ``isolated data island\" problem, where it trains a collective model from physically isolated clients while safeguarding the privacy of users' data. However, supervised federated learning necessitates that each client labels their data for training, which can be both time-consuming and resource-intensive, and may even be impractical for edge devices. Moreover, the training and transmission of deep models present challenges to the computation and communication capabilities of the clients.\nTo address these two inherent challenges in supervised federated learning, we propose a novel lightweight unsupervised federated learning approach that leverages unlabeled data on each client to perform lightweight model training and communication by harnessing pretrained vision-language models, such as CLIP. By capitalizing on the zero-shot prediction capability and the well-trained image encoder of the pre-trained CLIP model, we have carefully crafted an efficient and resilient self-training approach. This method refines the initial zero-shot predicted pseudo-labels of unlabeled instances through the sole training of a linear classifier on top of the fixed image encoder. Additionally, to address data heterogeneity within each client, we propose a class-balanced text feature sampling strategy for generating synthetic instances in the feature space to support local training. \nExperiments are conducted on multiple benchmark datasets. The experimental results demonstrate that our proposed method greatly enhances model performance in comparison to CLIP's zero-shot predictions and even outperforms supervised federated learning benchmark methods given limited computational and communication overhead.", "keywords": "unsupervised federated learning;vision-language model", "qa_pairs": [{"question": "How is the lightweight nature of the proposed framework substantiated in terms of parameter optimization, communication rounds, and comprehensive analysis of computation and communication costs?", "answer": "This paper addresses the novel unsupervised federated learning problem and introduces a lightweight framework to address it. The lightweight nature of the paper's approach is evident in the following aspects: To minimize the number of parameters optimized in each client and transmitted between clients and the server, the paper introduces the CLIP image encoder, fixing it during local training. Only the weight parameters of the linear layer classifier are updated and transmitted. To further reduce the number of communication rounds, the paper initializes the weight parameters of the linear layer classifier with text embeddings associated with class names. This results in a minimal number of local update epochs and communication rounds required for model convergence. A comprehensive analysis of both computation and communication costs is presented below. Comparing the training of the full model, including the image encoder and the classifier, with the ResNet-50 based CLIP image encoder and a linear layer classifier with 10 classes (e.g., CIFAR-10), the number of parameters in the former is approximately 38 million (M), while the latter is about 10 thousand (K), making a difference of almost 3800 times. Computation Costs: Training the full model typically takes over 100 rounds with 5 local update epochs for each round according to FedAvg and FedNTD. This requires updating a total of $38M \\times 100 \\times 5$ parameters for each client. In contrast, the paper's proposed lightweight method usually takes 10 rounds with 1 local update epoch for each round, involving the updating of only $10K \\times 10 \\times 1$ parameters for each client. The difference is approximately 190K times. Communication Costs: Transmitting the full model incurs a cost of $38M \\times 100 \\times 32$ bits of bandwidth, while transmitting only the linear model costs $10K \\times 10 \\times 32$ bits of bandwidth. The difference is about 38K times.", "category": "Method_Disambiguation"}, {"question": "What novel contributions does this study make in the context of unsupervised federated learning, and how do they address the limitations of prior work?", "answer": "This study introduces a novel setting of unsupervised federated learning and presents a lightweight framework designed to address this challenge effectively. It is the first to tackle federated learning with completely unlabeled data, unlike prior work such as FedUL[1], which requires knowledge of precise label frequencies for each client. The proposed framework incorporates the CLIP image encoder and text embeddings to enhance unsupervised federated learning, trains a cleverly-initialized linear classifier on each client to reduce computation and communication requirements, and introduces a resilient self-training approach with evolving pseudo-labels. Additionally, a class-balanced data generation approach is proposed to handle heterogeneous data distribution by generating synthetic feature instances from textual embeddings. Experimental results demonstrate enhanced performance of CLIP's zero-shot prediction and rapid convergence, significantly reducing computation and communication demands.\n\n[1]Nan Lu, Zhao Wang, Xiaoxiao Li, Gang Niu, Qi Dou, and Masashi Sugiyama. Federated learning from only unlabeled data with class-conditional-sharing clients. In ICLR, 2022.", "category": "Method_Disambiguation"}, {"question": "How does the proposed method of class prototype augmentation based on Gaussian noise differ from prior work like PASS [Zhu CVPR 2021], and how does it address the issue of class overlap mentioned in follow-up works like [Zhu CVPR 2022]?\n\n[Zhu CVPR 2021] Prototype Augmentation and Self-Supervision for Incremental Learning, CVPR,2021\n\n[Zhu CVPR 2022]. \"Self-sustaining representation expansion for non-exemplar class-incremental learning.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.", "answer": "The proposed method differs from PASS [Zhu CVPR 2021] by generating synthetic instances from textual embeddings produced by the pre-trained CLIP text encoder, rather than from prototypes during model training. The textual feature vectors are trained to be discriminative to enable zero-shot prediction. Additionally, controlling the variance of the Gaussian distribution effectively avoids feature distribution overlap across classes. Experiments with different $\\sigma$ values showed that unit variance produces the best performance. The innovation of the work lies in the introduction of a novel unsupervised federated learning setting and a lightweight framework, not in the use of Gaussian augmentation. It's worth noting that even the technique of feature-level Gaussian augmentation, which PASS employs, is not a novel introduction. Prior work, specifically by [DeVries and Taylor, 2017], had already employed this method for feature-level data augmentation. This technique is essentially a straightforward application of fundamental principles from the Gaussian distribution.\n\n[DeVries and Taylor, 2017] DeVries, Terrance and Taylor, Graham W. Dataset augmentation in feature space. arXiv 2017.", "category": "Method_Comparison"}, {"question": "Why does the testing accuracy of baselines drop as the number of communication rounds increases?", "answer": "The testing accuracy of baselines drops because, as shown in Figure 2, the predicted probability vectors from CLIP zero-shot prediction exhibit high entropy, resembling a uniform distribution. When the CLIP image encoder is fixed and the linear classifier is trained with strong supervision from ground-truth labels, the classifier weights diverge significantly from the textual embeddings. This divergence leads to biased and highly dynamic local classifiers, resulting in degraded performance. In contrast, the paper's approach uses resilient self-training with evolving pseudo-labels to stabilize the local model updating process and incorporates class-balanced synthetic data generation to address the heterogeneous distribution of local data.", "category": "Experimental_Analysis"}, {"question": "Why was the FedUL baseline not included in the comparison with the proposed approach?", "answer": "FedUL requires precise knowledge of label frequencies for each client, which makes it an unfair comparison for the proposed approach. Furthermore, FedUL involves transforming and recovering model parameters, whereas the proposed lightweight framework maintains a fixed state for the image encoder, making seamless adaptation impossible.", "category": "Motivation_Analysis"}, {"question": "How does the proposed lightweight framework ensure that devices can accommodate the model size of CLIP and have sufficient memory for inference?", "answer": "The proposed lightweight framework generates textual embeddings on the server side and distributes them to each client, eliminating the need for repeated inference from the CLIP text encoder on the client. Additionally, the smallest variant of the CLIP image encoder, RN50, is used, which ensures that devices capable of accommodating ResNet-50 for inference can also support the model in the proposed method.", "category": "Method_Mechanics"}, {"question": "Why were certain unsupervised federated learning baselines and methods, such as [1][2][3][4], not discussed or compared in the paper?\n\n[1] Collaborative Unsupervised Visual Representation Learning from Decentralized Data. ICCV’21\n\n[2] Divergence-aware Federated Self-Supervised Learning. ICLR’22.\n\n[3] Orchestra: Unsupervised Federated Learning via Globally Consistent Clustering. ICML’22\n\n[4] MocoSFL: Enabling Cross-client Collaborative Self-supervised Learning. ICLR’23", "answer": "The paper addresses the novel challenge of unsupervised federated learning with a lightweight framework that incorporates a pretrained CLIP image encoder to minimize computation and communication demands. The listed works [1][2][3][4]focus on self-supervised federated representation learning and propose techniques to train the image encoder in a federated manner, which is not directly relevant to the problem addressed in this paper.", "category": "Motivation_Analysis"}, {"question": "How does the proposed method differ from existing approaches like FedCLIP[1] that also adopt CLIP in federated learning?\n\n[1]Fedclip: Fast generalization and personalization for clip in federated learning.", "answer": "The proposed method differs from FedCLIP in that it is designed for unsupervised federated learning, where only unlabeled data is available in the clients, whereas FedCLIP addresses the traditional supervised federated learning setting with labeled client data. This distinction highlights the unique contribution and applicability of the paper's approach in scenarios with exclusively unlabeled client data.", "category": "Method_Comparison"}, {"question": "How does the choice of different CLIP image encoder variants impact the performance and computational requirements of the proposed lightweight unsupervised federated learning method?", "answer": "The CLIP model offers a range of image encoder variants, each striking a balance between zero-shot prediction ability and model size. In the context of the lightweight unsupervised federated learning setting and with the aim of minimizing computational requirements on local clients, the paper opted for the smallest variant, RN-50, as the fixed image encoder. It's noteworthy that while the proposed method's performance benefits from the CLIP zero-shot prediction model, incorporating larger variants of the CLIP image encoder can further enhance performance, starting from the corresponding zero-shot performance of the selected image encoder. However, this improvement comes at the cost of increased computational demands on the local clients.", "category": "Experimental_Analysis"}, {"question": "What is the impact of increasing the number of local update epochs on the performance of the proposed method across CIFAR-10 and CIFAR-100 datasets under homogeneous and heterogeneous settings?", "answer": "Increasing the number of local update epochs did not lead to a notable improvement in the performance of the proposed method. The overall best performance was achieved with only 1 local update epoch. This observation underscores the lightweight property of the proposed method and eliminates the necessity for tuning this hyper-parameter. The phenomenon can be attributed to the high entropy property of CLIP zero-shot prediction, emphasizing the need for stable local updating.\n\n| # epochs | CIFAR-10 |  | CIFAR-100 |  |\n|:---:|---|---|---|---|\n|  | s=2 | i.i.d. | s=10 | i.i.d. |\n| 1 | 72.0 | 74.0 | 43.3 | 43.2 |\n| 2 | 71.7 | 73.6 | 41.4 | 41.2 |\n| 3 | 72.0 | 73.1 | 37.3 | 37.4 |\n| 4 | 72.4 | 73.2 | 33.0 | 31.7 |\n| 5 | 73.0 | 73.2 | 28.6 | 26.6 |", "category": "Experimental_Exposition"}, {"question": "Do all the baselines adopt the same model architecture as the proposed method to ensure a fair comparison?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does each baseline utilize the pretrained CLIP image encoder (RN-50 variant) as the fixed encoder and subsequently train a linear layer classifier?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "OJoMzslBIa", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "ReweightOOD: Loss Reweighting for Distance-based OOD Detection", "authors": ["Sudarshan Regmi", "Bibek Panthi", "Yifei Ming", "Prashnna Kumar Gyawali", "Danail Stoyanov", "Binod Bhattarai"], "abstract": "Out-of-Distribution (OOD) detection is crucial for ensuring the safety and reliability of neural networks in critical applications. Distance-based OOD detection is based on the assumption that OOD samples are mapped far from In-Distribution (ID) clusters in embedding space. A recent approach for obtaining OOD-detection-friendly embedding space has been contrastive optimization of pulling similar pairs and pushing apart dissimilar pairs. It assigns equal significance to all similarity instances with the implicit objective of maximizing the mean proximity between samples with their corresponding hypothetical class centroids. However, the emphasis should be directed towards reducing the Minimum Enclosing Sphere (MES) for each class and achieving higher inter-class dispersion to effectively mitigate the potential for ID-OOD overlap. Optimizing low-signal dissimilar pairs might potentially act against achieving maximal inter-class dispersion while less-optimized similar pairs prevent achieving smaller MES. Based on this, we propose a reweighting scheme \\textbf{ReweightOOD}, that adopts the similarity optimization which prioritizes the optimization of less-optimized contrasting pairs while assigning lower importance to already well-optimized contrasting pairs. Such a reweighting scheme serves to minimize the MES for each class while achieving maximal inter-class dispersion. Experimental results on a challenging CIFAR100 benchmark using ResNet-18 network demonstrate that the proposed reweighting scheme improves the FPR metric by a whopping ~38\\% in comparison to the baseline. In various classification datasets, our method outperforms existing methods, making it a promising solution for enhancing OOD detection capabilities in neural networks.", "keywords": "Out of Distribution Detection", "qa_pairs": [{"question": "What are the key differences between the paper's work and CurricularFace[1] and AdaFace[2] in terms of hard samples re-weighting?\n\n[1] CurricularFace: Adaptive Curriculum Learning Loss for Deep Face Recognition \n\n[2] AdaFace: Quality Adaptive Margin for Face Recognition", "answer": "CurricularFace adopts a curriculum-style learning approach, wherein early/late training dynamics are taken into account. It emphasizes easy samples at the earlier stage while prioritizing the hard samples at the later stage. ReweightOOD does not impose curriculum-style learning at all, as it does not distinguish between early and late training phases. The reweighting function of ReweightOOD, in the form of a sigmoid, strategically suppresses easy samples from the outset, directing attention toward challenging samples. AdaFace, another proxy-based approach, adopts the quality of the image — approximated by feature norm — into the weighting of the sample's importance. In AdaFace, hard samples from higher-quality images receive increased weighting, while easy samples from lower-quality images are weighted higher. In contrast, the paper's approach focuses solely on the cosine similarity when reweighting pair instances. Moreover, ReweightOOD facilitates distinct treatment of positive and negative pairs due to the presence of a transformation function with separate hyperparameters. The utilization of two sets of linear transformations, specific to positive and negative pairs, enables flexible weighting tailored to achieving optimal OOD detection. Additionally, both CurricularFace and AdaFace adopt cross-entropy-based optimization while ReweightOOD uses similarity maximization/minimization optimization as shown in equation (3).", "category": "Method_Comparison"}, {"question": "What are the results of the CIDER method compared to other methods, including SupCon and the proposed approach, on the ImageNet100 dataset?", "answer": "The result of ImageNet100 including CIDER is presented below:\n\n| Method                 | iNaturalist | SUN  | Textures | Places | SSB Hard | Ninco | Openimage | Average **FPR** $↓$ |\n|------------------------|-------------|------|----------|--------|----------|-------|-----------|--------------------------------|\n| Baseline               | 3.07        | 2.39 | 4.57     | 5.47   | 35.39    | 29.15 | 7.05      | 12.44                          |\n| SupCon                 | 2.43        | 1.98 | 2.59     | 5.43   | 34.25    | 25.58 | 5.28      | 11.08                          |\n| CIDER                  | 5.07        | 1.85 | 1.77     | 5.70   | 37.35    | 27.93 | 5.73      | 12.20                          |\n| **(ReweightOOD) *The paper***| 2.18        | 1.97 | 2.73     | 5.29   | 32.00    | 24.63 | 5.06      | **10.55**                      |\n\nAs can be observed from the above Table, the paper's approach yields the minimum average FPR of 10.55.", "category": "Method_Comparison"}, {"question": "What are the MES score and average centroid dispersion for CIDER, and how do they compare to other methods?", "answer": "The MES Table with the results of CIDER is given below:\n\n| Method                 | Apples | Aquarium Fish | Baby | Bear | Beaver | Bed  | Bee | Beetle | Bicycle | Bottles | ... | Mean  |\n|------------------------|--------|---------------|------|------|--------|------|-----|--------|---------|---------|-----|-------|\n| Baseline               | 1.09   | 1.11          | 1.07 | 0.93 | 0.97   | 1.07 | 1.03| 1.07   | 1.10    | 1.07    | ... | 1.05  |\n| SupCon                 | 0.97   | 1.01          | 0.97 | 1.00 | 1.00   | 0.98 | 0.99| 0.97   | 1.04    | 1.03    | ... | 1.01  |\n| *CIDER*                | 0.96   | 0.94          | 0.89 | 0.91 | 0.86   | 0.88 | 0.87| 0.88   | 1.04    | 0.99    | ... | 0.92  |\n| **(ReweightOOD) *The paper***| 0.95   | 0.90          | 0.90 | 0.89 | 0.84   | 0.90 | 0.89| 0.91   | 0.99    | 0.96    | ... | **0.90** |\n\nAs evidenced from the above Table, the average MES of ReweightOOD is minimum among all. Moreover, the class centroid dispersion of ReweightOOD (0.63) is comparable to that of CIDER (0.64) despite not explicitly encouraging the inter-class dispersion.", "category": "Method_Comparison"}, {"question": "How do the hyperparameters for linear transformation need to be set for different backbones or datasets, given that ResNet18 uses (5, −2, 2, 1) and (5, −4, 2, 1)?", "answer": "Random hyperparameter optimization is used based on the validation set with Gaussian noise. The hyperparameter set (5, −4, 2, 1) shows consistent strong performance across both CIFAR100 and ImageNet100 datasets for all architectures tested in the paper.", "category": "Experimental_Setup"}, {"question": "How does the proposed approach perform in comparison to existing methods on ImageNet datasets? ", "answer": "The authors adopted the training setup of ImageNet100 and presented a comparison with ImageNet-1k.\n\n| Method                 | iNaturalist | SUN  | Textures | Places | SSB Hard | Ninco | Openimage | Average **FPR** $↓$ |\n|------------------------|-------------|------|----------|--------|----------|-------|-----------|--------------------------------|\n| SupCon                 | 61.93        | 68.85 | 25.43     | 75.28   | 90.79    | 78.82 | 55.64      | 65.25                          |\n| CIDER                   | 53.54        | 61.27 | 23.95     | 68.76   | 91.52    | 79.15 | 54.22      | 61.77                          |\n| **(ReweightOOD) *Ours***| 50.70        | 59.46 | 24.27     | 66.27   | 90.18   | 76.76 | 50.97      | **59.80**                      |\n\nIt can be observed that the proposed approach has noticeable gains over SupCon and CIDER demonstrating its superiority.", "category": "Method_Comparison"}, {"question": "Have experiments been conducted to evaluate the proposed method on the ImageNet-1k benchmark, and if so, what are the results compared to other methods?", "answer": "The experiments with the Imagenet-1k dataset following the setup of the ImageNet-100 experiment are shown below.\n\n| Method                 | iNaturalist | SUN  | Textures | Places | SSB Hard | Ninco | Openimage | Average **FPR** $↓$ |\n|------------------------|-------------|------|----------|--------|----------|-------|-----------|--------------------------------|\n| SupCon                 | 61.93        | 68.85 | 25.43     | 75.28   | 90.79    | 78.82 | 55.64      | 65.25                          |\n| CIDER                   | 53.54        | 61.27 | 23.95     | 68.76   | 91.52    | 79.15 | 54.22      | 61.77                          |\n| **(ReweightOOD) *The paper***| 50.70        | 59.46 | 24.27     | 66.27   | 90.18   | 76.76 | 50.97      | **59.80**                      |\n\nIt can be observed that the proposed approach has noticeable gains over SupCon and CIDER demonstrating its superiority.", "category": "Method_Comparison"}, {"question": "Does the proposed method demonstrate effectiveness on the ImageNet-1k benchmark, and how does it compare to SupCon and CIDER?", "answer": "The ImageNet-1k experiments adopting the ImageNet100 setup is presented below:\n\n  | Method                 | iNaturalist | SUN  | Textures | Places | SSB Hard | Ninco | Openimage | Average **FPR** $\\downarrow$ |\n  |------------------------|-------------|------|----------|--------|----------|-------|-----------|--------------------------------|\n  | SupCon                 | 61.93        | 68.85 | 25.43     | 75.28   | 90.79    | 78.82 | 55.64      | 65.25                          |\n  | CIDER                   | 53.54        | 61.27 | 23.95     | 68.76   | 91.52    | 79.15 | 54.22      | 61.77                          |\n  | **(ReweightOOD) *Ours***| 50.70        | 59.46 | 24.27     | 66.27   | 90.18   | 76.76 | 50.97      | **59.80**                      |\n\nIt can be observed that the proposed approach has noticeable gains over SupCon and CIDER demonstrating its superiority.", "category": "Method_Comparison"}, {"question": "How does the proposed method contribute to novelty in the field of OOD detection, given that designing different weights for hard samples is a common approach in other fields?", "answer": "The proposed method's novelty lies in being the first to study the reweighting approach's efficacy specifically in OOD detection, a domain where such work has been previously missing. The paper's method demonstrates superior performance compared to competitive approaches well suited for OOD detection like SupCon, CSI, and CIDER, and shows remarkable results against strong post-hoc methods.", "category": "Method_Disambiguation"}, {"question": "Why are the parameters of the two linear transformation layers set to (5, -4, 2, 1) or (5, -2, 2, 1), and were these hyperparameters studied in the ablation experiments?", "answer": "The hyperparameters were determined from the validation set. A sensitivity study of these hyperparameters was conducted using the ResNet18 network with the CIFAR100 dataset, as shown in the provided table.\n\n|$m_b$ | Average FPR | &#124; | $c_b$ | Average FPR | &#124; | $m_w$ | Average FPR | &#124; | $c_w$ | Average FPR | &#124;\n|----|--------------|-------|-----|--------------|-------|-----|--------------|-------|-----|--------------|-------|\n| 4  | 51.01        | &#124;| 1   | 44.96        | &#124;| 1   | 69.10        | &#124;| 1   | 38.36        | &#124;|\n| 5  | 38.36        | &#124;| 2   | 38.36        | &#124;| 2   | 38.36        | &#124;| 2   | 67.63        | &#124;|\n| 6  | 39.14        | &#124;| 3   | 52.66        | &#124;| 3   | 43.44        | &#124;| 3   | 70.84        | &#124;|", "category": "Experimental_Analysis"}, {"question": "Are the hyperparameters (5, −4, 2, 1) consistently effective across different backbones and datasets as tested in the experiments?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the experiments conducted on the ImageNet-1k benchmark to evaluate the OOD method's performance?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the proposed approach show noticeable performance gains over SupCon and CIDER on the ImageNet-1k benchmark?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the study present results using the ImageNet100 setup to demonstrate the performance of the proposed method against SupCon and CIDER?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were experiments on the complete ImageNet-1K dataset conducted as part of the paper?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "fnO5h1CFyh", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Learning Successor Representations with Distributed Hebbian Temporal Memory", "authors": ["Evgenii Aleksandrovich Dzhivelikian", "Petr Kuderov", "Aleksandr Panov"], "abstract": "This paper presents a novel approach to address the challenge of online hidden representation learning for decision-making under uncertainty in non-stationary, partially observable environments. The proposed algorithm, Distributed Hebbian Temporal Memory (DHTM), is based on factor graph formalism and a multicomponent neuron model. DHTM aims to capture sequential data relationships and make cumulative predictions about future observations, forming Successor Representation (SR). Inspired by neurophysiological models of the neocortex, the algorithm utilizes distributed representations, sparse transition matrices, and local Hebbian-like learning rules to overcome the instability and slow learning process of traditional temporal memory algorithms like RNN and HMM. Experimental results demonstrate that DHTM outperforms classical LSTM and performs comparably to more advanced RNN-like algorithms, speeding up Temporal Difference learning for SR in changing environments. Additionally, we compare the SRs produced by DHTM to another biologically inspired HMM-like algorithm, CSCG. Our findings suggest that DHTM is a promising approach for addressing the challenges of online hidden representation learning in dynamic environments.", "keywords": "Temporal Memory;Successor Representation;Hebbian Learning;Factor Graph;Multicomponent Neuron Model", "qa_pairs": [{"question": "What is the method used for learning the synaptic segment weights $w_{ul}$, and how is it represented in the algorithm?", "answer": "The synaptic segment weights $w_{ul}$ are learned through local Hebbian-like plasticity, representing the probability $p(s_l=1\\ |\\ c^u_{t-1}=1)$ that segment $s_l$ is active given cell $u$ was active at the previous time-step. The algorithm approximates this as an exponential moving average of segment $s_l$'s frequency activation, given $c^u_{t-1}=1$, using the formula $\\Delta w_{ul} = \\alpha * \\mathbb{I}[c^u_{t-1}=1] * (\\mathbb{I}[s_l=1] - w_{ul})$, where $\\alpha \\in [0, 1)$ is the learning rate.", "category": "Method_Mechanics"}, {"question": "What additional experiments have been conducted to more comprehensively showcase the capabilities of the DHTM model, and how do these experiments address the limitations highlighted in the initial single experimental result, where the performance is marginally inferior to the fchmm, one of the baseline models?", "answer": "Additional experiments include a 2D maze experiment to qualitatively show the adequacy of DHTM’s value function estimate, examples of learned Successor Features for the initial setup, and results in the AnimalAI environment to demonstrate scalability and environment agnosticism. Although the paper’s method is inferior to FCHMM according to some metrics, unlike FCHMM and other baselines, the paper’s method learns in fully online mode and doesn’t employ the error backpropagation algorithm. FCHMM requires the entire sequence to be available when updating learnable parameters. Thus, FCHMM provides a lower [surprise] bound, which the paper seeks to achieve using only online Hebbian learning. However, the hidden state representations resulting from FCHMM are dense, unlike those of the paper’s method, and this density hinders linear mapping of the hidden space to Successor Features.", "category": "Experimental_Analysis"}, {"question": "How does the proposed method address scalability challenges for infinite horizon or large POMDPs, and what are the limitations?", "answer": "The proposed method combines planning and TD learning to achieve n-step TD learning, eliminating the need for infinite horizon predictions by learning a linear function mapping from hidden space to Successor Features space. It predicts the first several steps to speed up TD learning and reduce bias, approximating the tail with the learned mapping. Scalability to large POMDPs is enabled by using multiple feature variables, which create a sparse distributed representation of observations. Sparsity of observations and connections is key to handling high dimensionality. Experiments in the AnimalAI environment demonstrate scalability with complex visual inputs. Scalability was the primary goal in designing DHTM. Nevertheless, the method may still face scalability limitations.", "category": "Method_Disambiguation"}, {"question": "Why is the target policy ignored in DHTM, and how does the method adapt to a changing policy?", "answer": "The target policy is not ignored in DHTM. Actions are included in the model by forcing some of the variables $H^k_{t-1}$ to represent actions, with the assumption that information about the action is included in the hidden state of the model. For example, if there are 4 actions, 4 states are set for one of the hidden variables, and its state is determined from the observation of the action. DHTM forms on-policy SRs, relying on the policy iteration theorem, and can be used to form SRs for any policy, provided it has learned the dynamics of the environment.", "category": "Method_Disambiguation"}, {"question": "How is the pseudo-surprise of SF computed in Section 4?", "answer": "To compute the pseudo-surprise of SF, the following steps are taken: 1. Normalize SF by summing it over corresponding variables $\\Phi$ and dividing SF by these sums, resulting in the SF-induced probability distribution $p(\\varphi^k)$ of $\\Phi^k$ variables. 2. Measure the average surprise over future observed states $\\varphi^k$ according to this distribution: $-\\log p(\\varphi^k=j)$, where $j$ is the observed state. Normalized SF represents the future observation (feature) profile for the current state. Pseudo-surprise indicates whether SF is consistent with the observed feature states. For example, if SF does not predict feature $j$ ($p(\\varphi^k=j) = 0$), but it is observed, this results in infinite surprise, indicating poor quality SF.", "category": "Method_Mechanics"}, {"question": "How is the SR representation defined in the context of the model, particularly in relation to the HMM states and actions?", "answer": "The SR representation includes actions by designating specific variables $H^k_{t-1}$ to represent actions. This means that information about actions is incorporated into the hidden state of the model. For instance, if there are 4 actions, 4 states are assigned to one of the hidden variables, and these states are determined based on the observed action.", "category": "Method_Mechanics"}, {"question": "Why can the 'independent' variable be represented in the form of Eq. (11), and how are the parameters such as $w_{ul}$ learned using Hebbian rules?", "answer": "Eq. (11) is a linear approximation of the log-likelihood of segment activation between two extreme cases. The synaptic weights $w_{ul}$ are learned to represent the probability that segment $l$ is active given that cell $u$ is active, denoted as $p(s_l=1 | c^u_{t-1}=1)$. These weights are estimated empirically by counting coincidences of segment and presynaptic cell activities. Segment $l$ learns to respond to a cluster of active cells, that’s why if $w_{ul} \\to 1$, that means cell $u$ is clustered, otherwise it is independent of the cluster that activates segment $l$.", "category": "Method_Mechanics"}, {"question": "How are actions incorporated into the Hidden Markov Model (HMM) as described in Section 3.1?", "answer": "Actions are included in the model by designating some of the variables $H^k_{t-1}$ to represent actions. This means that information about actions is assumed to be part of the hidden state of the model. For instance, if there are 4 actions, 4 states are set for one of the hidden variables, and these states are determined based on the observed action.", "category": "Method_Mechanics"}, {"question": "What are the specific biological correspondences of 'f' and 'm' within the context of sum–product message passing, and how do nodes transmit 'm' among each other?", "answer": "Factor value $f$ corresponds to a learnable parameter of a dendritic segment, specifically its efficacy, which determines how strongly the segment contributes to a cell's spiking. A segment can have several presynaptic cells and acts as a coincidence detector [1], explaining factors with multiple variables. Value $m$ represents the normalized spike frequency of the corresponding cell. \n\n[1] Stuart, Greg J., and Nelson Spruston. 2015. ‘Dendritic Integration: 60 Years of Progress’. Nature Neuroscience 18 (12): 1713–21. https://doi.org/10.1038/nn.4157.", "category": "Concept_Understanding"}, {"question": "Is there no loop in the paper’s graphical structure, so it needs only one iteration of message passing?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the empirical results based solely on a single custom environment without using any standard benchmarks or metrics?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "rcFXg2aqEj", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "LMDX: Language Model-based Document Information Extraction and Localization", "authors": ["Vincent Perot", "Kai Kang", "Florian Luisier", "Guolong Su", "Xiaoyu Sun", "Ramya Sree Boppana", "Zilong Wang", "Jiaqi Mu", "Hao Zhang", "Nan Hua"], "abstract": "Large Language Models (LLM) have revolutionized Natural Language Processing (NLP), improving state-of-the-art on many existing tasks and exhibiting emergent capabilities. However, LLMs have not yet been successfully applied on semi-structured document information extraction, which is at the core of many document processing workflows and consists of extracting key entities from a visually rich document (VRD) given a predefined target schema. The main obstacles to LLM adoption in that task have been the absence of layout encoding within LLMs, critical for a high quality extraction, and the lack of a grounding mechanism ensuring the answer is not hallucinated. In this paper, we introduce Language Model-based Document Information Extraction and Localization (LMDX), a methodology to adapt arbitrary LLMs for document information extraction. LMDX can do extraction of singular, repeated, and hierarchical entities, both with and without training data, while providing grounding guarantees and localizing the entities within the document. In particular, we apply LMDX to the PaLM 2-S LLM and evaluate it on VRDU and CORD benchmarks, setting a new state-of-the-art and showing how LMDX enables the creation of high quality, data-efficient parsers.", "keywords": "document understanding;information extraction;entity extraction;llm", "qa_pairs": [{"question": "What is the primary goal of the complexity in the prompt engineering approach, and how does it address the requirements of document information extraction?", "answer": "The complexity of the prompt engineering is not driven by a goal to reduce the number of tokens due to limited context length. The primary goal of the complexity in the prompt engineering approach is to communicate the layout information of the document to the model, which is necessary for high-quality extraction on visually-rich documents, and to derive the extracted entities’ bounding boxes, a major requirement in document information extraction. These goals are achieved through the coordinate tokens contained in the prompt. The chunking stage in the pipeline, detailed in Section 2.3, is included to allow ML practitioners to use the methodology on LLMs with context length limitations or on LLMs that cannot use their full context effectively[1], but it can be skipped for LLMs that support long context and can use it effectively.\n\n[1] Liu, Nelson F., Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. \"Lost in the middle: How language models use long contexts.\" arXiv preprint arXiv:2307.03172 (2023).", "category": "Method_Disambiguation"}, {"question": "What are the contributions of the LMDX methodology, aside from leveraging an LLM, to the performance advantages observed on the VRDU and CORD benchmarks?", "answer": "The performance advantages of LMDX are attributed to its prompting strategy, decoding approach, and base extraction training. For instance, the inclusion of coordinate tokens improved micro-F1 by 12-15% on VRDU Ad-Buy form, and base extraction training led to a 7-39% micro-F1 improvement, as shown in the ablation study in Table 4. The performance advantages come not only from the LLM, but also the prompting strategy which communicates the layout of the document to the model as well. LMDX furthermore localizes its predictions within the document.", "category": "Method_Disambiguation"}, {"question": "How does LMDX perform compared to a long-context LLM baseline like GPT-3.5 on the VRDU and CORD benchmarks?", "answer": "LMDX outperforms GPT-3.5 across all benchmarks on VRDU and CORD. GPT-3.5, which supports a context of 16k tokens, was used as a baseline due to lack of access to Claude 2. Unlike GPT-3.5's unlocalized predictions, LMDX introduces localized predictions within the document, contributing to its superior performance. The authors prompt the LLM with the raw OCR text along with extraction instructions for the entity values as shown below:\n\n{RAW_OCR_TEXT}\n\nGiven the document, extract the text value of the entities included in the schema in json format.\n- The extraction must respect the JSON schema.\n- Only extract entities specified in the schema. Do not skip any entity types.\n- The values must only include text found in the document.\n- Use null or [] for missing entity types.\n- Do not indent the json you produce.\n- Examples of valid string value format: \"$ 1234.50\", \"John Do\", null.\n- Examples of valid list value format: [\"$ 1234.50\", \"John Do\"], [].\n\nSchema: {\"file_date\": \"\", \"foreign_principle_name\": \"\", \"registrant_name\": \"\", \"registration_num\": \"\", \"signer_name\": \"\", \"signer_title\": \"\"}\n\n\nSee the results below:\n\n| Dataset / Model                          | GPT-3.5 | LMDX  |\n| :--------------------------------------- | :------: | :---: |\n| VRDU - Registration Form - Seen Template | 67.23%  | 73.81%|\n| VRDU - Registration Form - Mixed Template| 63.86%  | 71.65%|\n| VRDU - Registration Form - Unseen Template| 67.49%  | 74.94%|\n| VRDU - Ad-Buy Form - Mixed Template      | 30.05%  | 39.74%|\n| VRDU - Ad-Buy Form - Unseen Template     | 29.84%  | 39.33%|\n| CORD                                     | 58.25%  | 67.47%|\n\n\n", "category": "Method_Comparison"}, {"question": "Why was GPT-4 not included as a baseline in the evaluation, and what are the performance results of the alternative vision-text LLM, LLaVA-v1.5-13B, on document information extraction tasks?", "answer": "GPT-4 was not included as a baseline because the gpt-4-vision-preview model API was released after the paper submission deadline and is rate limited to 100 requests per day, while 10000 requests would be needed for the benchmarks. Instead, the authors evaluated LLaVA-v1.5-13B, an open-source vision-text LLM, which performed poorly on document information extraction tasks due to JSON errors, hallucinated entity values, and OCR errors. Unlike LMDX, LLaVA cannot specify entity locations within the document, failing to meet the desired properties of such a system.", "category": "Experimental_Exposition"}, {"question": "What were the experimental results for using contextual demonstrations to improve performance on this task?", "answer": "See the results on the CORD benchmark in the table below, where the authors prompt the LMDX base extractor model with N documents and extraction from the CORD training data using the following methods: Random N (randomly select N documents) and Nearest Neighbors (use sentenceT5 embeddings for nearest neighbor retrieval). Results show that few-shot helps as expected on CORD. There is a significant gain going from 0-shot to 1/3-shot, with smaller incremental returns as the number of shots increases. The nearest neighbors method has proven most effective, as it retrieves documents from similar templates, and is matching the performance of finetuning at N=10 examples.\n\n| N-Shot | 0      | 1      | 3      | 5      | 10     |\n| :----- | :----- | :----- | :----- | :----- | :----- |\n| Random | 67.47% | 74.96% | 84.88% | 86.47% | 87.26% |\n| Nearest Neighbors | 67.47% | 87.73% | 90.98% | 92.28% | 92.82% |\n", "category": "Experimental_Exposition"}, {"question": "How effective are coordinate tokens in generating prompts for document representation, and are line-level segments with 2 coordinates sufficient for various VRD data?", "answer": "Coordinate tokens are crucial for high-quality extraction, improving micro-F1 by an absolute 12-15% across data regimes on VRDU Ad-Buy Form Mixed, as shown in the ablation study in Table 4, where coordinate tokens are replaced by line index (See appendix A.6 for how the prompts look), this setup prevents the communication of any layout information to the model, yet still utilizes LMDX prompts and localizes the entities. Line-level segments with 2 coordinates are sufficient based on experiments, as attempts with 4 coordinates, word-level, and finer quantization buckets either showed no improvement or degraded performance. This is likely because additional coordinate tokens make the text sequence less similar to the data the LLM was pretrained on, making it harder to interpret LMDX-style sequences.", "category": "Experimental_Analysis"}, {"question": "How effectively does the LLM used in the study parse the provided schema JSON, and what are the error rates observed in the VRDU Ad-Buy Form Mixed dataset?", "answer": "Schema representation is critical. It takes a certain LLM size to properly parse the provided schema JSON. The authors initially started experimenting with PaLM 2-XXS, and the completions would include entity types not provided in the schema, and completely ignore some entity types. Starting at PaLM 2-Small (the LLM the paper used), the model is able to follow the schema near perfectly. The authors computed parse error rates on VRDU Ad-Buy Form Mixed, which shows the percentage of completions that were successfully parsed as JSON with the requested entity types present and values well formatted. See results below:\n\n| Dataset size | Invalid JSON | Invalid Entity Value Format | Entity Text Not Found |\n|--------------|--------------|-----------------------------|-----------------------|\n| 0            | 0.18%        | 0.04%                       | 0.59%                 |\n| 10           | 0.27%        | 0.04%                       | 0.44%                 |\n| 50           | 0.24%        | 0.00%                       | 0.17%                 |\n| 100          | 0.24%        | 0.00%                       | 0.13%                 |\n| 200          | 0.25%        | 0.00%                       | 0.09%                 |\n\nThe different possible errors are:\n\n* Invalid JSON Formatting: This error refers to cases for which Python's json.loads(completion) fails on a LLM's completion. The JSON parsing error rate is below 0.3% in all training settings.\n* Invalid Entity Value Format: This error refers to cases where the leaf entity value does not follow the expected `text-segment-1 XX|YY text-segment-2 XX|YY` format. The Invalid Entity Value Format Rate is below 0.05% in all training settings.\n* Entity Text Not Found: This error refers to cases where the segment identifier is valid, but the entity text does not appear on the predicted segments. The Entity Text Not Found error rate is below 0.6% in all training settings.\n\nDespite those rare errors, with the paper's sampling and decoding (aggregating) strategy, extraction is successful on every single document on the studied benchmarks. ", "category": "Experimental_Exposition"}, {"question": "Why did the paper choose to perform Top-K sampling and majority voting on individual chunks before merging, rather than applying Top-K sampling and majority voting across all chunks at once, even though the latter method might seem better for semantic integration quality?", "answer": "The authors chose to perform Top-K sampling for individual chunks and then merge because voting across all chunks directly would lead to lower quality extraction. For example, in a 3-chunk document with 2 completions per chunk, if the 'registrant_name' entity only appears in one chunk, then the following extraction would be obtained from the LLM:\n\n```\nChunk 1, Sample 1: { 'registrant_name': 'John 10|20' }\nChunk 1, Sample 2: { 'registrant_name': 'John 10|20' }\n\nChunk 2, Sample 1: { 'registrant_name': null }\nChunk 2, Sample 2: { 'registrant_name': null }\n\nChunk 3, Sample 1: { 'registrant_name': null }\nChunk 3, Sample 2: { 'registrant_name': null }\n```\n\nVoting across all chunks would result in 'registrant_name' being null (4 out of 6 completions = 67%). In contrast, voting on an individual chunk basis and then merging predictions ensures higher accuracy, such as choosing 'John' as the value for 'registrant_name' (2 out of 2 completions = 100%).", "category": "Motivation_Analysis"}, {"question": "How is the effect of hierarchical entity and entity localization evaluated in the study?", "answer": "With regards to hierarchical entity evaluation, this is directly evaluated: the paper highlights the F1 score on line_item entity (the hierarchical entity in VRDU) for all models and baselines in Table 2 and provides a discussion for it in Section 3.3 *Performance on Hierarchical Entities*. The table below synthesizes this information and shows the line_item F1:\n\n| Dataset Size | 0     | 10    | 50    | 100   | 200   |\n|--------------|-------|-------|-------|-------|-------|\n| FormNet      | N/A   | 5.72  | 19.06 | 18.80 | 21.86 |\n| LayoutLM     | N/A   | 6.95  | 19.50 | 21.26 | 23.90 |\n| LayoutLMv2   | N/A   | 9.96  | 20.98 | 23.52 | 25.46 |\n| LayoutLMv3   | N/A   | 5.92  | 19.53 | 22.08 | 24.51 |\n| LMDX         | 21.21 | 39.35 | 65.42 | 69.77 | 72.09 |\n\nOverall, previous SOTA methods (LayoutLM / FormNet) plateau around ~20-25% F1 due to their reliance on heuristics that do not benefit from additional training data, unlike LMDX. \n\nFor all models and baselines that can do entity localization, on VRDU Registration Form and Ad-Buy Form Mixed, the authors computed the localization accuracy, following the formula below:\n\n$$Accuracy_{Localization} = TotalEntityCount_{CorrectlyExtractedAndLocalized} \\div TotalEntityCount_{CorrectlyExtracted} $$\n\nSince LMDX localizes at the line level, localization verification is done at the line-level as well, i.e. localization is considered correct if the prediction bounding box is covered by the groundtruth line-level bounding box by more than 80%. See the results below.\n\nVRDU Registration Form Mixed:\n\n| Train Dataset Size / Model | LayoutLM | LayoutLMv2 | LayoutLMv3 | **LMDX (Ours)**   |\n|----------------------|----------|------------|------------|--------|\n| 0 doc                | N/A      | N/A        | N/A        | 93.21% |\n| 10 doc               |   98.71% |     99.00% |     99.20% | 99.75% |\n| 50 doc               |   99.69% |     99.54% |     99.39% | 99.87% |\n| 100 doc              |   99.63% |     99.72% |     99.72% | 99.92% |\n| 200 doc              |   99.69% |     99.75% |     99.67% | 99.87% |\n\nVRDU Ad-Buy Form Mixed:\n\n| Train Dataset Size / Model | LayoutLM | LayoutLMv2 | LayoutLMv3 | **LMDX (Ours)**   |\n|----------------------|----------|------------|------------|--------|\n| 0 doc                | N/A      | N/A        | N/A        | 88.18% |\n| 10 doc               |   92.60% |     93.95% |     90.68% | 94.51% |\n| 50 doc               |   95.24% |     95.64% |     95.28% | 98.28% |\n| 100 doc              |   95.09% |     95.72% |     95.88% | 98.69% |\n| 200 doc              |   95.38% |     95.78% |     95.95% | 98.65% |\n\nThis evaluates the localization quality independently of the extraction quality. ", "category": "Experimental_Analysis"}, {"question": "What are the technical and computational challenges in using open-source large language models (LLMs) with publicly accessible optical character recognition (OCR) tools for this study?", "answer": "Using open-source LLMs is not particularly technically difficult provided there is sufficient compute resources for supervised fine-tuning (SFT) and an open-source model with adequate input/output sequence length support. However, the experiments in this study are computationally expensive, requiring 81 fine-tuning runs, which is why the focus was on a single LLM, PaLM-2 S. Publicly available OCR tools from the benchmarks were already used in the study.", "category": "Method_Disambiguation"}, {"question": "What alternative prompts, schemas, or target formats did the authors try during development, and how were the final settings chosen?", "answer": "The authors tried multiple prompt formats during development, including:\n- 2 coordinates vs 4 coordinates for each segment: 4 coordinates led to a small quality decrease compared to 2 coordinates.\n- line level vs word level segment: word-level segment performed worse than line-level segment. This is likely due to the fact that word-level segments change the text sequence so much, it starts being very different from what the LLM was initially pretrained on (most tokens start being coordinate tokens), and as such the LLM struggles more to interpret the sequence.\n- 100 coordinate quantization buckets vs 1000 quantization buckets: this was overall neutral quality-wise but 1000 used more tokens overall in its prompt and completions.\n- Indent the json for the schema and completion : this was quality-neutral. Hence no indentation was chosen to save on prompt/completion tokens (a 14% save, leading to faster and cheaper inference).\n- More detailed instructions in the prompt: This was quality-neutral after base entity extraction training.\n-  Trying document first versus last in the prompt: Placing the document first provided a slight quality gain.\n\nFinal settings (2 coordinates, line-level segments, 100 quantization buckets, no JSON indentation) were chosen based on optimal performance with PaLM 2-Small, used in experiments on CORD and VRDU benchmarks.", "category": "Experimental_Setup"}, {"question": "Did the experiments show that the LLM's JSON parsing errors were consistently below 0.3% for invalid JSON formatting across different dataset sizes?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the study find that the Entity Text Not Found error rate was below 0.6% in all training settings?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the LLaVA-v1.5-13B vision-text LLM perform well on document information extraction tasks according to the computed baseline numbers?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the code for the PaLM2-derived models provided in the paper?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the extraction successful on every single document in the studied benchmarks despite rare parsing errors?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was a multi-modal vision-text LLM, such as GPT-4, tested as a baseline in the experiments?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are vision-text LLMs like LLaVA capable of specifying the entity location within a document for information extraction tasks?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "PIl69UIAWL", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "GraphLLM: Boosting Graph Reasoning Ability of Large Language Model", "authors": ["Ziwei Chai", "Tianjie Zhang", "Liang Wu", "Kaiqiao Han", "Xiaohai Hu", "Xuanwen Huang", "Yang Yang"], "abstract": "The advancement of Large Language Models (LLMs) has remarkably pushed the boundaries towards artificial general intelligence (AGI), with their exceptional ability on understanding diverse types of information,  including but not limited to images and audio.  Despite this progress, a critical gap remains in empowering LLMs to proficiently understand and reason on graph data. Recent studies underscore LLMs' underwhelming performance on fundamental graph reasoning tasks.  In this paper, we endeavor to unearth the obstacles that impede LLMs in graph reasoning, pinpointing the common practice of converting graphs into natural language descriptions (Graph2Text) as a fundamental bottleneck. To overcome this impediment, we introduce GraphLLM, a pioneering end-to-end approach that synergistically integrates graph learning models with LLMs. This synergy equips LLMs with the ability to proficiently interpret and reason on graph data, harnessing the superior expressive power of graph learning models. Our empirical evaluations across four fundamental graph reasoning tasks validate the effectiveness of GraphLLM. The results exhibit a substantial average accuracy enhancement of 54.44%, alongside a noteworthy context reduction of 96.45% across various graph reasoning tasks.", "keywords": "Large Language Models (LLMs); reasoning on graph; graph neural networks;", "qa_pairs": [{"question": "How is the encoder-decoder initialized and updated with the guidance of the pre-trained LLM, and what role does the pre-trained LLM play in this process?", "answer": "The encoder-decoder is randomly initialized, and the nodes' textual descriptions are embedded using the frozen embedding table from the pre-trained LLM. The guidance from the pre-trained LLM comes from backpropagating the encoder-decoder's gradient from the LLM during end-to-end training, which allows the encoder-decoder and the graph transformer to learn alignment with the pre-trained LLM.", "category": "Method_Mechanics"}, {"question": "Why is a fixed template not used for generating each node's text features in the Maximum Triplet Sum task, and how are the text features generated instead?", "answer": "A fixed template is not used to prevent the model from solving the task by relying on correlations rather than understanding and extracting key information. For example, if the key information appears in a fixed position in the text, the node understanding transformer may recognize the specific positional embedding and extract the information from that position rather than understanding it. Meanwhile, non-template text is considered a more realistic setting. Therefore, in the Maximum Triplet Sum task, the paper randomly selects different sentences (generated by GPT-3.5-Turbo) describing a person and randomly inserts a sentence containing the age to form each node's text feature.", "category": "Motivation_Analysis"}, {"question": "Can GraphLLM be used as a general graph-to-prefix encoder to improve the performance of large language models (LLMs) on different fundamental tasks?", "answer": "Yes, GraphLLM was tested as a general framework for different fundamental tasks. When jointly trained on Substructure Counting and Maximum Triplet Sum using the same settings as in the paper, GraphLLM achieved accuracies of 0.9946±0.0021 on Substructure Counting and 0.8472±0.0210 on Maximum Triplet Sum, demonstrating its preliminary general capability.", "category": "Experimental_Analysis"}, {"question": "Why were simple artificial tasks chosen for benchmarking the reasoning abilities of LLMs instead of more complex, real-world datasets?", "answer": "The authors chose simple artificial tasks for benchmarking because it is a well-established practice in the field, as exemplified by datasets like GSM8K, which is widely employed to evaluate LLMs' mathematical reasoning abilities. Real-world benchmarks, such as ogbn-arXiv, may pose lower demands on LLMs' graph reasoning abilities compared to the artificial tasks used in the study. For instance, determining the arXiv category of a paper based on its abstract is a simple task for LLMs (LLMs can even continue writing the paper or provide review comments based on the abstract), making it challenging to measure the benefits of graph reasoning abilities. Building a real-world benchmark that accurately measures these capabilities requires community-wide effort, so the authors opted for artificial datasets to achieve precise measurements at this early stage.\n", "category": "Motivation_Analysis"}, {"question": "Why is the design of Graph2Text intentionally structured to maximize the number of tokens, particularly in the context of the node understanding encoder?", "answer": "The design of Graph2Text maximizes the number of tokens to include key information such as the person's age and other descriptions generated by GPT-3.5-Turbo. This approach aims to evaluate the node understanding capabilities of GraphLLM and Graph2Text-based LLMs, specifically their ability to extract essential details from seemingly unrelated text. Additionally, the authors assessed Graph2Text with GPT-4 few-shot CoT. In Maximum Triplet Sum, the edge between people is described using ‘friendship’, maintaining an accuracy of 93.33%. In Shortest Path, the background is modified to cities connected with roads, where it incurs a cost to pass through a city. However, the accuracy decreases from 86.67% to 83.33%.\n", "category": "Motivation_Analysis"}, {"question": "Why were artificial graph tasks chosen instead of real-world tasks like node classification and link prediction to demonstrate the effectiveness of the proposed method, given that more real-world tasks are especially important for a technical paper where new approaches are introduced, as opposed to the previous observation-based papers such as NLGraph and GPT4Graph?", "answer": "NLGraph[1] primarily serves as a benchmarking study. The NLGraph paper itself mentions that 'we propose NLGraph (Natural Language Graph), a comprehensive benchmark of graph-based problem solving designed in natural language.' Given the paper's objective to evaluate and boost the graph reasoning abilities of large language models, **it is** logical to design the paper's tasks in a manner that closely mirrors the NLGraph benchmark. Additionally, it's important to note that the use of 'artificial' tasks for benchmarking the reasoning abilities of LLMs is a well-established practice in the paper's field. An example of this is the GSM8K dataset[2], widely employed to evaluate mathematical reasoning skills. This particular dataset consists of 'simple' problems, all of which can be accurately solved using arithmetic operations at the middle school level. Somewhat paradoxically, the established node classification benchmarks like ogbn-arXiv (with textual feature) might pose lower demands on LLMs' graph reasoning abilities compared to the tasks presented in this text. For LLMs, determining the arXiv category of a paper based on its abstract is a very simple task. (LLMs can even continue writing the paper or provide review comments based on the abstract.) As a result, measuring the benefits of graph reasoning abilities for this task has become quite challenging (see Table 16 & 17 in [3]). In the era of LLMs, building a real-world benchmark that truly accurately measures the graph reasoning capabilities of these models requires the effort of the entire community. After careful consideration, at this early stage, the authors have opted for artificial datasets to secure somewhat precise measurements.\n\n[1] Can Language Models Solve Graph Problems in Natural Language? (NeurIPS 23) https://arxiv.org/abs/2305.10037\n\n[2] https://paperswithcode.com/dataset/gsm8k\n\n[3] Exploring the Potential of Large Language Models (LLMs) in Learning on Graphs https://arxiv.org/abs/2307.03393", "category": "Motivation_Analysis"}, {"question": "Can GraphLLM provide explanations for its predictions, and what evidence supports its ability to do so?", "answer": "Yes, GraphLLM can provide explanations for its predictions. In additional experiments on the Shortest Path task, GraphLLM (LLaMA-7B-2) was tasked with generating the optimal path as an explanation for the answer. Under a dataset division of 6000/2000/2000 graph instances for training/validation/test, GraphLLM output valid and optimal paths for 94.55% of the test graph instances, compared to 57.70% achieved by LoRA fine-tuning of LLaMA-7B-2. This provides preliminary evidence of GraphLLM's ability to articulate explanations for its predictions.", "category": "Experimental_Analysis"}, {"question": "How does the end-to-end policy-based GraphLLM compare to a single model for solving diverse tasks in LLMs?", "answer": "Leveraging a single model to address diverse tasks in LLMs has recognized strength. However, the paper's experimentation reveals that prompting open-source LLMs and GPT-3.5-turbo for these tasks yields unsatisfying accuracy, rendering it impractical to rely on a single model for solving multiple problems. Even fine-tuning the LLM for a specific downstream task fails to yield satisfactory performance. In contrast, GraphLLM exhibits the ability to effectively address individual downstream tasks. The additional experiments demonstrate that GraphLLM possesses a preliminary capacity to handle multiple tasks concurrently. For instance, joint training of GraphLLM on Substructure Counting and Maximum Triplet Sum results in an accuracy of 0.9946±0.0021 on Substructure Counting and 0.8472±0.0210 on Maximum Triplet Sum. Similarly, when training GraphLLM to count the number of C-C-O triangles and C-O-O triangles in Substructure Counting, it achieves simultaneous accuracies of 0.9925±0.0046 on C-C-O triangles and 0.9996±0.0001 on C-O-O triangles.", "category": "Method_Comparison"}, {"question": "How does GraphLLM's context length reduction compare to other methods such as Selective Context, and what are the potential trade-offs in terms of information loss and accuracy?", "answer": "The effectiveness of compression techniques such as Selective Context[1] in shortening prompt length and enhancing the efficiency of Graph2Text methods. However, it's important to highlight the substantial differences between Graph2Text and GraphLLM. GraphLLM achieves a 96.45% reduction in context length compared to the original context, and when half of the original tokens are filtered, the reduction is still 92.9%. In contrast, even when filtering 50% of the context in Graph2Text, the remaining 50% is much longer than GraphLLM's instruction. However, the filtering process may result in the removal of key information related to node details or the graph structure, potentially leading to additional accuracy decrease in the baseline.\n\n[1] Compressing Context to Enhance Inference Efficiency of Large Language Models (EMNLP 23) https://arxiv.org/abs/2310.06201", "category": "Method_Comparison"}, {"question": "What are the formal definitions of Q, K, and V in the context of prefix tuning, and how are they used in the transformer layers? Specifically, where are these tokens concatenated? How does this process relate to the graph learning scenario?", "answer": "In prefix tuning, $Q_l$, $K_l$, and $V_l$ denote the query, key, and value matrices generated by the input tokens of the $l$-th transformer layer. The input to the transformer consists solely of the 'instruction' and is independent of graph-related elements. Only $K_l$ and $V_l$ are prepended with the prefix tokens $P_l$, while $Q_l$ remains unaffected. The prepended $K_l^{\\prime}$ and $V_l^{\\prime}$, along with the original $Q_l$, are then fed into the attention mechanism. \nThe prefix tokens $P_l \\in \\mathbb{R}^{K \\times d^{\\mathrm{M}}}$ are prepended to the original keys and values $(K_l, V_l \\in \\mathbb{R}^{* \\times d^{\\mathrm{M}}})$ along the first dimension, resulting in $K_l^{\\prime}, V_l^{\\prime} \\in \\mathbb{R}^{(K+*) \\times d^{\\mathrm{M}}}$, where $K$ is the number of prefix tokens, $*$ denotes the length of the input instruction, and $d^{\\mathrm{M}}$ is the dimension of the LLM. \nThe relationship between prefix tuning and graph learning is explained in Section 3.4, 'Graph-enhanced Prefix Tuning for LLMs'. The prefix tokens are a combination of graph representations and learned prefix embeddings: $P = G W_{\\mathrm{U}} + B$, where $G$ is the graph representation from the graph learning module, $B$ is the learned prefix embedding, and $W_\\mathrm{U}$ is a matrix for dimension conversion. The relationship between prefix tuning and graph learning is further elaborated in Section 3.4, where the prefix tokens are described as a combination of graph representations and learned prefix embeddings.", "category": "Concept_Understanding"}, {"question": "What is the principal novelty of the paper in combining graph transformers and prefix tuning of LLMs, and what specific implementation challenges were addressed?", "answer": "The principal novelty lies in the idea of collaborating graph learning models and LLMs as a unified system, which is a first in this domain. Furthermore, even combining graph transformers with prefix tuning is non-trivial in its implementation. For example, the main implementation challenge addressed is how to make the graph transformer generate prefixes of specific dimensions. Initial attempts to apply dimensional transformation to the graph transformer’s output failed, leading to a design where the graph transformer’s input is derived from a query of the same dimension as the prefix, a proposal made for the first time.", "category": "Method_Disambiguation"}, {"question": "Why was P-tuning v2 not considered for comparison or integration in the study, given its perceived superior performance over Prefix Tuning?", "answer": "P-tuning v2 is often perceived as a general enhancement over Prefix Tuning. The P-Tuning v2 paper itself mentions that 'P-Tuning v2 is an implementation of [Prefix Tuning] optimized and adapted for NLU (Natural Language Understanding)'. Originally, Prefix Tuning was designed for NLG (Natural Language Generation) purposes, while P-Tuning v2 adjusted this approach for NLU, employing classification heads for sequence labeling tasks instead of verbalizers. The impact of LLMs largely stems from their generative capabilities, and in the current landscape (as of 2023), most LLM-based research focuses on formulating tasks as NLG challenges. The use of classification heads post-LLM, a key differentiator between P-tuning v2 and traditional Prefix Tuning, is not a common approach in this domain.", "category": "Method_Disambiguation"}, {"question": "What experiments or analyses were conducted to investigate the impact of different instructions on language model performance, and how were the instructions chosen?", "answer": "Although varying instructions can influence results, the magnitude of this influence may be limited. The authors conducted experiments on Maximum Triplet Sum and Shortest Path using GPT-4 few-shot CoT prompting with different instructions. For Maximum Triplet Sum, the accuracy remained at 93.33% under the new instruction, while for Shortest Path, the accuracy decreased from 86.67% to 83.33%. In the case of fine-tuning, such as through prefix-tuning, the learned prefixes serve as optimized prompts for downstream tasks, thereby mitigating the performance gap resulting from different instructions. For prompting, finding the optimal prompt remains an open question due to the impracticality of testing all possible prompts.", "category": "Experimental_Exposition"}, {"question": "Why were graph neural network-based methods not included as baselines in the experiments for enhancing graph reasoning capabilities of Large Language Models (LLMs)?", "answer": "Graph neural network-based methods are not a meaningful baseline for the issue explored in this paper. Firstly, the primary claim of this paper is to enhance the graph reasoning capabilities of Large Language Models (LLMs). To support this claim, there is no need to compare with methods outside of LLMs. Previous benchmarking work has also not considered graph neural networks as a type of baseline. The purpose of this paper has never been to achieve better performance with LLMs on certain graph reasoning tasks than graph neural networks/graph transformers; these tasks can be exactly solved with algorithms consisting of just a few lines of code. The paper also does not seek to use Large Language Models (LLMs) for the improvement of graph representation learning. Moreover, the graph reasoning task the paper addresses (refer to Definition 2.2) is intrinsically a natural language generation task, requiring input and output in human language. For example, in the substructure counting task, the desired output is a natural language statement like 'There is 1 [substructure name]'. If the task is not transformed or the graph neural network-based methods are not adapted, even achieving a comparison becomes impossible. Adapting graph neural network-based methods to generate such natural language outputs presents a significant challenge, diverging from the scope of the paper's investigation.", "category": "Method_Disambiguation"}, {"question": "Why were only basic graph reasoning tasks evaluated in the study—including substructure counting, maximum triplet sum, shortest path, and bipartite graph matching—instead of more complex tasks, such as multi-hop reasoning on large graphs?", "answer": "The tasks evaluated are similar in difficulty to those in a benchmark study[1] on LLMs’ graph reasoning abilities. The primary aim is to understand what hinders LLMs’ graph reasoning by starting with basic tasks. A natural thought is that, to investigate this issue, the paper should start with the most basic graph reasoning tasks. Surprisingly, current LLMs struggle even with these simple tasks. While addressing noise graphs and complex reasoning are important for further applications, the current focus is on the primary research question.\n\n[1] Can Language Models Solve Graph Problems in Natural Language? (NeurIPS 23) https://arxiv.org/abs/2305.10037", "category": "Motivation_Analysis"}, {"question": "Why does the paper not compare GraphLLM to established graph neural networks like GNNs, GCNs, and graph transformers, which are optimized for graph reasoning, to better evaluate its performance relative to the state-of-the-art in graph representation learning? For example, how does GraphLLM compare to a graph transformer without the attached LLM on the tested graph reasoning tasks? This could isolate the benefits of the LLM integration. Similarly, comparisons to GNN variants on standard benchmarks could reveal where GraphLLM excels or lags behind existing graph-specific architectures. The improvements over Graph2Text may simply be because LLMs struggle with learning from text descriptions of graphs. Graph models designed for relational reasoning may be more competitive. In essence, without comparing to specialized graph reasoning models, it is hard to discern if GraphLLM's integration of LLMs and graph networks is truly advancing state-of-the-art performance. Adding these comparisons would give a clearer sense of GraphLLM's capabilities relative to the field of graph representation learning overall.", "answer": "Graph neural network-based methods are not a meaningful baseline for the issue explored in this paper. Firstly, the paper’s primary aim is to enhance the graph reasoning abilities of LLMs, a goal that does not necessitate a comparison with non-LLM methods. Previous benchmarking work[1] has also not considered graph neural networks as a type of baseline. As corroborated by previous benchmarking work[1], graph neural networks have not been considered a standard baseline in this context. The paper’s focus is not on outperforming graph neural network-based methods in graph reasoning tasks, such as shortest path problems, which can be precisely solved with simple algorithms with a few lines of code. Instead, the pivotal question explored is: in the trend where generative LLMs are increasingly seen as a universal agent due to their powerful reasoning capabilities, how does the graph reasoning ability of LLMs perform, being one of the key reasoning skills for solving many problems like path planning, and what future developments could lie ahead? Considering this, the performance metrics of graph neural network-based methods do not contribute additional insights to the paper’s research question. Even if the paper were to attempt a comparison with graph neural networks, the graph reasoning task the paper addresses (as outlined in Definition 2.2) is intrinsically a natural language generation task, necessitating inputs and outputs in human language. For instance, in a substructure counting task, the expected output is a natural language statement like ‘There is 1 [substructure name]’. If the paper does not transform the task or adapt the graph neural network-based methods, even achieving a comparison becomes impossible. The essential question is how to make a fair comparison between generative LLMs that take natural language for input and output, and a graph neural network? Transforming natural language outputs into some kind of classification category seems a possible solution, but it should be noted that an inherent advantage of LLMs is their ability to interact and reason in human language, making such a comparison meaningless.\n\n[1] Can Language Models Solve Graph Problems in Natural Language? (NeurIPS 23) https://arxiv.org/abs/2305.10037\n\n", "category": "Method_Disambiguation"}, {"question": "Does prefix tuning require separate embeddings for each prefix value and key, resulting in a tensor of shape L x 2K x d?", "answer": "No, prefix tuning in the paper's implementation uses a single graph-enhanced prefix token $p \\in \\mathrm{R}^{d^\\mathrm{M}}$ that is linearly transformed by $W_V$ and $W_K$ to obtain the prefix value and key. This reduces the required tensor shape to $L \\times K \\times d^{\\mathrm{M}}$, saving trainable parameters while maintaining effectiveness.", "category": "Method_Disambiguation"}, {"question": "What substructure is consistently identified in task 1, and what additional substructures were tested in further experiments?", "answer": "In the paper, it is always the C-C-O triangles being identified. Additional experiments were conducted to have GraphLLM simultaneously identify C-C-O triangles as well as C-O-O triangles. After joint training, GraphLLM achieves accuracies of 0.9925±0.0046 on C-C-O triangles and 0.9996±0.0001 on C-O-O triangles.", "category": "Experimental_Exposition"}, {"question": "Why were graph Transformer or Graph Neural Networks (GNNs) not compared with the proposed method to evaluate whether Large Language Models (LLMs) have advanced the field?", "answer": "Graph neural network-based methods are not a meaningful baseline for the issue explored in this paper. Firstly, the paper's primary aim is to enhance the graph reasoning abilities of LLMs, a goal that does not necessitate a comparison with non-LLM methods. Previous benchmarking work[1] has also not considered graph neural networks as a type of baseline. As corroborated by previous benchmarking work[1], graph neural networks have not been considered a standard baseline in this context. The paper's focus is not on outperforming graph neural network-based methods in graph reasoning tasks, such as shortest path problems, which can be precisely solved with simple algorithms with a few lines of code. Instead, the pivotal question the paper aims to explore is: in the trend where generative LLMs are increasingly seen as a universal agent due to their powerful reasoning capabilities, how these LLMs fare in graph reasoning. Considering this, the performance metrics of graph neural network-based methods do not contribute additional insights to the paper's research question. Even if the paper were to attempt a comparison with graph neural networks, the graph reasoning task the paper addresses (as outlined in Definition 2.2) is intrinsically a natural language generation task, necessitating inputs and outputs in human language. For instance, in a substructure counting task, the output is expected to be a natural language statement like 'There is 1 [substructure name]'. If the paper does not transform the task or adapt the graph neural network-based methods, even achieving a comparison becomes impossible. The question is how to make a fair comparison between a generative LLMs that take natural language for input and output, and a graph neural network? Transforming natural language outputs into some kind of classification category seems a possible solution, but it should be noted that an inherent advantage of LLMs is their ability to interact and reason in human language, making such a comparison evidently meaningless.\n\n[1] Can Language Models Solve Graph Problems in Natural Language? (NeurIPS 23) https://arxiv.org/abs/2305.10037", "category": "Method_Disambiguation"}, {"question": "How does the necessity of end-to-end training and the limitation to open-sourced LLMs impact the applicability of the proposed method, especially in scenarios where fine-tuning is required?", "answer": "In scenarios where users need an LLM like LLaMA 2 7B to solve tasks such as shortest path problems, zero/few-shot prompting yields unsatisfactory results (Exact Match Accuracy: 0.1089). Fine-tuning is thus a necessary cost. While the graph2text approach incurs this cost without satisfactory outcomes, the paper's method demonstrates superior applicability. The limitation to open-sourced models is due to the closed-sourcing practice, not the paper's method, and using closed-source model APIs is often impractical due to data security and cost concerns.\n", "category": "Method_Disambiguation"}, {"question": "What is the principal novelty of the paper, considering that all three components are heavily based on the existing literature, i.e., node understanding is a standard Transformer, structure understanding is essentially equivalent to Ma et al., 2023[1], and prefix tuning follows the common practice of LLMs? It would strengthen the paper if these components could be customized to further enhance performance.\n\n[1] Liheng Ma, Chen Lin, Derek Lim, Adriana Romero-Soriano, Puneet Kumar Dokania, Mark Coates, Philip H. S. Torr, and Ser Nam Lim. Graph inductive biases in transformers without message passing. In International Conference on Machine Learning, 2023.\n", "answer": "The paper’s principal novelty lies in the collaboration of graph learning models and LLMs as a unified system, a first in this domain. Furthermore, even combining graph transformers with prefix tuning is non-trivial in the implementation. The paper has indeed made customizations. For example, the most critical issue is how to make the graph transformer generate prefixes of specific dimensions. The paper’s initial experimental attempts indicate that simply applying a dimension transformation to the output of the graph transformer does not work. Therefore, the graph transformer’s input is designed to be derived from a query of the same dimension as the prefix, a design the paper proposes for the first time.", "category": "Method_Disambiguation"}, {"question": "What experimental evidence supports the claim that graph Transformers decouple node information and structural information, and what is the key distinction between GAT and Graph Transformer in this context?", "answer": "The experimental evidence supporting the claim is provided in Appendix C.2, titled 'Ablation Study on Graph Transformer'. The key distinction is that GAT treats the graph structure as an inductive bias for aggregating node features, while Graph Transformer considers the graph structure as positional encodings.", "category": "Method_Comparison"}, {"question": "Is the same substructure (C-C-O triangles) always identified in task 1 as per the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the input to the LLM (after the prefix) consistently the same for all instances of the same task?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did additional experiments involve GraphLLM identifying both C-C-O and C-O-O triangles simultaneously?", "answer": "True", "category": "Claim_Verification"}, {"question": "After joint training, does GraphLLM achieve an accuracy of 0.9925±0.0046 on C-C-O triangles and 0.9996±0.0001 on C-O-O triangles?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "VP20ZB6DHL", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Chain-of-Verification Reduces Hallucination in Large Language Models", "authors": ["Shehzaad Dhuliawala", "Mojtaba Komeili", "Jing Xu", "Roberta Raileanu", "Xian Li", "Asli Celikyilmaz", "Jason E Weston"], "abstract": "Generation of plausible yet incorrect factual information, termed hallucination, is an unsolved issue in large language models. We study the ability of language models to deliberate on the responses they give in order to correct their mistakes. We develop the Chain-of-Verification (CoVe) method whereby the model first (i) drafts an initial response; then (ii) plans verification questions to fact-check its draft; (iii) answers those questions independently so the answers are not biased by other responses; and (iv) generates its final verified response. In experiments, we show CoVe decreases hallucinations across a variety of tasks, from list-based questions from Wikidata, closed book MultiSpanQA and longform text generation.", "keywords": "Large Language Model;Hallucination", "qa_pairs": [{"question": "How does CoVE's reduction in the number of factual outputs, as shown in Table 1 and Table 3, align with its goal of reducing hallucinations while minimizing the impact on truthful facts?", "answer": "CoVE reduces the number of facts as a consequence of reducing hallucinations. The paper's focus is on CoVE's ability to selectively remove fictitious facts with minimal impact on truthful facts. The FactScore metric, normalized by the number of facts, ensures that shorter generations do not automatically imply fewer hallucinations. Additionally, CoVE significantly outperforms a baseline that reduces facts by chopping off passage ends, demonstrating its effectiveness in hallucination mitigation.", "category": "Experimental_Analysis"}, {"question": "What baselines were added to compare the effectiveness of CoVe in mitigating hallucinations in longform generations, and how does CoVe perform against these baselines? Considering that there are many relevant works based on prompting, to reduce the hallucination of LLMs [1,2,3] or improve LLMs’ reasoning [4,5].", "answer": "Three new state-of-the-art baselines were added: SelfCheckGPT (NLI and LLM) and ChatProtect. CoVe outperforms both SelfCheckGPT (with NLI and LLM) and ChatProtect- three state-of-the-art papers that deal with hallucinations in longform generations.\n\n| Base Models        | Method                 | FACTSCORE  | Avg # facts |\n|--------------------|------------------------|------------|-------------|\n| GPT3               | Few-shot               | 45.3       | 15.6        |\n| GPT3 + ChatGPT     | ChatProtect            | 48.5       | 14.6        |\n| GPT3 + DeBERTA     | SelfCheckGPT-NLI       | 60.6       | 6.0         |\n| GPT3 + InstructGPT | SelfCheckGPT-LLM       | 61.7       | 6.3         |\n| GPT3 + InstructGPT | CoVe - Factor + revise | **68.6**       | 9.0         |", "category": "Method_Comparison"}, {"question": "Why are zero-shot results for LLaMA65B not included in the comparison, and how does this affect the comparison between LLaMA2 and LLaMA65B?", "answer": "LLaMA65B is not an instruction-tuned model and produces gibberish when prompted zero-shot, which is expected behavior for a base language model. When prompted zero-shot with ‘Let’s think step-by-step,’ the model generated erroneous steps—for example, suggesting to check information on the web instead of generating the required answer—and did not produce the required answers. In any case, these serve as baselines for CoVe rather than being interesting in themselves for comparison. The CoVe experiments also use few-shot prompting with LLaMA65B, hence the baseline few-shot LLaMA65B comparison. LLaMA 2 (which should be superior) is included as a baseline simply to show that zero-shot instruction-tuned model baselines do not actually fare better.", "category": "Experimental_Analysis"}, {"question": "Why are the results for the Factor+revise method only presented in Table 3 and not in Tables 1 and 2? Does this method perform differently on the setups in Tables 1 and 2?", "answer": "The Factor+revise method adds an additional step where the LLM can check for inconsistencies between the generated content and the answered verification question. Consider the example:\n\n'Trump received a BS in Business from the University of Pennsylvania' and the verification question and output:\n\nWhat degree did Donald Trump graduate with from the University of Pennsylvania in 1968? Answer: A BS in Economics.\n\nThe Factor+revise method adds a separate step that has the LLM check if these two outputs are consistent, inconsistent, or partially consistent and then recomposes the output based on the consistent parts. For the set QA tasks, checking for inconsistencies and removing an answer from the list is more trivial, as it is just an entity match. Example:\n\nName some politicians born in New York, NY A: Hillary Clinton\n\nWhere was Hillary Clinton born? A: Chicago (Chicago != New York)\n\nCoVe - 2 step is a variant of CoVe-Factored where the Verification questions are evaluated in one step (thus reducing the need to prompt the LLM several times for each verification question). This is feasible to use on tasks such as set QA but as the number of verification questions grows answering them all can be difficult in the fixed context size. The paper hence doesn't experiment with this approach on the harder benchmarks.", "category": "Experimental_Analysis"}, {"question": "How does the technical contribution of CoVE differ from the work in 'Measuring and Narrowing the Compositionality Gap in Language Models' (cited as [1]), particularly in terms of focus, methodology, and novelty?\n\n[1] https://arxiv.org/abs/2210.03350", "answer": "The paper 'Measuring and Narrowing the Compositionality Gap in Language Models' (which is cited) as the title suggests concentrates on compositional questions with multiple hops such as 'Who won the Master’s Tournament the year Justin Bieber was born?' – and breaking into subquestions. It does not focus on hallucinations (does not mention that word at all). It uses a method called self-ask to build this subquestion decomposition. In contrast, CoVE uses questions to verify hallucinations rather than decompose multi-hop questions. Additionally, CoVE introduces and compares joint, 2-step, factored, and factor+revise approaches, which are not addressed in the cited paper.", "category": "Method_Comparison"}, {"question": "What is the significance of the Wikipedia QA datasets (Wikidata and Wiki-category list) in the paper, given that they are created from simple templates and appear toyish?", "answer": "The Wiki-Category benchmark is a subset of an existing benchmark called Quest[1], and Wikidata uses real data from Wikipedia, though the queries are synthetically constructed. Additionally, the paper reports results on the MultiSpanQA task, which uses real queries from humans. List-form answers are natural in assistant-based applications and are useful for understanding LLM hallucinations and testing approaches efficiently. List generations mimic longform generations, with the advantage that they can be automatically, efficiently, and perfectly evaluated, whereas in other tasks evaluation is much more tricky. While questions like 'Name some politicians born in Boston' may seem trivial, commercially deployed LLM systems often fail to produce correct answers.\n\n\n[1] https://aclanthology.org/2023.acl-long.784.pdf", "category": "Method_Disambiguation"}, {"question": "Why was biography generation chosen as the only generation task for addressing hallucinations, given that it is less challenging than most real-world applications?", "answer": "Biography generation was chosen because (i) it is known to produce hallucinations for state-of-the-art models, as shown in experiments with both open source and commercial products; and (ii) evaluation methods, particularly the FactScore metric, were recently developed for the biography task, which is why several other papers also focus on biographies (see new baselines). While exploring further benchmark tasks, such as those in scientific, legal, and medical domains, would be interesting, these applications require deeper analysis.\n", "category": "Motivation_Analysis"}, {"question": "How does the FactScore metric account for the length of generated outputs, and what evidence supports the robustness of CoVe against shorter generations with fewer facts?", "answer": "The FactScore metric is normalized by the number of facts, so shorter outputs do not automatically result in fewer hallucinations or a higher FactScore. Additionally, a baseline was created by reducing the number of facts in passages to match the best model, and CoVe significantly outperformed this baseline, demonstrating its effectiveness in hallucination mitigation.", "category": "Experimental_Analysis"}, {"question": "Does the model require external information, such as retrieval, to identify factual errors in its own outputs, and what are the limitations of relying on retrieval for this purpose? Considering a key assumption behind CoVE is that even when the model is not able to generate factual outputs, it might still be able to identify nonfactual statements. A similar observation is mentioned by [1], but they argue that retrieval is necessary.\n\n[1] https://www.semanticscholar.org/paper/Language-Models-Hallucinate%2C-but-May-Excel-at-Fact-Guan-Dodge/45653ad43124f02dc2cf2db3357be1d1d78ddb18?utm_content=title&utm_medium=unfurl&utm_source=slackbot", "answer": "The paper's work and prior studies such as SelfCheckGPT and ChatProtect demonstrate that mitigating hallucinations is possible without retrieval. While guaranteeing veracity requires verified information sources, retrieval-based approaches like RAG also cannot ensure accuracy without assuming all internet data is true. Additionally, errors can still occur in the RAG pipeline, as evidenced by the PerplexityAI result in the paper.", "category": "Method_Disambiguation"}, {"question": "How is the performance of the model in plan verification, execution verification, and verified response measured, particularly in few-shot and zero-shot CoVe settings?", "answer": "Measuring the performance of intermediate parts of the chain is challenging due to the absence of labels, but end-to-end performance can be measured, or data can be hand-annotated. For the Wikidata task, the accuracy of verification questions is manually measured, and Llama 65B achieves around 70% accuracy in answering factoid questions related to professions and birthplaces, as described in the paper under “Shortform verification questions.”", "category": "Experimental_Setup"}, {"question": "What is the time complexity of the proposed CoVe method compared to other hallucination mitigation approaches?", "answer": "CoVe requires 4 types of LLM prompt calls corresponding to its 4 steps: (1) Generate baseline, (2) Plan verifications, (3) Execute verifications, and (4) Final response. Step 3 involves multiple LLM calls, but these can be executed in parallel using batching. A comparative analysis shows that CoVe has a comparable inference overhead to other hallucination mitigation approaches. Further details are provided in Section 8 of the appendix.", "category": "Method_Comparison"}, {"question": "How does the computational cost of the proposed CoVe method compare to other existing methods aimed at reducing LLM hallucinations?", "answer": "CoVe essentially requires 4 types of LLM prompt calls, corresponding to the 4 steps of the method (1. Generate baseline, 2. Plan verifications, 3. Execute verifications, 4. Final response). Step 3 in the factored variant of the approach is actually several LLM calls, one for each planned verification question, however, they can all be called in parallel and hence use batching. A comparative analysis shows that CoVe has a comparable inference overhead to other hallucination mitigation approaches. Further details are provided in Section 8 of the appendix.\n\n| Base Models        | Method                 | Worst case LLM calls                      |\n|--------------------|------------------------|-------------------------------------------|\n| GPT3               | Few-shot               | 1                                         |\n| GPT3 + ChatGPT     | ChatProtect            | 1 + num_sents x repeats x 3               |\n| GPT3 + InstructGPT | SelfCheckGPT-LLM       | num_samples + 1 + num_sents x num_samples |\n| GPT3 + InstructGPT | CoVe - Factor + revise | 1 + num_sents + 2 x num_facts             |", "category": "Method_Comparison"}, {"question": "What criteria were used to select few-shot examples for the list-based QA benchmarks and the biographies dataset, and how might using different examples affect the results?", "answer": "For the list-based QA benchmarks, few-shot prompts were randomly selected from the train set and kept constant across variants. For the biographies dataset, prompts were crafted based on biographies of the most viewed people on Wikipedia. Few-shot examples were used to show the desired verification responses. Changing the few-shot prompts can lead to variance in the final output, so the same prompts were used for all baseline approaches in the comparative analysis.", "category": "Experimental_Setup"}, {"question": "What is the estimated additional computational overhead introduced by the proposed CoVe method compared to other hallucination mitigation methods?", "answer": "CoVe essentially requires 4 types of LLM prompt calls, corresponding to the 4 steps of the method (1. Generate baseline, 2. Plan verifications, 3. Execute verifications, 4. Final response). Step 3 in the factored variant of the approach is actually several LLM calls, one for each planned verification question, however, they can all be called in parallel and hence use batching. The analysis shows that CoVe has a comparable inference overhead to other hallucination mitigation approaches. Further details are provided in Section 8 of the appendix.", "category": "Method_Comparison"}]}
{"id": "tG5mpAM7ZK", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Extending to New Domains without Visual and Textual Oracles", "authors": ["Daiqing Qi", "Handong Zhao", "Aidong Zhang", "Sheng Li"], "abstract": "To avoid the high cost of collecting visual data from all test domains in domain adaption task, recent work takes advantage of the pre-trained large-scale vision language models and augment training data with only text descriptions (e.g.,“a photo/painting/sketch...”) of each test domain. However, in many real-world ap- plications, such text information of test domains is not always available in ad- vance. Moreover, even if we can verbalize all test domains, it is laborious for existing work (Dunlap et al., 2023) to train a different augmentation network for each possible unseen domain. To overcome these challenges, we benefit from the multimodal embedding space of a pre-trained vision-language model and propose to acquire training-free and domain-invariant augmentations with text descrip- tions of arbitrary crafted unseen domains, which not necessarily match test do- mains. Beyond achieving state-of-the-art results, compared with existing works that require trainable augmentation networks, our approach is also notably more time-efficient, and exhibits a more solid theoretical support.", "keywords": "multimodal learning;vision-language model;domain adaptation;domain generalization", "qa_pairs": [{"question": "Do the CLIP features consistently satisfy the conditions of equation 4, and can empirical analysis provide insights into potential failure cases?", "answer": "It can be illustrated with an example, where the source domain is 'painting' and target domain is 'photo'. Given an image-text pair from source data, say, (apple_painting.jpg, 'a painting of an apple'), and a text description from target domain: 'a photo of an apple', intuitively, equation 4 means: in the CLIP embedding space, 'a painting of an apple' is closer to 'a photo of an apple' than to apple_painting.jpg. In other words, embeddings of the **same** modality should be closer to each other than embeddings of the same semantic meaning but different modality. This is what the paper expect. Liang et al.[1] reveal that, when a vision-language model is pre-trained with a contrastive loss, the embeddings are clustered per modality, which naturally guarantees what the paper expect. \n\n**And in practice, this always holds.** It is reasonable because the CLIP model is well-trained with 400M data, during which time the embeddings are very well clustered per modality. Refer to Fig. 1b [1]. Intuitively, if there is a failure case, one could imagine a blue point appearing among the red points, which indicates a text embedding is closer to one or some image embeddings than to other text embeddings. In the paper's practice with the pre-trained CLIP model, it never happens. Liang et al. further explain why they are so separated from each other from three aspects: cone effect, initializations, and contrastive learning objective.\n\n[1] Liang et al., Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS 2022.", "category": "Experimental_Analysis"}, {"question": "What empirical analysis was conducted to verify the assumptions in Appendix A regarding the modality gap in CLIP embeddings, and how do the results support the validity of these assumptions?", "answer": "This section introduces the following modality gap geometry metrics. The distance between the modality clusters in the CLIP embedding space is referred to as the modality gap. The instance-level modality gap $g$ is defined as the difference between image and text embeddings for a single pair. The class-level modality gap is defined as $g_{c}$, where for a given class $c$, the expectation of $g$ is calculated over all instances of class $c$ across different domains. Finally, the domain-level modality gap $g_{d}$ is calculated, where given a specific domain, the expectation of $g$ is calculated over all instances of domain $d$ across different classes. Formally, given an instance pair ($x$, $y$), $g = x - y$, where $x$ and $y$ are CLIP features of an image and a text. There are also: $g_{c} = x_c - y_c$, where $x_c = E_d[ x_c^d ]$ and $y_c = E_d[ y_c^d ]$. Note that $E_d[ x_c^d ]$ means the expectation of all instances of a given class $c$ across all domains. And $g_{d} = x_d - y_d$, where $x_d = E_c[ x_d^c ]$ and $y_d = E_c[ y_d^c ]$. Note that $E_c[ x_d^c ]$ means the expectation of all instances of a given domain $d$ across all classes. It is hoped that the modality gap $g_{c} $ grants the assumptions presented in Appendix A. Intuitively, if satisfied, given data from one source domain, training-free augmentation can be performed as $g_{c} $ is valid across different domains. More specifically, the two assumptions are: \n\n1. *The modality gap between corresponding image and text embeddings can be approximated by a constant vector, particularly given a class across different domains. This is verified by computing distributions over ||$g$|| (magnitude) and $cos(g, E_g[g])$ (direction).* \n\n2. *The modality gap is approximately orthogonal to the span of image embeddings and text embeddings, and image embeddings and text embeddings have a nearly zero mean in the subspace orthogonal to the modality gap. This is verified by computing distributions over $cos(x - E_x[x], E_g[g])$ (orthogonality) and $E_x[x - x^T g’g’]_i$ (center), where $g’ = E_g[g] / || E_g[g] ||$ and $i \\in [d]$. The subscript $i$ denotes indexing the $i$-th dimension of the vector.* The results are listed:\n\n | VL Model - Dataset   | Magnitude - *Instance* | Magnitude - *Class* | Magnitude - *Domain* | Direction - *Instance* | Direction - *Class* | Direction - *Domain* | Orthogonality     | Center           | \n| -------------------- | ---------------------- | ------------------- | -------------------- | ---------------------- | ------------------- | -------------------- | ----------------- | ---------------- | \n| CLIP - DomainNet     | 1.32 $\\pm$ 0.034       | 1.19 $\\pm$ 0.027    | 1.32 $\\pm$ 0.029     | 0.76 $\\pm$ 0.051       | 0.89 $\\pm$ 0.023    | 0.77 $\\pm$ 0.049     | 0.01 $\\pm$  0.107 | 0.00 $\\pm$ 0.024 |\n | CLIP - CUB Paintings | 1.29 $\\pm$ 0.035       | 1.18 $\\pm$ 0.029    | 1.30 $\\pm$ 0.038     | 0.70 $\\pm$ 0.051       | 0.88 $\\pm$ 0.029    | 0.74 $\\pm$ 0.045     | 0.00 $\\pm$  0.095 | 0.00 $\\pm$ 0.026 | \n\nIt can be observed that the orthogonality and center are close to 0 with small standard deviations, which means assumption 2 is well satisfied. For magnitude and direction, the class-level modality gap (**across domains**) better satisfies assumption 1, since the paper achieves higher scores for direction on both datasets. It meets the expectation that the class-level modality gap $g_{c}$ better satisfies the assumptions. It indicates that, given data from one source domain, the paper's training-free augmentation can be performed because the modality gap $g_{c}$ of a given class $c$ is valid across different domains.\n", "category": "Experimental_Analysis"}, {"question": "What does the symbol '1x' for Time in the tables represent?", "answer": "'$n$ x' indicates that the stage-2 training time of a particular model is $n$ times that of the LADS baseline. In other words, the training time of LADS is considered as the reference unit time. For instance, in the following example:\n| Model         | Average | ID    | OOD   | Training-free (Stage-1) | Time (Stage-2) |\n| ------------- | ------- | ----- | ----- | ----------------------- | -------------- |\n| LADS          | 74.99   | 85.33 | 64.85 | ✗                       | 1x             |\n| TEAM-*invar.* | 74.67   | 85.21 | 64.12 | ✓                       | 0.23x          |\n\nWhen the model is TEAM-*invar.*, $0.23$x is used because its training time is 0.23 times that of LADS. Therefore, when the model is LADS, $1$x is used because its training time is the same as that of itself.\n", "category": "Concept_Understanding"}, {"question": "How is it possible to directly apply the linear probe trained on domain-invariant and original features to the target feature without any augmentation?", "answer": "The questions are in fact standard pipelines for data augmentation based domain generalization methods [1,2,3], which are also adopted by LADS [3]. As the paper's TEAM model is also a data augmentation based method, the paper simply follows existing works. Recap the paper's methodology first: The paper has two model versions: TEAM-full and TEAM-invariant. In TEAM-full, $k$ crafted domain names are provided. The paper utilizes a training-free method to obtain augmented features for $k$ provided domains, which are then mixed with the source domain for training on (k+1) domains during stage-2. The evaluation is conducted directly on test domains images without prior image augmentation to a specific domain. This follows the standard approach in data augmentation based domain generalization (DG) methods [1,2,3], which LADS also follows: The trained model is applied to test domain data without any augmentation. In the 'invariant' version, augmentation is not performed for $k$ domains. Instead, augmentation is performed for a single 'invariant' domain. The key is, it is still a data augmentation method, which means the obtained invariant embeddings are still 'augmented' data, just like any augmented data for $k$ domains in TEAM-full. The difference from TEAM-full is that, here the paper sets $k$ = 1 (i.e., the so-called 'invariant domain'). Therefore, the training and evaluation process are the same as in TEAM-full and LADS.\n\n\n[1] Li et al., A Simple Feature Augmentation for Domain Generalization, ICCV 2021\n\n[2] Vidit, Vidit, Martin Engilberge, and Mathieu Salzmann. \"CLIP the Gap: A Single Domain Generalization Approach for Object Detection.\" CVPR, 2023.\n\n[3] Lisa et al., Using Language to Extend to Unseen Domains, ICLR 2023", "category": "Method_Mechanics"}, {"question": "Why does the TEAM model mix augmented data with original data for training in Stage 2 instead of using only augmented embeddings?", "answer": "The questions are in fact standard pipelines for data augmentation based domain generalization methods [1,2,3], which are also adopted by LADS [3]. As the paper's TEAM model is also a data augmentation based method, the paper simply follows existing works. Recap the paper's methodology first: The paper has two model versions: TEAM-full and TEAM-invariant. In TEAM-full, $k$ crafted domain names are provided. The paper utilizes a training-free method to obtain augmented features for $k$ provided domains, which are then mixed with the source domain for training on $(k+1)$ domains during stage-2. The evaluation is conducted directly on test domain images without prior image augmentation to a specific domain. This follows the standard approach in data augmentation based domain generalization (DG) methods [1,2,3], which LADS also follows: Augmented data is mixed with original data for training. \n\nIn addition, the authors have done ablation studies on DomainNet: the paper's model istrained on DomainNet with: ① a mix of augmented data and original data. ② augmented data only. ③ original data only.\n\n | Model                            | Average   | ID        | OOD       | \n| -------------------------------- | --------- | --------- | --------- |\n| ① original data + augmented data | **96.18** | **95.61** | **96.70** | \n| ② augmented data only            | 95.60     | 95.21     | 95.99     | \n| ③ original data only             | 94.39     | 95.03     | 93.75     |\n\n[1] Li et al., A Simple Feature Augmentation for Domain Generalization, ICCV 2021\n\n[2] Vidit, Vidit, Martin Engilberge, and Mathieu Salzmann. \"CLIP the Gap: A Single Domain Generalization Approach for Object Detection.\" CVPR, 2023.\n\n[3] Lisa et al., Using Language to Extend to Unseen Domains, ICLR 2023.", "category": "Method_Disambiguation"}, {"question": "What is the significance of exploring the conditions under which a valid solution for $z$ exists in the normalized embedding space, and how does it relate to the pre-trained vision-language (VL) model?", "answer": "Obtaining a solution for $z$ is not hard mathematically. Actually, the paper's intention is not to showcase the paper's skills of solving the equations, but rather to explore the conditions under which a valid solution is available in the normalized embedding space (i.e., training-free feasibility) (Eq. 17 in Appendix B), and its relationship with the pre-trained vision-language (VL) model (Lemma 1). The paper is more interested in $\\lambda_{1}$ in Eq. 17 (Appendix B), which determines if the solved $z$ is a valid solution in the embedding space. The exploration is critical because it theoretically demonstrates that, if aligning with the global direction (Eq. 21), the condition for the existence of a valid $z$ is more stringent, requiring the cosine similarity of any negative image-text pair is always less than that of the positive pair. This implies that the pre-trained VL model should be perfect (Appendix C). On the other hand, if aligning with the paper's modality direction, the condition for the existence of $z$ is more relaxed: once the embeddings are clustered per modality in the VL model’s embedding space, a solution is readily available. Fortunately, (Liang et al.) [1] validate that the condition is generally satisfied with a VL model trained with a contrastive loss. \n\nThe paper's ablation studies in Tab. 3 also validate this point: aligning modality direction proves to be more effective: \n\n| Augmentation Method                 | Average   | ID        | OOD       |\n| ----------------------------------- | --------- | --------- | --------- | \n| LADS                                | 74.99     | 85.33     | 64.85     |\n| Training-free + Align Global Dir.   | 74.67     | 85.21     | 64.12     | \n| Training-free + Align Modality Dir. | **77.16** | **86.61** | **67.71** |\n\n[1] Liang et al., Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS 2022", "category": "Method_Disambiguation"}, {"question": "Why does the performance of the proposed method on CUB and DomainNet in Table 1 and Table 2 not show a significant margin over the baselines?", "answer": "Table 1 shows results under the LADS setting with exact test domain names, while Table 2 presents main results in a text-driven domain generalization setting without exact test domain names. In this challenging setting, competing CLIP-based baselines like WiSE-LP and LADS have close performance, making further improvements difficult. However, the paper's method distinguishes itself in AVG and OOD scores in Table 2:\n\n| Method  | OOD (CUB-Paintings) | OOD (DomainNet) | Average (CUB-Paintings) | Average (DomainNet) |\n| ------- | ------------------- | --------------- | ----------------------- | ------------------- | \n| WiSE-LP | 64.80               | 93.68           | 73.27                   | 94.44               | \n| LADS    | 64.85               | 94.65           | 74.99                   | 94.97               | \n| The paper's    | **67.71**           | **96.70**       | **77.16**               | **96.18**           | \n\nThis is achieved without parametric augmentation or training in the first phase, contributing to time efficiency, which is a crucial aspect beyond accuracy. The paper's focus is on out-of-domain (OOD) performance, a major metric for domain generalization tasks, like existing works [1,2,3]. To summarize, the paper's model achieves obvious results on the paper's practical setting (Tab. 2) over computing baselines (their performances are very close to each other while the paper achieves a more obvious improvement) without training during stage-1 and less time during stage-2 *with “invar.” mode.\n\n[1] Li et al., A Simple Feature Augmentation for Domain Generalization, ICCV 2021\n\n[2] Vidit, Vidit, Martin Engilberge, and Mathieu Salzmann. \"CLIP the Gap: A Single Domain Generalization Approach for Object Detection.\" CVPR, 2023.\n\n[3] Lisa et al., Using Language to Extend to Unseen Domains, ICLR 2023", "category": "Experimental_Analysis"}, {"question": "What specific ablation studies have been conducted to demonstrate the effectiveness of the augmentation method and the robustness across different sets of prompts?", "answer": "The ablation studies include: (1) augmentation modes and (2) invariant modes in Tab. 3, choices of backbone vision-language models in Tab. 11, and different sources of prompt templates in Tab. 10 and Tab. 11. These studies use pre-defined templates such as ImageNet Templates and prompts from different LLMs (e.g., ChatGPT, CPT4, New Bing, and Bard) to demonstrate robustness across different sets of prompts. Additionally, a new table was added to show the model's performance on DomainNet with a mix of augmented data and original data, augmented data only, and original data only.\n\n| Model                            | Average   | ID        | OOD       | \n| -------------------------------- | --------- | --------- | --------- | \n| ① original data + augmented data | **96.18** | **95.61** | **96.70** | \n| ② augmented data only            | 95.60     | 95.21     | 95.99     | \n| ③ original data only             | 94.39     | 95.03     | 93.75     |", "category": "Experimental_Analysis"}, {"question": "What are the key differences between the proposed method (TEAM) and CLIP the GAP (CGAP)[1] in terms of augmentation strategies, methods, domain prompts, training stage data, and tasks?\n\n[1] Vidit, Vidit, Martin Engilberge, and Mathieu Salzmann. \"CLIP the Gap: A Single Domain Generalization Approach for Object Detection.\" CVPR, 2023. (https://openaccess.thecvf.com/content/CVPR2023/papers/Vidit_CLIP_the_Gap_A_Single_Domain_Generalization_Approach_for_Object_CVPR_2023_paper.pdf)", "answer": "Although the paper may share similarity in working under domain generalization with CGAP, the paper is fundamentally different in methods and tasks. In fact, CGAP is mostly similar to LADS rather than to the paper's model. The paper only shares the similarity with CGAP on using domain prompts to perform augmentation. But the key is how to use them, i.e., augmentation strategies and methods. It significantly differs from CGAP in this sense, which are the major contributions of this paper, leading to consistently better performance on both LADS settings and the paper's setting with much less time cost. The major differences are summarized in the table below:\n\n| Model       | Augmentation Strategy (Stage-1)                        | Augmentation Method (Stage-1)                                | Domain Prompt                                                | Training Stage Data (Stage-2)             | Task             |\n| ----------- | ------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | ----------------------------------------- | ---------------- |\n| CGAP        | Train multiple augmentation functions                  | Align global direction with trainable augmentation functions | A given pre-defined list                                     | Augmented and original features           | Object Detection |\n| LADS        | Train multiple augmentation functions                  | Align global direction with trainable augmentation functions | A given pre-defined list (exactly match test domains)       | Augmented and original features           | Classification   |\n| TEAM (the paper's) | *Training-free (calculated with Eq. 9, nonparametric)* | *Align modality direction with nonparametric training-free method* | *(1) Domain-invariant representations (2) a given pre-defined list* | *Domain-invariant and original features* | *Classification* |\n\nIn fact, CGAP is mostly similar to LADS in most columns, except for the 'task,' while it is different from the paper in all above aspects. Specifically, CGAP still follows the pipeline of LADS: they first perform optimization for trainable augmentations {$A_j$} (Eq. 3) during the first stage, which are later used for augmenting source images during the training stage. And they all align 'global directions' during the training of augmentation functions.", "category": "Method_Comparison"}, {"question": "What are the fundamental differences between the observations in this paper and those in CLIP the GAP (CGAP)[1] regarding the alignment of domains of the same object?\n\n[1] Vidit, Vidit, Martin Engilberge, and Mathieu Salzmann. \"CLIP the Gap: A Single Domain Generalization Approach for Object Detection.\" CVPR, 2023.(https://openaccess.thecvf.com/content/CVPR2023/papers/Vidit_CLIP_the_Gap_A_Single_Domain_Generalization_Approach_for_Object_CVPR_2023_paper.pdf)", "answer": "The fundamental difference lies in the alignment approach: CGAP aligns the **global direction** (Eq. 21), following methods like LADS, Styleclip[2], and Stylegan-nada[3], while this paper proposes to align the **modality direction** (Eq. 3), motivated by Liang et al.[4] This difference, though seemingly minor, leads to notable improvements, including more solid theoretical support, better preservation of class information, and milder assumptions for an analytical solution. These benefits are validated by ablation studies (Tab. 3) and further enhanced by the domain-invariant augmentation proposed in Section 4.2 and Fig. 3, which produces domain-invariant features and further reduces the training time (Stage-2 time in Tab.2).\n\n| Augmentation Method                 | Average   | ID        | OOD       |\n| ----------------------------------- | --------- | --------- | --------- |\n| LADS                                | 74.99     | 85.33     | 64.85     |\n| Training-free + Align Global Dir.   | 74.67     | 85.21     | 64.12     |\n| Training-free + Align Modality Dir. | **77.16** | **86.61** | **67.71** |\n\n[1] Patashnik et al., Styleclip: Text-driven manipulation of stylegan imagery, ICCV 2021\n\n[2] Gal et al., “Stylegan-nada: Clip-guided domain adaptation of image generators,” ACM Transactions on Graphics (TOG)\n\n[3] Liang et al., Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS 2022", "category": "Method_Comparison"}, {"question": "What is the justification for the significance of the proposed method given the marginal improvement over the baseline LADS in Table 1?", "answer": "Tab. 1 shows the results under the LADS setting, where exact test domain names are given, and LADS is expected to perform well. Tab. 2 presents the paper's setting, which is the more interesting part, as exact test domain names are not provided. Without exact test domains, the performance gap between LADS and the paper becomes more pronounced: **2.86% (OOD)** and **2.17% (AVG)** in CUB-Paintings, and **2.15% (OOD)** and **2.05% (AVG)** in DomainNet. Note that the paper achieves better results without any augmentation networks or training during stage-1, whereas LADS requires training an augmentation network for each given domain prompt. The paper's method is significantly more efficient. Additionally, the paper's method requires less time in stage-2 training (four times faster with 'invar.' mode in Tab. 2). The paper's model is also more robust across different sets of domain prompts (Appendix E and F provide more experiments and details), while LADS only performs well when the exact test domain names are given, which is impractical. As mentioned in the paper, out-of-domain (OOD) performance is the primary metric for the paper, as it focuses on the domain generalization task, where test domain accuracy is the major metric, similar to existing works [1,2,3]. To summarize, the paper's model achieves significant improvements in the paper's practical setting (Tab. 2) over computing baselines (their performances are very close to each other, while the paper achieves a more obvious improvement) without training during stage-1 and less time during stage-2 with 'invar.' mode.\n\n| Method  | OOD (CUB-Paintings) | OOD (DomainNet) | Average (CUB-Paintings) | Average (DomainNet) |\n| ------- | ------------------- | --------------- | ----------------------- | ------------------- |\n| WiSE-LP | 64.80               | 93.68           | 73.27                   | 94.44               |\n| LADS    | 64.85               | 94.65           | 74.99                   | 94.97               |\n| The paper's    | **67.71**           | **96.70**       | **77.16**               | **96.18**           |\n\nEfficiency is also one of the paper's key improvements.\n\n[1] Vidit, Vidit, Martin Engilberge, and Mathieu Salzmann. \"CLIP the Gap: A Single Domain Generalization Approach for Object Detection.\" CVPR, 2023.\n\n[2] Li et al., A Simple Feature Augmentation for Domain Generalization, ICCV 2021\n\n[3] Lisa et al., Using Language to Extend to Unseen Domains, ICLR 2023", "category": "Experimental_Analysis"}, {"question": "Do the CLIP features always satisfy the conditions of equation 4 in the empirical analysis conducted by the authors?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "o4AydSd3Lp", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Harnessing Discrete Representations for Continual Reinforcement Learning", "authors": ["Edan Jacob Meyer", "Adam White", "Marlos C. Machado"], "abstract": "Reinforcement learning (RL) agents make decisions using nothing but observations from the environment, and consequently, heavily rely on the representations of those observations. Though some recent breakthroughs have used vector-based categorical representations of observations, often referred to as discrete representations, there is little work explicitly assessing the significance of such a choice. In this work, we provide a thorough empirical investigation of the advantages of representing observations as vectors of categorical values within the context of reinforcement learning. We perform evaluations on world-model learning, model-free RL, and ultimately continual RL problems, where the benefits best align with the needs of the problem setting. We find that, when compared to traditional continuous representations, world models learned over discrete representations accurately model more of the world with less capacity, and that agents trained with discrete representations learn better policies with less data. In the context of continual RL, these benefits translate into faster adapting agents. Additionally, our analysis suggests that the observed performance improvements can be attributed to the information contained within the latent vectors and potentially the encoding of the discrete representation itself.", "keywords": "reinforcement learning;continual reinforcement learning;discrete representations;representation learning", "qa_pairs": [{"question": "What are the key conclusions and insights from the experiments in Section 3.3.3, and how are they actionable in a general sense?", "answer": "The key conclusion from Section 3.3.3 is that the importance lies in the sparse and element-wise binary nature of the one-hot encoding scheme for discrete latents, rather than the 'discreteness' itself. This insight highlights the importance of designing representations with appropriate sparsity for improved performance in continual learning tasks, making it actionable in a general sense.", "category": "Experimental_Analysis"}, {"question": "Can the results of using discrete representations, as demonstrated in the Minigrid testbed, be extended to more visually complex environments?", "answer": "The paper's decision to use Minigrid as a testbed was based on its goal to perform a more extensive empirical evaluation surrounding the successful use of discrete representations, and for that, it needed to be able to run large amounts of experiments and have a high level of control over its environment. Using Minigrid allowed the paper to go beyond reporting only performance curves but also to compare the state distributions induced by the different models (Figure 2), to vary the tasks in a way that it could simulate a continual learning problem (Figure 6), and even visualize distributions over rolled out trajectories (Figures 12 & 13). VQ-VAEs have been used to successfully compress and generate complex images in a number of works since the first paper on the architecture in 2017 (van den Oord et al.[1]). The authors of the original VQ-VAE paper even provide results to show how a VQ-VAE can reconstruct observations and predict believable, multi-step rollouts in DMLab specifically (Figures 5 & 7). The work since then has only gotten more impressive, with follow-up work using VQ-VAEs to generate high-resolution images (e.g., Razavi et al., 2019[2], Esser et al., 2021[3]) and videos (e.g., Yan et al., 2021[4], Walker et al., 2021[5]). The complexity of these datasets surpasses what is generally used in RL and gives high confidence that discretization works very well even in complex domains. Additionally, Hafner et al. (2020)[6] perform a direct comparison between continuous and discrete representations on over 50 Atari games (Figure 5) and find that discrete representations perform much better. The paper's setup is even similar to DreamerV2 in that both works use autoencoders to learn representations, and DreamerV2 uses a special case of a VQ-VAE (where the codebook is one-hot).\n\n[1]https://proceedings.neurips.cc/paper_files/paper/2017/file/7a98af17e63a0ac09ce2e96d03992fbc-Paper.pdf\n\n[2] https://proceedings.neurips.cc/paper/2019/file/5f8e2fa1718d1bbcadf1cd9c7a54fb8c-Paper.pdf\n\n[3]https://arxiv.org/abs/2012.09841\n\n[4]https://arxiv.org/abs/2104.10157\n\n[5]https://arxiv.org/abs/2103.01950\n\n[6]https://arxiv.org/abs/2010.02193\n", "category": "Experimental_Analysis"}, {"question": "Why do the methods perform similarly in section 3.3.2, and how does this relate to the simplicity of the environment?", "answer": "The methods perform similarly in section 3.3.2 because, in a simple environment, all representation methods achieve near-optimal performance. However, as the environment becomes more complex, discrete representations excel due to their ability to learn more with less modeling capacity. This motivates their use in continual reinforcement learning, as continual reinforcement learning is all about how to maximize reward when the agent cannot perfectly model the world (because the world is continually changing from the agent’s perspective).", "category": "Experimental_Analysis"}, {"question": "Given that Minigrid is inherently discrete and unlike most real-world problems, how can the results of the experiments be generalized to more complex domains?", "answer": "The use of Minigrid allowed for rigorous experiments that would not have been possible in more complex environments, such as visualizing distributions over rollouts. Several related works, including Hafner et al. (2020)[1], Robine (2020)[2], Micheli et al. (2023)[3], and Hafner et al. (2023)[4], have demonstrated strong results in more complex domains. Specifically, Hafner et al. (2020)[1] compared continuous and discrete representations over 50 Atari games and found that discrete representations performed much better. Additionally, the experiments in Section 4 show that the results in the episodic RL setting translate well into the continual RL setting, where the results are even better. \n\n[1]https://arxiv.org/abs/2010.02193\n\n[2]https://arxiv.org/abs/2010.05767\n\n[3]Vincent Micheli, Eloi Alonso, and Franc¸ois Fleuret. Transformers are sample-efficient world models. In International Conference on Learning Representations (ICLR), 2023.\n\n[4]https://arxiv.org/abs/2301.04104", "category": "Experimental_Analysis"}, {"question": "Why was Minigrid chosen as the testbed for the experiments, and how do the results from Minigrid generalize to more complex, real-world scenarios?", "answer": "The paper's decision to use Minigrid as a testbed was based on the goal to perform a more extensive empirical evaluation surrounding the successful use of discrete representations, and for that, it needed to be able to run large amounts of experiments and have a high level of control over the environment. Using Minigrid allowed the paper to go beyond reporting only performance curves but also to compare the state distributions induced by the different models (Figure 2), to vary the tasks in a way that it could simulate a continual learning problem (Figure 6), and even visualize distributions over rolled out trajectories (Figures 12 & 13). While Minigrid is inherently discrete, several works have demonstrated that discrete representations generalize well to more complex domains. For example, Hafner et al., 2020[1] directly compared continuous and discrete representations on over 50 Atari games and found that discrete representations perform much better. The paper's findings complement these works by providing insights into why discrete representations are effective and identifying promising directions such as continual learning.\n\n[1]https://arxiv.org/abs/2010.02193", "category": "Motivation_Analysis"}, {"question": "How does the limited breadth of domains in the experiments align with the paper's goal of providing a comprehensive empirical study?", "answer": "The authors chose to limit the breadth of experiments to a single domain to ensure thoroughness and depth in their study. This approach allowed them to examine many properties of the representations while minimizing variables, perform multiple runs, and achieve high confidence in the validity of their results. While the limited breadth of domains, the paper’s focus on depth and rigor in Minigrid experiments does provide valuable insights. Additionally, related works[1][2] have already demonstrated the applicability of discrete representations in more complex domains, supporting the generalizability of their findings.\n\n[1]https://arxiv.org/abs/2010.02193\n\n[2]https://arxiv.org/abs/2301.04104", "category": "Experimental_Analysis"}, {"question": "What empirical evidence supports the authors' interpretations regarding the role of sparsity in continual learning performance?", "answer": "The authors conducted additional experiments on sparsity, which are detailed in Appendix A.6. They found a clear, positive correlation between sparsity and continual learning performance up to a certain point, beyond which excessive sparsity becomes harmful. They also added a discussion in Section 3.3.3, supported by experimental results, to clarify their interpretations. These findings align with prior works, such as [Pan et al., 2021][1], which also identified an optimal level of sparsity for performance.\n\n[1]https://arxiv.org/abs/1911.08068", "category": "Experimental_Analysis"}, {"question": "Was the connection between the success of discrete representations in continual RL and representation sparsity empirically measured, and what were the findings?", "answer": "Yes, additional experiments were conducted to measure the connection between sparsity and continual learning performance. The results, included in Appendix A.6, show a clear, positive correlation between sparsity and performance up to an optimal level, beyond which too much sparsity becomes detrimental. These findings align with prior work such as [Pan et al., 2021][1], and Section 3.3.3 has been updated to include additional discussion on how the results support the paper's interpretations on sparsity.\n\n[1]https://arxiv.org/abs/1911.08068", "category": "Experimental_Exposition"}, {"question": "Why were regularization-focused baselines, particularly VAEs, not included in the study, and how does this relate to the representational capacity of the compared methods?", "answer": "The authors initially considered using Variational Autoencoders (VAEs) as a regularization-focused baseline. However, they found that no set of VAE hyperparameters could match the performance of a simple vanilla Autoencoder (AE) for model learning. The regularization imposed by VAEs hindered the model's ability to reconstruct observations, resulting in a significant performance gap. Consequently, the authors decided to exclude VAEs and proceed with vanilla AEs. While it is possible that a VAE might perform better in the continual reinforcement learning (RL) setting, the paper's previous results in the model learning setting lowered the expectations for its effectiveness.\n", "category": "Method_Disambiguation"}, {"question": "Why were only MiniGrid environments used in the experiments?", "answer": "The authors chose MiniGrid as a testbed to perform an extensive empirical evaluation of discrete representations. This allowed them to run large amounts of experiments, maintain a high level of control over the environment, and go beyond performance curves by comparing state distributions (Figure 2), simulating continual learning problems (Figure 6), and visualizing distributions over rolled-out trajectories (Figures 12 & 13).", "category": "Motivation_Analysis"}, {"question": "Can VQ-VAE-based methods effectively perform discretization in complex environments, and what evidence supports this capability?", "answer": " VQ-VAEs have been used to successfully compress and generate complex images in a number of works since the first paper on the architecture in 2017 (van den Oord et al.[1]). The authors of the original VQ-VAE paper even provide results to show how a VQ-VAE can reconstruct observations and predict believable, multi-step rollouts in DMLab specifically (Figures 5 & 7). The work since then has only gotten more impressive, with follow-up work using VQ-VAEs to generate high-resolution images (e.g., Razavi et al., 2019[2], Esser et al., 2021[3]) and videos (e.g., Yan et al., 2021[4], Walker et al., 2021[5]). The complexity of these datasets surpasses what is generally used in RL, indicating that discretization works effectively even in complex domains.\n\n[1]https://proceedings.neurips.cc/paper_files/paper/2017/file/7a98af17e63a0ac09ce2e96d03992fbc-Paper.pdf\n\n[2]https://proceedings.neurips.cc/paper/2019/file/5f8e2fa1718d1bbcadf1cd9c7a54fb8c-Paper.pdf\n\n[3]https://arxiv.org/abs/2012.09841\n\n[4]https://arxiv.org/abs/2104.10157\n\n[5]https://arxiv.org/abs/2103.01950", "category": "Method_Disambiguation"}, {"question": "How do the findings from experiments conducted in the Minigrid environment generalize to more complex, continuous domains?", "answer": "The experiments in Minigrid were rigorous and only possible due to its discrete nature, which allows for detailed visualizations. For generalization to more complex domains, several existing works (Hafner et al., 2020[1], Robine, 2020[2], Micheli et al., 2023[3], Hafner et al., 2023[4]) have demonstrated strong results in continuous domains. Hafner et al. (2020)[1] compared continuous and discrete representations in over 50 Atari games, finding discrete representations superior. The paper's setup, similar to DreamerV2, uses autoencoders for learning representations, and DreamerV2 employs a VQ-VAE variant. The paper's findings complement these works by explaining why discrete representations excel and identifying promising directions like continual learning.\n\n[1]https://arxiv.org/abs/2010.02193\n\n[2]https://arxiv.org/abs/2010.05767\n\n[3]Vincent Micheli, Eloi Alonso, and Franc¸ois Fleuret. Transformers are sample-efficient world models. In International Conference on Learning Representations (ICLR), 2023.\n\n[4]https://arxiv.org/abs/2301.04104", "category": "Experimental_Analysis"}, {"question": "What is the rationale behind using Minigrid as a testbed for evaluating discrete representations, and how do the findings align with prior work in more complex environments like Atari?", "answer": "The decision to use Minigrid as a testbed was based on the goal to perform a more extensive empirical evaluation surrounding the successful use of discrete representations, and for that, it was necessary to be able to run large amounts of experiments and have a high level of control over the environment. Using Minigrid allowed the paper to go beyond reporting only performance curves but also to compare the state distributions induced by the different models (Figure 2), to vary the tasks in a way that it could simulate a continual learning problem (Figure 6), and even visualize distributions over rolled out trajectories (Figures 12 & 13). While experiments in different environments would certainly be nice to have, there is existing prior work that experiments with discrete representations (including comparisons to continuous representations) in Atari (Hafner et al., 2020), and the paper's results align with the findings in that work. Most importantly, Hafner et al. (2020) perform a direct comparison between continuous and discrete representations on over 50 Atari games (Figure 5) and find that discrete representations perform much better. The paper's setup is even similar to DreamerV2 in that both works use autoencoders to learn representations, and DreamerV2 uses a special case of a VQ-VAE (where the codebook is one-hot). The paper's findings are complementary to these other works; they show that this is possible, and the paper helps understand why that is the case and identifies promising directions (continual learning).\n\n[1]https://arxiv.org/abs/2010.02193\n", "category": "Motivation_Analysis"}, {"question": "Why is the efficiency of representation convergence important in the context of continual learning, given that both discrete and continuous representations converge almost immediately in the episodic RL task?", "answer": "The efficiency of representation convergence is crucial for continual learning because agents need to adapt to constantly changing environments. In the Door Key environment, the VQ-VAE-based agent takes ~80k steps to converge, while the AE-based agent takes ~250k steps. Saving 170k steps is a significant improvement for continual RL, where efficiency is paramount. The long-term goal is to develop agents that can learn in just several episodes, making such efficiency gains a step in the right direction.", "category": "Method_Disambiguation"}, {"question": "What causes the larger error bars and the dramatic increase in error when transitioning from 512 to 1024 dimensions in Figure 3?", "answer": "The larger error bars in Figure 3 are caused by a bimodal distribution resulting from a few diverging runs in the end-to-end method. These divergences occur because the model must backpropagate multiple loss terms through both the transition model and encoder over multiple time steps, complicating the optimization process. In contrast, other methods train the autoencoder and transition models separately and do not backpropagate over multiple time steps, simplifying optimization. The paper's work aims to be a thorough study that improves the understanding of this area, and reporting all results, including 'not clean' ones, is an important part of helping others understand the realistic state of these problems.", "category": "Experimental_Analysis"}, {"question": "Why is the gap between FTA and VQ-VAE much larger in the door key environment compared to the crossing environment?", "answer": "The difference in scale is due to the evaluation metric used. In the crossing environment, the objective causes the agent to spread out, whereas in the door key environment, the objective concentrates the agent into fewer states. This results in a higher KL divergence in the door key environment when the agent is off by a few steps, as shown in Figures 12 and 13, where there is more overlap in the crossing environment.", "category": "Experimental_Analysis"}, {"question": "How dependent are the results of the paper on the pretraining distribution used for the learned representations, particularly in domains like Atari or Mujoco where random policies sample a limited state distribution?", "answer": "The results are not entirely dependent on the pretraining distribution because the representations are continually trained as the policy changes, allowing them to adapt to distribution shifts. Experiments without the pretraining phase also worked but resulted in less stable RL performance, indicating a plasticity issue. This issue could potentially be addressed by works on the stability-plasticity dilemma(e.g. Dohare et. al 2023[1], Nikishin et. al, 2023[2]), though this has not yet been tested.\n\n[1]https://arxiv.org/abs/2306.13812\n\n[2]https://openreview.net/forum?id=O9cJADBZT1", "category": "Experimental_Analysis"}, {"question": "Was Section 3.3.3 updated to include additional discussion on how the experimental results support interpretations on sparsity?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "WZ6NY4JfFX", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Revisiting the Role of Language Priors in Vision-Language Models", "authors": ["Zhiqiu Lin", "Xinyue Chen", "Deepak Pathak", "Pengchuan Zhang", "Deva Ramanan"], "abstract": "Vision-language models (VLMs) are impactful in part because they can be applied to a variety of visual understanding tasks in a zero-shot fashion, without any fine-tuning. We study currently popular *generative VLMs* that are trained for next-word generation given the image. We explore their zero-shot performance on the illustrative task of image-text retrieval across 8 popular vision-language benchmarks. Our first observation is that they can be repurposed for discriminative tasks (such as image-text retrieval) by simply computing the match score of generating a particular text string given an image. We call this probabilistic score the *Visual Generative Pre-Training Score* (VisualGPTScore). While the VisualGPTScore produces near-perfect accuracy on some retrieval benchmarks, it produces poor accuracy on others. We analyze this behavior through a probabilistic lens, pointing out that some benchmarks inadvertently capture unnatural language distributions by creating adversarial but unlikely text captions. In fact, we demonstrate that even a ``blind'' language model that ignores any image evidence can sometimes outperform all prior art, reminiscent of similar challenges faced by the visual-question answering (VQA) community many years ago. We derive a probabilistic post-processing scheme that controls for the amount of linguistic bias in generative VLMs at test time without having to retrain or fine-tune the model. We show that the VisualGPTScore, when appropriately debiased, is a strong zero-shot baseline for vision-language understanding, oftentimes producing state-of-the-art accuracy.", "keywords": "Vision-Language Models;Visio-Linguistic Compositionality;Cross-Modal Retrieval;Computer Vision;Natural Language Processing", "qa_pairs": [{"question": "What is the motivation behind using the marginal $P_{train}(t)^\\alpha$ in the proposed calibration approach, and how does it relate to point-wise mutual information (PMI) in improving the calibration of generative VLMs?", "answer": "The proposed calibration approach, involving dividing by the marginal $P_{train}(t)^\\alpha$, is motivated by the analysis in Section 3 and is not directly based on PMI. Here is a concrete example (cf. Figure 1) for ease of understanding: in some recent compositionality benchmarks such as SugarCrepe, a negative caption such as '*people are cooking in a kitchen*' is more common in the VLM’s trainset (thus a higher $P_{train}(t)$) than the positive caption '*people are posing in a kitchen*'. As such, dividing by $P_{train}(t)$ forces the model to select the caption that matches better to the image (rather than the caption with a larger marginal prior). The paper's Appendix A shows that the paper's approach $\\frac{P(t|i)}{P(t)^\\alpha}$ happens to be the same as smoothed estimates of pointwise mutual information (PMI$^k$), which also reduces the effect of marginal priors from training data. ", "category": "Motivation_Analysis"}, {"question": "How is the measurement of $P(t^i_{positive})$ with respect to $P(t^i_{negative})$ conducted for each individual test instance in 'blind models'?", "answer": "$P_{train}(text)$ is measured for each individual text independently using Monte Carlo estimation (Equation 9): $P_{train}(text) \\approx \\frac{1}{n} \\sum_{k=1}^n P_{train}(text | image_k)$, where $k$ represents the number of random Gaussian noise images sampled.", "category": "Method_Mechanics"}, {"question": "Does measuring $\\alpha^{*}$ require test-time privileged information, and if not, how is the optimal alpha determined in practice?", "answer": "The method does not require test-time privileged information. In practice, a small validation set from the target task is used to search for the optimal alpha, or practical assumptions can be made without tuning alpha, such as choosing $\\alpha=1$. Table 2-a demonstrates that both approaches consistently improve performance without using test-time privileged information.", "category": "Method_Mechanics"}, {"question": "How does the proposed method address the concern of high computational costs in large-scale image-text retrieval? Considering that Image-text retrieval, often utilized in search engines, demands high time efficiency. With this method, for every new image, the VLM must process all texts, which incurs substantial computational costs. In contrast, traditional methods like CLIP pre-compute text embeddings and only require a dot product with each image embedding.", "answer": "The proposed method can be practical for applications such as evaluating text-to-image generation, where it measures the similarity score between a text prompt and a generated image. It complements CLIPScore[1,2], which struggles with understanding complex texts involving compositions of objects, attributes, and their relations. Additionally, it is useful for recent image-text retrieval benchmarks like ARO and SugarCrepe, which evaluate VLMs’ compositional reasoning capabilities. The method's focus is on studying linguistic priors in generative VLMs and vision-language benchmarks.\n\n[1] Ruiz et al. “DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation”.\n\n[2] Brooks et al. “InstructPix2Pix: Learning to Follow Image Editing Instructions”.", "category": "Method_Disambiguation"}, {"question": "What is the time-efficiency comparison between ITCScore, ITMScore, and VisualGPTScore when using the same BLIP model for large-scale inference (1000 images x 5000 texts)?", "answer": "The time-efficiency comparison is as follows: ITCScore has an inference cost of O(|image| + |text|) and takes 2 minutes on Flickr30K, while both ITMScore and VisualGPTScore have an inference cost of O(|image| x |text|) and each take 3 hours on Flickr30K.", "category": "Method_Comparison"}, {"question": "How does the deterioration of OTS scores in Table 7 as the dataset size increases, and the widening gap with ITM even when using the optimal alpha, indicate an inherent limitation of VisualGPTScore?", "answer": "The deterioration of OTS scores and the widening gap with ITM in Table 7 highlight the inherent linguistic priors of VisualGPTScore, which make it a biased estimator of $P_{train}(image | text)$. These issues are due to the “long-tailed” problems that arise from learning from imbalanced datasets, where certain sentences are more common than others. A detailed discussion on this limitation is provided in Appendix C, titled *“Is VisualGPTScore a biased estimator?”*.", "category": "Experimental_Analysis"}, {"question": "What empirical evidence supports the suitability of using Gaussian noise images as 'null' inputs in generative VLMs?", "answer": "Gaussian noise images are selected as 'null' inputs because they empirically help calibrate models, as shown in [1], which uses a language model to approximate P(text) = P(text | “N/A”). Their null text prompt “N/A” is also not seen during training. Additionally, both Gaussian noise images and training set images were tested and found to perform similarly, as evidenced in Table 4.\n\n[1]Zhao et al. “Calibrate Before Use: Improving Few-Shot Performance of Language Models”", "category": "Experimental_Analysis"}, {"question": "How do the null text prompts used in the sentence classification task, such as 'N/A', 'abc', and random strings, affect the calculation of marginals, and what additional experiments were conducted to explore the impact of different content-free inputs on performance? Since table 3 in [1] demonstrates that the choice of content-free input does affect accuracy.\n\n[1]Zhao et al. “Calibrate Before Use: Improving Few-Shot Performance of Language Models”", "answer": "Here are some concrete examples from [1] for the sentence classification task, which used null input strings such as “N/A”, “abc”, or even “dasjhasjkdhjskdhds” to approximate the marginal:\n\n- P(answer) ≈ P(answer | “N/A was born in”)\n- P(answer) ≈ P(answer | “abc was born in”)\n- P(answer) ≈ P(answer | “dasjhasjkdhjskdhds was born in”)\n\nWhile these null text strings themselves might be included in the pretraining corpora, they are certainly out-of-distribution for calculating the above marginals.  The authors conduct tests using **mean training image** (in place of Gaussian noise images) and find they yield almost identical performance. Concretely, by sampling random sets of 100 or 1000 images from the trainset and calculating their average (which visually resembles Gaussian noise images too). The table below shows that both **mean training images** and Gaussian noise images yield similar performance on Winoground (following Table 4’s setup):\n\n| Method               | Gaussian Noise | Mean of 100 train images | Mean of 1000 train images |\n|----------------------|-----------|------------|-------------|\n| $\\alpha=\\alpha_{val}^*$  |    36.4±0.1       |      36.2±0.4       |      36.3±0.3        |\n\n[1]Zhao et al. “Calibrate Before Use: Improving Few-Shot Performance of Language Models”", "category": "Experimental_Analysis"}, {"question": "How does the proposed method demonstrate performance improvements across various tasks, specifically in the context of complementing CLIPScore, beyond the evaluated text-image retrieval tasks?", "answer": "Recent compositionality benchmarks like Winoground are specifically designed for image-text evaluation within the context of image-text retrieval. In these benchmarks, human annotators assign a score of 1 to an (image, text) pair if it matches, or 0 if it does not. This means that **a higher performance on these benchmarks corresponds to a better alignment with human judgements**. To demonstrate the effectiveness of the paper’s proposed method, the Winoground results for two recent text-to-image evaluation algorithms, LLMScore [1] and VPEval [2], both of which claim to complement CLIPScore, are presented. The results demonstrate that the paper's proposed score exhibits stronger compositional understanding than these heavily engineered methods that use expensive ChatGPT during inference.\n\n| Method                     | Text Score | Image Score | Group Score |\n|----------------------------|------------|-------------|-------------|\n| CLIPScore                  | 28.5       | 10.8        | 12.8        |\n| VisualGPTScore |  36.5       | 21.5        | 16.8        |\n| LLMScore (GPT4 + BLIPv2)     | 21.3       | 17.8        | 12.5        |\n| VPEval (GPT3.5 + BLIPv2)     | 12.8       | 11.0        | 6.3         |\n\n[1]Lu et al. LLMScore: Unveiling the Power of Large Language Models in Text-to-Image Synthesis Evaluation. 2023.\n\n[2] Cho et al. Visual Programming for Text-to-Image Generation and Evaluation. 2023.", "category": "Experimental_Analysis"}, {"question": "What is the novel contribution of 'VisualGPTScore' compared to prior works that also measure similarity using the average log likelihood of strings given a fixed context, particularly in the context of generative VLMs?", "answer": "The novel contribution of 'VisualGPTScore' lies in revisiting and addressing the language priors and biases in generative VLMs, an issue that has been neglected in recent image-text retrieval benchmarks. Unlike prior works, such as [1], which focus on training the generative score from scratch, 'VisualGPTScore' emphasizes its zero-shot application, especially on how to improve its performance under a substantial distribution (marginal) shift between VLMs’ pretraining data and downstream tasks. Additionally, [1] only reports results for text-to-image retrieval tasks, omitting image-to-text retrieval, although it uses the same COCO and Flickr30K benchmarks, likely due to language biases. 'VisualGPTScore' addresses this issue and provides algorithms for debiasing the generative score, significantly improving performance on benchmarks like COCO and Flickr30K.\n\n[1] Antonie Miech et al. “Thinking Fast and Slow: Efficient Text-to-Visual Retrieval with Transformers”.", "category": "Method_Comparison"}, {"question": "Why is the gap between test and training (p_te vs p_tr) relevant for zero-shot models, even though the model does not use p_tr during training?", "answer": "The term $P_{train}$ in the paper refers to the pretraining datasets of vision-language models (VLMs), such as LAION for BLIPv1 and BLIPv2. Even in zero-shot applications, where the model does not use training data, popular benchmarks like ARO and SugarCrepe are designed to evaluate VLMs in a zero-shot manner and do not provide a training split.", "category": "Method_Disambiguation"}, {"question": "Why was BLIP chosen for the experiments instead of more recent vision-language models like LLaVA or BLIPv2, which can directly compute the language marginals?", "answer": "The language marginal ($P_{train}(t)$) is derived from pretraining captions of vision-language models (VLMs). LLaVA and BLIPv2 cannot approximate this marginal without sampling images, as their language models are pretrained on different text corpora than VLMs’ pretraining captions. Additionally, results from BLIPv2, specifically its stage-1-Q-Former and stage-2-FlanT5-XL components (a state-of-the-art captioning model), are included in the paper, demonstrating that the $\\alpha$-debiasing solution improves all BLIP model variants even with a fixed $\\alpha$=1.\n", "category": "Motivation_Analysis"}, {"question": "How does the proposed method, which uses an optimal alpha determined through cross-validation, compare to other baseline methods that calibrate model predictions before inference?", "answer": "The proposed method can be viewed as a calibration approach for the language prior's influence. Other baselines for calibration, such as Platt scaling or adjusting the softmax temperature, would not affect image-to-text retrievals. Additionally, using a fixed alpha = 1 without cross-validation still provides a consistent and reasonable performance increase across all balanced benchmarks.", "category": "Method_Comparison"}, {"question": "Did the study compare the proposed debiasing solution with state-of-the-art generative vision-language models like BLIP-2?", "answer": "Yes, the study included comparisons with BLIP-2 (both stage-1-QFormer and stage-2-FlanT5) in Appendix D. The results, shown in Table 9, demonstrate that the debiasing solution improves performance on balanced benchmarks across all model variants even with a fixed alpha = 1. Additionally, image-to-text retrieval results(text score) for Winoground and EqBen are provided.\n\n| Winoground                          | $\\alpha=0$ | $\\alpha=1$ | $\\alpha=\\alpha^*$ |\n| ------------------------------ | ---------- | ---------- | ----------------- |\n| BLIP-1                         | 27.0       | 33.0       | 36.5              |\n| BLIP-2 (stage-1-QFormer)       | 24.3       | 29.3       | 33.0              |\n| BLIP-2 (stage-2-FlanT5)        | 25.3       | 31.5       | 34.3              |\n\n| EqBen                          | $\\alpha=0$ | $\\alpha=1$ | $\\alpha=\\alpha^*$ |\n| ------------------------------ | ---------- | ---------- | ----------------- |\n| BLIP-1                         | 9.6       | 19.8       | 19.8              |\n| BLIP-2 (stage-1-QFormer)       | 12.2       | 21.9       | 22.2              |\n| BLIP-2 (stage-2-FlanT5)        | 8.5       | 22.0       | 22.0              |", "category": "Method_Comparison"}, {"question": "Why does the $\\alpha$-debiasing method operate at the dataset level ($P_{train}$ vs. $P_{test}$) rather than the instance level, and shouldn’t the value of $\\alpha^{*}$ vary based on the specific target ($t$)?", "answer": "The $\\alpha$-debiasing method($\\frac{P_{train}(text | image)}{P_{train}(text)^\\alpha}$ operates at the dataset level because it addresses the train-test shift between the VLM’s training set (web corpora) and testing benchmarks (e.g., SugarCrepe), as discussed in Section 3. Without additional training, measuring an $\\alpha$ that varies based on specific target texts is unclear. The paper focuses on training-free debiasing.", "category": "Method_Disambiguation"}, {"question": "How does the proposed method perform on zero-shot classification tasks, and why does it not naturally support VQA tasks?", "answer": "The proposed method, designed to calculate the similarity score between an image and a caption, does not naturally support VQA tasks. However, it was evaluated on ImageNet1K for zero-shot classification, where the $\\alpha$-debiasing solution with a fixed $\\alpha=1$ nearly doubled the performance from 18.6% to 36.2%. By sampling an additional one-shot validation set (and repeat using 3 random seeds), a near-optimal $\\alpha$ was found, achieving 40.0% without retraining or finetuning the model.\n\n| Method                                  | Result      |\n|-----------------------------------------|-------------|\n| ITCScore                                | 31.7       |\n| ITMScore                                | 37.4       |\n| VisualGPTScore ($\\alpha=0$)             | 18.6       |\n| VisualGPTScore ($\\alpha=1$)             | 36.2       |\n| VisualGPTScore ($\\alpha^*_{val}=0.65\\pm0.01$) | 40.0$\\pm$0.1  |\n| VisualGPTScore ($\\alpha^*_{test}=0.69$)  | 40.2       |\n\nIn this experiment, by sampling only a single Gaussian noise image for the calculation of $P_{train}(t)$, a negligible inference cost is incurred, equivalent to the testing of one additional image.\n", "category": "Experimental_Exposition"}, {"question": "Why do negative captions rarely occur in web corpora, and how does this affect the performance of 'blind models'?", "answer": "Negative captions rarely occur in web corpora, particularly in benchmarks like ARO. This explains why 'blind models' can outperform state-of-the-art methods on many recent popular benchmarks. The experiments suggest that existing benchmarks are not sufficiently challenging and contain correlations that can be exploited using language priors.", "category": "Experimental_Analysis"}, {"question": "What specific types of distribution shifts are observed across the datasets, and how do examples such as ImageNet1K and ARO illustrate these shifts?", "answer": "The paper focuses on the train-test marginal shift in $P(text)$, which is common in vision-language benchmarks with balanced $P_{test}(text)$. For example, in ImageNet1K, $P_{test}(text)$ is the same for all 1000 classes,  as each class has the same number of testing images. However, VLM’s training data  (image-text pairs found on the Web) tend to be naturally imbalanced. For instance, $P_{train}(cat)$ is much higher than $P_{train}(frog)$ because people tend to upload more cat images than frog images to the Web. As such, to account for the fact that $P_{test}(cat) = P_{test}(frog)$ in balanced test sets such as ImageNet, the authors propose dividing the paper's score by $P_{train}(text)^\\alpha$ (e.g., $\\alpha$=1) to improve performance. This strategy is effective for other balanced VLM benchmarks such as Winoground and EqBen (as shown in Table 3). \n\nHowever, for popular testing benchmarks such as ARO, negative captions are constructed to be implausible and thus rare in VLM's training data too. For example, the negative caption $t_{negative}$ of  'white a duck spreads its wings in while the water' is equally unlikely to occur in Web data (i.e., $P_{train}(t_{negative}) \\approx 0 = P_{test}(t_{negative})$). Therefore, dividing by $P_{train}(text)^\\alpha$ (where $\\alpha$>0) hurts performance because it increases the match score for negative captions. \n\nIt is believed that balanced benchmarks such as Winoground and EqBen may be better testbeds for compositional reasoning (as evidenced by the lower SOTA performance), and so conjecture that debiasing (with a val-tuned $\\alpha$ or even by naively assuming $\\alpha$ = 1) will continue to be useful.", "category": "Experimental_Exposition"}]}
{"id": "ueqTjOcuLc", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Exploring Collaboration Mechanisms for LLM Agents: A Social Psychology View", "authors": ["Jintian Zhang", "Xin Xu", "Shumin Deng"], "abstract": "As Natural Language Processing (NLP) systems are increasingly employed in  intricate social environments, a pressing query emerges: \\emph{Can these NLP systems mirror human-esque collaborative intelligence, in a multi-agent society consisting of multiple large language models (LLMs)?} \nThis paper probes the collaboration mechanisms among contemporary NLP systems by melding practical experiments with theoretical insights. We fabricate four unique 'societies' comprised of LLM agents, where each agent is characterized by a specific 'trait' (easy-going or overconfident) and engages in collaboration with a distinct 'thinking pattern' (debate or reflection). \nEvaluating these multi-agent societies on three benchmark datasets, we discern that LLM agents navigate tasks by leveraging diverse social behaviors, from active debates to introspective reflections. \nNotably, certain collaborative strategies only optimize efficiency (using \\textit{fewer} API tokens), but also outshine previous top-tier approaches. \nMoreover, our results further illustrate that LLM agents manifest human-like social behaviors, such as conformity or majority rule, mirroring foundational Social Psychology theories. \nIn conclusion, we integrate insights from Social Psychology to contextualize the collaboration of LLM agents, inspiring further investigations into the collaboration mechanism for LLMs. \nWe commit to sharing our code and datasets (already submitted in supplementary materials), hoping to catalyze further research in this promising avenue.", "keywords": "Multi-Agent;Large Language Model;Society of Mind;Social Psychology;Human Group Dynamics", "qa_pairs": [{"question": "How can confidence be placed in the statistical significance of the differences in performance on MMLU S1 (e.g., better on min but not on avg), ensuring they are not due to chance?", "answer": "To ensure the differences in performance are statistically significant and not by chance, the authors conducted additional ANOVA analysis, which is detailed in the main experiment of section 3.1 and included in Table 4, Appendix F.1 of the paper. The p-value associated with the 'collaborative strategy' is significantly below the 0.05 threshold, indicating its substantial impact. In contrast, the p-values for 'society' and 'collaborative strategy and society' are greater than 0.05, corroborating that the influence of collaborative strategy outweighs that of society and is significantly correlated with performance.", "category": "Experimental_Analysis"}, {"question": "What is the utility of having a society composed of agents with different traits if there are no marked differences among them?", "answer": "The composition of agents with different 'traits' (easy-going or overconfident) in the society, even without marked differences, provides a nuanced understanding of social dynamics, with Sherif’s Classic Autokinetic Effect Study being a representative piece of research. It serves to model real-world scenarios where subtle variations in behavior or capability can significantly impact collective outcomes. For example, as can be seen from Table 2, the performance exhibited by different societies is not always exactly consistent and still present differences. The paper's design of society’s agent composition initially did not account for the lack of significant differences among different societies. However, this phenomenon became apparent after the paper's simulation experiments. Besides, The assertion that collaborative strategies overshadow the influence of agent composition may be premature. Moreover, the paper's experiments in the Appendices (as illustrated in Figure 8, a word frequency analysis was conducted on two societies with the most distinct compositions, S1 and S4.) reveal that all societies tend to harmonize. This suggests a propensity of agents to exhibit easy-going rather than overconfident traits. It is speculated that this could be due to issues with the prompt or the model, involving two potential scenarios: (1) The model itself is not at fault; rather, the ineffectiveness of the prompt is the issue. (2) The prompt is adequate, but the model does not adhere to it. Scenario (2) is considered the more probable cause, as the prompt is found to be effective enough. Furthermore, this phenomenon is related to the alignment of LLMs, which discourages the adoption of traits like overconfidence that conflict with human preferences. This indicates a need for LLMs to strike a balance between ‘following instructions’ and ‘maintaining non-toxicity’, instead of only trying to mirror human preferences.", "category": "Method_Disambiguation"}, {"question": "Why might the collaborative strategy $p_0p_0p_1$ be necessary for achieving optimal performance, even when $p_0p_0p_0$ is used initially?", "answer": "The collaborative strategy $p_0p_0p_1$ is not always superior to $p_0p_0p_0$, but it achieves performance close to $p_0p_0p_0$ in some cases. On MATH S3, $p_0p_0p_1$ performs worse, but for the MMLU dataset, it is often the optimal choice. $p_0p_0p_1$ increases the proportion of 'correcting mistakes' while maintaining the rate of 'changing correct answers'. Figure 12 shows that $p_0p_0p_1$ achieves a consistency of 2.8 in the first two rounds, with minor changes in answers triggered by one reflection ($p_1$), akin to how new trials in human society(assuming the stability of human society) lead to discoveries and innovations.", "category": "Experimental_Analysis"}, {"question": "How are the agents initialized and conditioned, and what specific LLM settings were used to implement overconfident and easy-going agents in this study?Is the prompt indicating how agents should behave sufficient to condition the agents effectively?", "answer": "The paper enhanced the experiments to validate the effectiveness of the Prompt and Model, maintaining alignment with the main experiments detailed in this paper. Detailed experimental setup is available in Appendix C: for the prompt, refer to Table 3 in the Appendix; and for complete dialogue examples, refer to Figures 9 and 10 in the Appendix. The experiments used the gpt-3.5-turbo model, with both overconfident and easy-going agents using 0-shot prompting. Effectiveness of the Prompt: Analysis of the word cloud in Figure 8 suggests that the prompt designed for easy-going traits is successful. Consequently, the paper's attention turns to confirming the effectiveness of the prompt for overconfident traits. In the absence of a robust validation approach, the authors draw upon the paper's existing experimental results and practical experience to assess the prompt's effectiveness from four key angles: (1)Behavior Explanation: Two behaviors, ‘confident in your answer’ and ‘persuades other agents to believe in you’, were identified, which align with the behavioral display of overconfidence. (2) Refusal to Accept: The authors employ a role-play method to prompt the model, and then ask if the model comprehends the prompt (if a prompt induces harmful content output, the model will refuse it). When checking the logs, the model didn't refuse the prompts and replied affirmatively with 'OK'. (3) Re-asking: Building on the previous step, the authors posed another query: 'What should you do if someone else's answer differs from yours?'. The model's reply, 'I will convince them', proves the effectiveness of the paper's prompt tailored for overconfident traits. (4) Example Analysis: Delving deeper, the authors contextualize the 're-asking' with a practical scenario. Even if the model's own response is wrong and the paper's prompted answer is correct, the model still insists on its viewpoint. This observation further substantiates the effectiveness of the paper's prompt tailored for overconfident traits. In summary, these experiments collectively affirm the effectiveness of the prompts the paper has developed for both easy-going and overconfident traits.", "category": "Experimental_Analysis"}, {"question": "What would a non-simplified collaboration setup entail, and how might it influence the observations and conclusions presented in the paper?", "answer": "A more complex collaboration setup might involve a larger number of agents with a greater diversity of traits and interaction rounds. It is hypothesized that such a setup would provide more nuanced insights into the dynamics of agent collaboration and competition. Including more complex societal structures, such as hierarchical or network-based interactions, could also significantly impact the results and the generalizability of observations. Previous research [1] outlines several branches of psychological research, such as social psychology, group psychology, moral psychology, and the psychology of judgment and decision-making. Due to constraints in assessment, context length, modality, and datasets, many aspects of machine psychology cannot be immediately studied. Therefore, in the foreseeable future, more specific setups are anticipated in two main dimensions: \n\n(1) A Richer Society: \n- An environment with feedback. Feedback plays a crucial role in human interaction. Studying machine society based on feedback phenomena and agents’ responses is a fascinating research direction. \n- A richer variety of agent personalities, more role-playing, agents driven by different LLMs, larger-scale machine societies, and collaboration between machine societies would be more challenging settings. \n\n(2) More Detailed Collaborative Strategies: \n- Expanding methods of reflection. Reflection can take various forms, such as convergent thinking, divergent thinking, critical thinking, etc., which can be referred to [2]. Notably, Chain of Thought [3] can also be seen as a form of self-reflection. \n- Human-computer interaction is a future trend. It’s promising to introduce human-computer collaboration to observe human and agent behaviors. \n- Introducing an external third-party agent to determine the appropriate thinking pattern and timing for each participating agent. Moreover, these research extensions on ‘Non-Simplified Collaboration’ could have potential impacts, such as: \n - They might uncover more efficient collaboration techniques that swiftly pinpoint correct answers, diverging from the methods currently discussed in the paper. \n - Furthermore, employing agents powered by various LLMs could result in fascinating behavioral dynamics, such as increased overconfidence during collaboration. It’s also possible that reflection doesn’t invariably result in model hallucinations, opening new avenues for exploration.\n\n[1] Machine Psychology: Investigating Emergent Capabilities and Behavior in Large Language Models Using Psychological Methods\n\n[2] https://www.magneticmemorymethod.com/types-of-thinking\n\n[3]Personality traits, emotional intelligence and decision-making styles in Lebanese universities medical students\n\n", "category": "Experimental_Analysis"}, {"question": "What is the motivation for using agents backed by the same language model, and how would using different models influence the current experimental setup?", "answer": "*The motivation behind using agents backed by the same language model*: \n- To ensure consistency in language understanding and response generation capabilities across agents. This homogeneity allows the paper to isolate the effects of the designed societal roles and collaborative strategies. \n- Many studies [1][2] have also adopted settings of using the same language model, and the paper just follow the commonly used settings. \n- Access to APIs of other LLMs like Bard and Claude is limited, and sufficient computing resources to conduct all the experiments on open-source LLMs were not available at the time of writing. Therefore, the same language model had to be used tentatively.\n\n[1] Improving Factuality and Reasoning in Language Models through Multiagent Debate.\n\n[2] Generative Agents: Interactive Simulacra of Human Behavior.", "category": "Motivation_Analysis"}, {"question": "How would using different models influence the current experimental setup?", "answer": " *The influence of using different models on the current experimental setup*: Employing different models could introduce variability in basic language processing capabilities, which would be an interesting dimension to explore in understanding how model-specific characteristics influence collaborative outcomes. Different models use varied training methods and data, so they differ in the breadth and depth of knowledge, and in their ability to follow instructions. \n- Regarding the experimental setup, the paper's framework still supports this, and it can be easily implemented with a few lines of code. \n- Regarding the experimental results, ideally, agents might show overconfident behavior in collaboration, leading to absolute obedience between agents. Additionally, agents could complement each other through collaboration to arrive at more accurate answers. ", "category": "Experimental_Analysis"}, {"question": "Why was the societal setup simplified in terms of the number of traits and the size of the society, and how does this limitation align with the goals of the study?", "answer": "The simplification of societal setups and trait numbers was a necessary limitation for this preliminary investigation, as acknowledged in the Limitations section in Appendix E. Human societies are vast and diverse, necessitating efficient modes of collaboration. For machine societies, an increase in scale brings about an increase in the length of context, which currently remains a significant challenge.  A complex society functions through specialization, where small groups work collaboratively, much like how a large problem needs to be broken down into smaller problems for resolution. The paper's current research can only offer insights on how to solve small-scale problems.", "category": "Method_Disambiguation"}, {"question": "What evidence or analysis supports the claim that collaborative strategies overshadow the influence of agent composition?", "answer": "The effectiveness of collaborative strategies is supported by word cloud analysis in Figure 8, which demonstrates that the model exhibits easygoing traits even when prompted to display overconfident traits. Additionally, Appendix F.4 provides a detailed analysis of the overconfident prompt across four angles: Behavior Explanation, Refusal to Accept, Re-asking, and Example Analysis. These analyses substantiate the effectiveness of the prompts tailored for overconfident traits. ", "category": "Experimental_Analysis"}, {"question": "What are the real-world applications and generalizability of the findings beyond the datasets used in the study?", "answer": "The findings have two potential real-world applications: (1) They provide insights for problem-solving in multi-agent frameworks, where conclusions about the number of rounds, the number of agents, and thinking patterns can inform collaborative strategy design. (2) The experimental framework can serve as a reference for psychological researchers and LM designers, For instance, psychologists can adapt the paper's framework to derive conclusions that, combined with theories of human social psychology, can help address some issues in LLMs and design better machine society architectures. Some conclusions, such as the necessity of collaboration in complex tasks and the heterogeneity of collaboration in simple tasks, are generalizable. However, some conclusions are time-sensitive and will evolve as LMs improve.", "category": "Experimental_Analysis"}, {"question": "What is the primary takeaway or key message from the experiments, considering the significant variance observed in the results?", "answer": "Despite the significant variance in accuracy, the primary takeaway is the importance of collaboration and the impact of different collaborative strategies and thinking patterns on agent behavior. Key findings include that reflection can introduce new opinions but may lead to hallucinations due to lack of feedback, while debates can foster consensus and mitigate hallucinations. The experiments were repeated five times, and ANOVA analysis confirmed the homogeneity of variances, as detailed in Appendix E and F. Additional experiments are provided in the appendix for further insights.", "category": "Experimental_Exposition"}, {"question": "What meaningful conclusions can be drawn from the results of the W-T metric?", "answer": "W-T is a commonly used metric in many papers. This paper focuses on two variants of W-T: $\\underline{W-T}$ and \\uuline{$W-T$}, denoting the total number of occurrences where the performance exceeds $p_0p_0p_0$ strategy under the same $\\underline{collaborative~strategy}$ / \\uuline{$society$}. The W-T metric results show two key conclusions: (1) Performance is not significantly related to society, as the optimal society varies across the three datasets. (2) The strategy $p_0p_0p_1$ is superior to other strategies except for $p_0p_0p_0$. A significance analysis further confirms that the impact of society is not significant, aligning with the conclusions from the \\uuline{$W-T$}.", "category": "Experimental_Exposition"}, {"question": "Were experiments conducted involving two agents, and what were the results of strategies involving p0p0 and p0p1?", "answer": "Experiments involving two agents with easy-going traits were conducted five times on three datasets. The results, presented in Appendix F.3 (Table 8 and Table 9), show that for two-round collaboration, $p_0p_0$ and $p_0p_1$ are the best and second-best strategies, respectively. However, these results are not as effective as the best results achieved with three agents performing three-round collaboration.", "category": "Experimental_Exposition"}, {"question": "How did the paper address the issue of significant variance in the experimental results and what consistent findings were identified despite this variance?", "answer": "The authors addressed the variance by repeating the experiments five times and performing ANOVA analysis to confirm the homogeneity of variances. Despite the variance, they identified consistent findings on the importance of collaboration, the impact of different collaborative strategies, and thinking patterns on agent behavior. For example, reflection can bring new opinions but may lead to hallucinations due to the lack of feedback, while debates can mitigate hallucinations and bring consensus.", "category": "Experimental_Exposition"}, {"question": "Is it the case that adding more debates will not lead to performance improvement once a society achieves strong consistency after a round of collaboration?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the number of rounds closely related to consistency?\n\n", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the analysis of the word cloud in Figure 8 suggest that the prompt designed for easy-going traits is successful?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the paper's acknowledgment of the simplification in societal setup and trait numbers included in the Limitations section of Appendix E?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is three agents a more suitable choice among 2-4 agents?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper claim to provide insights on solving large-scale societal problems in its current research?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are overconfident and easy-going agents in the study initialized using 0-shot prompting?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "ysue5S6cVS", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Confidence-driven Sampling for Backdoor Attacks", "authors": ["Pengfei He", "Han Xu", "Jie Ren", "Yingqian Cui", "Shenglai Zeng", "Yue Xing", "Jiliang Tang", "Makoto Yamada", "Mohammad Sabokrou"], "abstract": "Backdoor attacks aim to surreptitiously insert malicious triggers into DNN models, granting unauthorized control during testing scenarios. Existing methods lack robustness against defense strategies and predominantly focus on enhancing trigger stealthiness while randomly selecting poisoned samples. Our research highlights the overlooked drawbacks of random sampling, which make that attack detectable and defensible. The core idea of this paper is to strategically poison samples near the model's decision boundary and increase defense difficulty. We introduce a straightforward yet highly effective sampling methodology that leverages confidence scores. Specifically, it selects samples with lower confidence scores, significantly increasing the challenge for defenders in identifying and countering these attacks. Importantly, our method operates independently of existing trigger designs, providing versatility and compatibility with various backdoor attack techniques. We substantiate the effectiveness of our approach through a comprehensive set of empirical experiments, demonstrating its potential to significantly enhance resilience against backdoor attacks in DNNs.", "keywords": "Backdoor attack;sampling method", "qa_pairs": [{"question": "Is the proposed method inefficient for training surrogate models on large datasets, and what measures are taken to ensure time efficiency?", "answer": "The proposed method is not inefficient for training surrogate models on large datasets. First, the surrogate model does not need to be trained till the end; shorter training epochs (e.g., 60 epochs) are sufficient for selection and reduce the risk of overfitting which may reduce the poisoning effect. An ablation study about training epochs for the surrogate model in the CIFAR-10 dataset and ResNet18 model in Table below:\n\nEpochs|20|60|80|100\n--------|----|--|---|-----\nNo defense|94.4|93.6|93.2|92.1\nSS|14.6|23.2|22.9|23.6\nStrip|12.1|26.2|26.7|25.9\nNC|13.5|24.6|24.2|24.8\nABL|10.8|31.3|31.5|30.7\n\nAccording to this table, when the surrogate model is trained for few epochs (20), a good estimation for the boundary cannot be obtained, thus the performance is much decreased; when trained for more epochs (80), the performance is similar to that at 60 epochs.\n\nSecond, many existing attacking methods, such as hidden-trigger[2] and clean-label[3], utilize a surrogate model. When applying the proposed method to these attacks, surrogate models of these attacks can be directly leveraged for selection, thus avoiding inefficiency. Moreover, for some large datasets such as ImageNet, pre-trained models exist that can be directly used by the proposed method. The authors conduct experiments on Tiny-ImageNet where the surrogate model is trained for 60 epochs within 2 hours. The paper’s method consistently improves the attack’s resistance against various defenses.", "category": "Method_Disambiguation"}, {"question": "How does the proposed method (CBS) perform compared to the I-BAU[1] defense and random sampling in terms of success rate (ASR) and clean accuracy (Clean Acc) on the CIFAR-10 dataset and ResNet18 model?\n\n[1] Zeng, Yi, et al. \"Adversarial Unlearning of Backdoors via Implicit Hypergradient.\" International Conference on Learning Representations. 2021.", "answer": "The proposed method (CBS) is compared to random sampling (Random) and the I-BAU defense on the CIFAR-10 dataset and ResNet18 model. The results show that CBS achieves lower success rates (ASR) and comparable or slightly better clean accuracy (Clean Acc) across various attacks (BadNet, Blend, WaNet, and Hidden-trigger) compared to random sampling. Additionally, CBS is tested alongside another defense, Neural Cleanse[2], for comparison. While I-BAU is more powerful than other defenses[3], it compromises clean performance. However, CBS improves the stealthiness of the backbone methods under this powerful defense.\n\n| Attacks        | Random |     | CBS  |     |\n|----------------|--------|-----|------|-----|\n|                | ASR    | Clean Acc | ASR | Clean Acc |\n| **I-BAU**      |        |     |      |     |\n| BadNet         | 2.3    | 91.3 | 9.5  | 91.4 |\n| Blend          | 3.2    | 90.7 | 8.6  | 90.5 |\n| WaNet          | 1.7    | 90.5 | 8.2  | 90.6 |\n| Hidden trigger | 0.8    | 90.6 | 8.1  | 90.4 |\n| **NC**         |        |     |      |     |\n| BadNet         | 1.1    | 93.5 | 24.6 | 93.1 |\n| Blend          | 82.5   | 93.7 | 81.7 | 94.0 |\n| WaNet          | 8.9    | 94.1 | 13.4 | 93.8 |\n| Hidden trigger | 6.3    | 92.7 | 8.7  | 93.5 |    \n\n[2]Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks\n\n[3] SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics", "category": "Method_Comparison"}, {"question": "How does the proposed method perform against outlier-detection defenses compared to other backdoor defense methods?", "answer": "The proposed method improves resistance against both outlier-detection defenses (e.g., Spectral Signature[1], Activation Clustering[2], SPECTRE[3]) and non-outlier-detection defenses (e.g., Neural Cleanse[4], Fine Pruning[5], Anti-Backdoor Learning[6]), and no drop in performance against non-outlier-detection defenses is observed. For example, in Table 1, BadNet+CBS achieves a success rate of 24.6% against Neural Cleanse (a non-outlier-detection defense) and 23.7% against STRIP (an outlier-detection defense).\n\n[1] Spectral Signatures in Backdoor Attacks\n\n[2] Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering\n\n[3] SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics\n\n[4] Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks\n\n[5] Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks\n\n[6] Anti-Backdoor Learning: Training Clean Models on Poisoned Data", "category": "Method_Comparison"}, {"question": "Do the visualizations in Figure 1 for dirty-label attacks also apply to clean-label attacks?", "answer": "Yes, similar observations apply to clean-label attacks, as demonstrated by Clean-label attack in [1]. Visualizations for clean-label attacks will be included in the appendix. Additional evidence is provided in existing work [2], and as shown in the introduction and Figure 1 in [3], poison and clean samples consistently form two separate clusters in the latent space, even for the Clean-label attack[1].\n\n[1] Label-consistent backdoor attacks.\n\n[2] Anti-Backdoor Learning: Training Clean Models on Poisoned Data\n\n[3] Revisiting the Assumption of Latent Separability for Backdoor Defenses\n", "category": "Experimental_Analysis"}, {"question": "Why does the proposed method achieve lower ASRs compared to Random and FUS under the condition of 'No Defenses' in Tables 1, 2, and 3?", "answer": "The proposed method selects samples close to the decision boundary, which results in a more constrained selection space compared to random sampling. This targeted selection leads to a slightly reduced performance in the absence of defenses. High-confidence samples for their true classes, when used as poisoning samples, cause a significant shift in the decision boundary, typically leading to a higher success rate. Additionally, boundary samples with lower confidence have a smaller success rate than samples further from the boundary, as theoretically illustrated in Section 4.3, which indicates that our methods slightly compromise the performance without defenses.. This indicates a trade-off between clean performance and stealthiness, with the proposed method being stronger against defenses.", "category": "Experimental_Analysis"}, {"question": "Under what practical scenarios does the attacker have 'edit access' to the entire training set, and how does this assumption impact the proposed method?", "answer": "The attacker can have 'edit access' to the entire training set in scenarios where the attacker is a model provider or a training platform provider, as described in scenarios 1 and 2 in [1]. In such cases, the attacker can modify any data in the training set as desired, allowing the proposed method to be adapted for selection to improve the stealthiness of the backdoor. However, the proposed method does not necessarily assume full edit access. When the attacker is only allowed to insert triggers to part of the training data, they can still select samples from their own data using the proposed method while keeping other users' data clean. This was validated through an additional experiment on CIFAR-10, where 10% of the training set was allocated as the attacker's private data, and the attacker was not allowed to perturb the remaining 90%. It is assumed that the attacker selects 10% from their own data to insert triggers. To apply the paper’s method, a ResNet18 is trained as the surrogate model, and samples are selected from the private data. The poisoning effect is evaluated on both ResNet18 and VGG16. The BadNet attack and defenses such as Spectral Signature (SS), Strip, and Neural Cleanse (NC) are considered. The results (success rates), detailed in the following Table, suggest that the paper’s approach can indeed enhance the stealth of backbone attacks, even when the full training set is not accessible.\n\n| Defenses    | Model   | Random | CBS  | \n|-------------|---------|--------|------| \n| No defenses | ResNet18| 99.9   | 93.7 | \n|             | VGG16   | 99.8   | 95.2 | \n| SS          | ResNet18| 1.2    | 15.7 | \n|             | VGG16   | 3.5    | 10.4 | \n| NC          | ResNet18| 2.7    | 14.4 | \n|             | VGG16   | 3.6    | 12.8 | \n| Strip       | ResNet18| 1.5    | 13.2 | \n|             | VGG16   | 1.3    | 10.9 |\n\n[1] Backdoor Learning: A Survey", "category": "Experimental_Analysis"}, {"question": "Under what practical scenarios does the attacker have access to a surrogate model, and is the use of surrogate models in poisoning attacks a common and effective practice?", "answer": "**First**, there are practical scenarios where surrogate models are available. When the attacker is a model provider or a training platform provider, he has access to the model information and thus can easily utilize a surrogate model for selection via the paper's proposed method. **Second**, it is a common assumption to have a surrogate model. The previous works in [1][2][3] also consider poisoning attacks in the same setting to leverage a surrogate model for generating poisoning samples. Notably, the architecture or parameters of the surrogate model do not need to be the same as the victim model. Table 1 and 2 present experiments where poisons are generated from a ResNet18 model, and the paper tests the poisoning performance on the victim model VGG16. These results show that poisons can be transferred to different models. Existing work such as [4] also provides empirical evidence for the effectiveness of surrogate models under the black-box setting. In Table 4 in [4], poisons are generated from the surrogate model ResNet18 and transferred to MobileNet-V2, VGG11. The poisons are still effective. Therefore, the usage of surrogate models is practical.\n\n[1] Hidden Trigger Backdoor Attacks\n\n[2] Label-Consistent Backdoor Attacks\n\n[3] Backdoor Attack with Imperceptible Input and Latent Modification\n\n[4]Sleeper Agent: Scalable Hidden Trigger Backdoors for Neural Networks Trained from Scratch", "category": "Experimental_Analysis"}, {"question": "What types of defenses are included in the evaluation, and how does the proposed method perform against them?", "answer": "The evaluation includes both outlier-based and non-outlier-based defenses, such as NC, ABL, and FP. The proposed method improves the attacks’ resistance against both types of defenses without a drop in performance against non-outlier-detection defenses. Detailed results are presented in the main paper (Tables 1-3) and the Appendix (Tables 4-7).", "category": "Experimental_Exposition"}, {"question": "Why was the defense method mentioned in [1] not implemented, and which defense method was focused on instead?\n\n[1] Rethinking Backdoor Attacks\n", "answer": "The defense method mentioned in [1] was not implemented due to its complexity, which requires training approximately 100,000 models on randomly selected 50% subsets of the poisoned dataset, making the process exceedingly time-intensive. Instead, the focus was on the defense proposed in [2], which was tested on CIFAR-10 with model ResNet18, evaluating the performance of attacks BadNet and Clean-label attack, and comparing the proposed method with random sampling.\n\n[2] Incompatibility Clustering as a Defense Against Backdoor Poisoning Attacks", "category": "Method_Disambiguation"}, {"question": "What are the results of the proposed method compared to random sampling under the defense proposed in [2]?\n\n[2] Incompatibility Clustering as a Defense Against Backdoor Poisoning Attacks", "answer": "Under the defense proposed in [2], which was tested on CIFAR-10 with model ResNet18, evaluating the performance of attacks BadNet and Clean-label attack, and comparing the proposed method with random sampling. The proposed method (CBS) significantly improves the stealthiness of backbone methods compared to random sampling. The results show higher success rates (ASR) and comparable or slightly improved clean accuracy (Clean Acc) for both BadNet and Clean-label attacks. For example, for ISPL+B with BadNet, the ASR increased from 0.3 with random sampling to 13.1 with CBS, while the Clean Acc remained stable at 93.3.\n\n| Attacks |      | Random |       | CBS   |       | \n|---------|------|--------|-------|-------|-------| \n|         |      | ASR    | Clean Acc | ASR  | Clean Acc | \n| ISPL+B  | BadNet | 0.3    | 93.1    | 13.1 | 93.3      | \n|         | LC     | 0.9    | 92.2    | 10.3 | 92.4      | \n| NC      | BadNet | 1.1    | 93.5    | 24.6 | 93.1      | \n|         | LC     | 8.9    | 92.7    | 12.6 | 92.6      |", "category": "Method_Comparison"}, {"question": "What is the reason for the observed performance drop in the absence of defenses, as shown in Table 1?", "answer": "The performance drop occurs because the method selects samples close to the decision boundary, resulting in a more constrained selection space compared to random sampling. This targeted selection leads to a slightly reduced performance in the absence of defenses. Additionally, high-confidence samples used as poisoning samples cause a significant shift in the decision boundary, typically increasing the success rate. Theoretically, boundary samples (with lower confidence) have a smaller success rate than samples further from the boundary (with higher confidence), indicating a slight compromise in performance without defenses. This reflects a trade-off between clean performance and stealthiness, with the method being stronger against defenses.", "category": "Experimental_Analysis"}, {"question": "Has the method been evaluated on larger datasets such as Tiny-ImageNet, and what are the results?", "answer": "Yes, the method has been evaluated on Tiny-ImageNet. The results, presented in detailed tables, show that CBS consistently improves the stealthiness of different attacks across various defenses and attack types.\n\nPerformance on Type I backdoor attacks:\n|               | Attacks        | random | CBS  |\n|---------------|----------------|--------|------|\n| No defenses   | BadNet         | 89.5   | 83.1 |\n|               | Blended        | 83.4   | 81.6 |\n|               | Adaptive-Blend | 67.2   | 66.2 |\n|               | Adaptive-Patch | 84.5   | 81.7 |\n| SS            | BadNet         | 0.4    | 18.5 |\n|               | Blended        | 37.2   | 46.3 |\n|               | Adaptive-Blend | 59.4   | 65.1 |\n|               | Adaptive-Patch | 75.3   | 78.5 |\n| Strip         | BadNet         | 0.6    | 12.2 |\n|               | Blended        | 46.2   | 54.6 |\n|               | Adaptive-Blend | 58.4   | 63.3 |\n|               | Adaptive-Patch | 69.5   | 72.8 |\n\nPerformance on Type II backdoor attacks:\n|               | Attacks        | random | CBS  |\n|---------------|----------------|--------|------|\n| No defenses   | Hidden-trigger | 59.7   | 54.3 |\n| NC            | Hidden-trigger | 5.3    | 11.5 |\n| FP            | Hidden-trigger | 8.4    | 12.1 |\n| ABL           | Hidden-trigger | 1.8    | 4.2  |\n\nPerformance on Type III backdoor attacks:\n|               | Attacks        | random | CBS  |\n|----------------|---------|-------|------|\n| No defenses | WaNet | 98.5 | 96.1 |\n|             | LiRA  | 99.3 | 96.4 |\n|             | WB    | 98.2 | 92.7 |\n| NC          | WaNet | 5.4  | 10.7 |\n|             | LiRA  | 6.3  | 10.2 |\n|             | WB    | 9.6  | 15.1 |\n| FP          | WaNet | 9.5  | 13.6 |\n|             | LiRA  | 8.7  | 13.8 |\n|             | WB    | 10.2 | 16.5 |", "category": "Experimental_Exposition"}, {"question": "Has the method been evaluated on defenses that do not depend on outlier detection, and what are the results?", "answer": "Yes, the method has been evaluated on defenses that do not depend on outlier detection, including Neural Cleanse[3], Fine Pruning[4], Anti-Backdoor Learning[5]. The results, presented in the main text (Table 1-3) and the Appendix (Table 4-7), show that the method effectively improves the attacks' resistance against non-outlier-detection defenses. Additional tests on CIFAR-10 with ResNet18, comparing the method with random sampling, demonstrate significant improvements in stealthiness under these defenses.\n\n| Attacks |      | Random |       | CBS   |       |\n|---------|------|--------|-------|-------|-------|\n|         |      | ASR    | Clean Acc | ASR  | Clean Acc |\n| ISPL+B  | BadNet | 0.3    | 93.1    | 13.1 | 93.3      |\n|         | LC     | 0.9    | 92.2    | 10.3 | 92.4      |\n| NC      | BadNet | 1.1    | 93.5    | 24.6 | 93.1      |\n|         | LC     | 8.9    | 92.7    | 12.6 | 92.6      |\n\n[3] Neural Cleanse: Identifying and Mitigating Backdoor Attacks in Neural Networks\n\n[4] Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks\n\n[5] Anti-Backdoor Learning: Training Clean Models on Poisoned Data", "category": "Experimental_Exposition"}, {"question": "How is the selection of poisoned samples performed using the FUS technique?", "answer": "FUS[2]  is a method to improve the efficiency of poisoning by a rational selection of poisoned samples. The selection strategy is based on Sections 2.2, 5.1, 5.2, 5.3, 5.4 of the original paper, where FUS is compared with random sampling for fixed poison rates (referred to as mixing ratio in the paper). In this case, FUS is compared with random sampling and CBS for fixed poisoned rates.\n[2] Data-Efficient Backdoor Attacks \n\n\n", "category": "Method_Mechanics"}, {"question": "Is it a common assumption in previous works[1][2][3] to use a surrogate model for generating poisoning samples in poisoning attacks?\n\n[1] Hidden Trigger Backdoor Attacks\n\n[2] Label-Consistent Backdoor Attacks\n\n[3] Backdoor Attack with Imperceptible Input and Latent Modification\n", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the method assume the attacker always has full edit access to the entire training set?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does existing work[4] provide empirical evidence for the effectiveness of surrogate models under black-box settings?\n\n[4]Sleeper Agent: Scalable Hidden Trigger Backdoors for Neural Networks Trained from Scratch", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the table include both success rate (ASR) and clean accuracy (Clean Acc) for comparing the method with random sampling under different attacks and defenses?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was an experiment conducted to validate the method's feasibility when the attacker only has access to a subset of the training data?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "fvhJu0FODp", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "Avalon's Game of Thoughts: Battle Against Deception through Recursive Contemplation", "authors": ["Shenzhi Wang", "Chang Liu", "Zilong Zheng", "Siyuan Qi", "Shuo Chen", "Qisen Yang", "Andrew Zhao", "Chaofei Wang", "Shiji Song", "Gao Huang"], "abstract": "Recent breakthroughs in large language models (LLMs) have brought remarkable success in the field of LLM-as-Agent. Nevertheless, a prevalent assumption is that the information processed by LLMs is consistently honest, neglecting the pervasive deceptive or misleading information in human society and AI-generated content. This oversight makes LLMs susceptible to malicious manipulations, potentially resulting in detrimental outcomes. This study utilizes the intricate Avalon game as a testbed to explore LLMs' potential in deceptive environments. Avalon, full of misinformation and requiring sophisticated logic, manifests as a \"Game-of-Thoughts\". Inspired by the efficacy of humans' recursive thinking and perspective-taking in the Avalon game, we introduce a novel framework, Recursive Contemplation (ReCon), to enhance LLMs' ability to identify and counteract deceptive information. ReCon combines formulation and refinement contemplation processes; formulation contemplation produces initial thoughts and speech, while refinement contemplation further polishes them. Additionally, we incorporate first-order and second-order perspective transitions into these processes respectively. Specifically, the first-order allows an LLM agent to infer others’ mental states, and the second-order involves understanding how others perceive the agent’s mental state. After integrating ReCon with different LLMs, extensive experiment results from the Avalon game indicate its efficacy in aiding LLMs to discern and maneuver around deceptive information without extra fine-tuning and data. Finally, we offer a possible explanation for the efficacy of ReCon and explore the current limitations of LLMs in terms of safety, reasoning, speaking style, and format, potentially furnishing insights for subsequent research.", "keywords": "Large language model;Avalon game;strategic communnication;deception", "qa_pairs": [{"question": "Why is the statement 'we use ReCon for the good side' used instead of 'we use CoT for the good side,' and how does this choice impact the experimental outcomes?", "answer": "The statement 'authors use ReCon for the good side' is intentional. Authors initially tested the good side using the CoT method against the evil side with ReCon, but due to the evil side's advantage, the good side consistently lost, resulting in uniform results. To achieve more discriminative outcomes, authors used ReCon on the good side when assessing methods for the evil side.", "category": "Experimental_Exposition"}, {"question": "Why were 20 Avalon game rounds selected for the experiment, and is the reliability of the results supported by statistical tests?", "answer": "The rationale for selecting 20 Avalon game rounds is explained in `Global Q1`, and a statistical test of the results is provided in `Global Q2`. For more details, refer to the Global Responses (https://openreview.net/forum?id=fvhJu0FODp&noteId=Y7twEtlQGy).", "category": "Motivation_Analysis"}, {"question": "How do authors ensure the reliability and reproducibility of the MULTI-DIMENSIONAL EVALUATION results when only GPT-4 is used for evaluation?", "answer": "To ensure reliability, automated evaluation of LLMs is widely recognized and popular [1]. authors analyzed over 2300 LLM responses using GPT-4, which enhances the evaluation's reliability. Authors also acknowledge the value of human evaluation and are conducting a human agreement evaluation using a method similar to AlpacaEval, with results to be published soon. For reproducibility, authors set the GPT-4 evaluator's temperature to 0 and documented the GPT-4 version in the paper's 'Reproducibility Statement.' Evaluating over 2300 responses also reduces randomness under the law of large numbers, making the GPT-4 evaluation reproducible.\n[1] Chang, Yupeng, et al. 'A survey on evaluation of large language models.' arXiv preprint arXiv:2307.03109 (2023).", "category": "Method_Mechanics"}, {"question": "How does the ReCon method address the limitations of LLMs in understanding and responding to prompts, particularly in terms of utilizing a theory of mind for other agents in the game?", "answer": "Authors acknowledge the viewpoint that LLMs are fundamentally linguistic and that higher-order cognitive processes are emergent properties. However, authors wish to clarify two critical points: 1. Having the capability does not necessarily equate to achieving potential. For instance, while any cognitively sound individual possesses the intelligence to play the Avalon game, the quality of play varies significantly without the right reasoning mechanisms, even for the same person. In the realm of LLMs, their emergent nature only ensures the capability to engage in activities like playing Avalon but does not guarantee reaching the pinnacle of gameplay potential. It is here that thinking mechanisms (e.g., Chain-of-Thought [1, 2], Tree-of-Thought [3], Graph-of-Thought [4], Reflexion [5], ReAct [6], Voyager [7], and  proposed ReCon) play a pivotal role; they are designed to unleash the potential of LLMs in complex scenarios. 2. It is an oversimplification to categorize all methods developing thinking mechanisms (such as [1-7], and  proposed ReCon) as mere 'prompt engineering.' As emphasized earlier, the essence of these methods lies in the insights derived from the underlying thinking processes, not the prompts themselves. To label these innovative algorithms and ReCon as just 'mumbling different incantations at a mysterious creature to see how it responds'—for example, understanding Chain-of-thought as just adding an incantation like 'let's think step by step' [2]—would unfairly dismiss their value and the significant advancements they represent. This perspective overlooks the strategic and thoughtful design embedded in these mechanisms, which are crucial for advancing understanding and application of LLMs in complex scenarios. Additionally, the paper's experiments show that current cognitive methods for LLMs, like CoT, are ineffective in deceptive environments such as Avalon, whereas the paper's ReCon method proves successful. The success of ReCon is not due to brute-force attempts with various prompts but rather stems from cognitive insights related to the theory of mind [8] and communicative theory [9].\n[1] Wei, Jason, et al. \"Chain-of-thought prompting elicits reasoning in large language models.\" Advances in Neural Information Processing Systems 35 (2022): 24824-24837.\n\n[2] Kojima, Takeshi, et al. \"Large language models are zero-shot reasoners.\" Advances in neural information processing systems 35 (2022): 22199-22213.\n\n[3] Yao, Shunyu, et al. \"Tree of thoughts: Deliberate problem solving with large language models.\" arXiv preprint arXiv:2305.10601 (2023).\n\n[4] Besta, Maciej, et al. \"Graph of thoughts: Solving elaborate problems with large language models.\" arXiv preprint arXiv:2308.09687 (2023).\n\n[5] Shinn, Noah, et al. \"Reflexion: Language agents with verbal reinforcement learning.\" Thirty-seventh Conference on Neural Information Processing Systems. 2023.\n\n[6] Yao, Shunyu, et al. \"React: Synergizing reasoning and acting in language models.\" arXiv preprint arXiv:2210.03629 (2022).\n\n[7] Wang, Guanzhi, et al. \"Voyager: An open-ended embodied agent with large language models.\" arXiv preprint arXiv:2305.16291 (2023).\n\n[8] Yuan, Luyao, et al. \"In situ bidirectional human-robot value alignment.\" Science robotics 7.68 (2022): eabm4183.\n\n[9] Tomasello, Michael. Origins of human communication. MIT press, 2010.", "category": "Method_Mechanics"}, {"question": "How does the ReCon method address concerns about its generalizability to future LLMs with different training sets, architectures, or fine-tuned guardrails?", "answer": "ReCon was evaluated on two distinct LLMs, ChatGPT and Claude, which differ in training sets, architectures, and fine-tuned guardrails. Identical thinking processes and prompts were applied, showing superior outcomes in both, suggesting generalizability  (Figure 4 of the paper). The thought processes in ReCon, such as recursive thinking and perspective-taking, are also relevant to human intelligence, indicating potential effectiveness even as LLMs evolve. The only exception might be future LLMs with inferior generative abilities.", "category": "Method_Mechanics"}, {"question": "Is the choice of 20 Avalon game rounds as a test size sufficient for demonstrating the efficacy of the ReCon framework and its components?", "answer": "Yes, the choice of 20 Avalon game rounds as a test size is sufficient. Each game encompasses numerous dialogues, making this a considerable volume of data. A study recently accepted by EMNLP [1] introduced a dataset containing 20 Avalon games, which is equivalent to the number of test rounds per method in the study. Additionally, the experimental costs amounting to around $20,000 further emphasize the scale of the 20 Avalon games. The high success rate depicted in Figure 4 of the paper, coupled with thorough statistical tests on authors' success rate improvement, substantiates this claim.\n[1] Stepputtis, Simon, et al. 'Long-Horizon Dialogue Understanding for Role Identification in the Game of Avalon with Large Language Models'. Findings of the Association for Computational Linguistics: EMNLP, 2023.\n", "category": "Experimental_Analysis"}, {"question": "What statistical tests were applied to evaluate the effectiveness of ReCon compared to CoT, and what were the results?", "answer": "Authors applied Barnard's test [1] to the results in Figure 4 to determine if there is a statistically significant performance difference between the methods (ReCon and its ablated versions) and CoT. The justification for selecting Barnard's test is its suitability for analyzing 2x2 contingency tables, as described in [Wikipedia](https://en.wikipedia.org/wiki/Barnard's_test). This is relevant for the comparison of methods with CoT, where the numbers of successes and failures in each method form such contingency tables, making Barnard's test an appropriate choice. The outcomes, specifically the p-values, are presented in the following table. P-values less than 0.05, indicating statistical significance, are highlighted with an asterisk '*'. \n\n| p-value | ChatGPT | Claude | Evil | \n| -------------------------------------------- | -------------------------- | ---------- | ---------- | \n| ReCon w/o First-Order Perspective Transition | $0.0321^*$ | $0.6497$ | $0.0114^*$ |\n | ReCon w/o Second-Order Pespective Transition | $0.0427^*$ | $0.3427$ | $0.0074^*$ | \n| ReCon w/o Refinement Contemplation | $0.0183^*$ | $0.0830$ | $0.5484$ | \n| ReCon w/o Formulation Contemplation | $0.0005^*$ | $0.5324$ | $0.0074^*$ | \n| ReCon (authors) | ${3.4782\\times 10^{-5}}^*$ | $0.0484^*$ | $0.0016^*$ | \n\nThe table above reveals that, except for their performance on Claude, nearly all the methods demonstrate statistically significant superiority over CoT, underscoring the effectiveness of approaches in ReCon. Moreover, to further confirm the robustness of ReCon's designs on Claude, authors increased the number of test games from approximately 20 to about 40. This expansion aims to assess whether ReCon and its variations continue to outperform CoT with a larger set of test games. The results, displayed in the subsequent table, indicate that while success rates may vary with the number of games, the findings observed with around 20 games remain consistent when increased to approximately 40 games. This consistency affirms the reliable effectiveness of ReCon's designs in the context of Claude.\n\n | Claude's success rate (v.s. baseline CoT) | ~20 rounds | ~40 rounds |\n | -------------------------------------------- | ---------- | ---------- | \n| Baseline | $47.4$% | $45.0$% |\n | ReCon w/o First-Order Perspective Transition | $55.0$% | $52.6$% | \n| ReCon w/o Second-Order Pespective Transition | $63.2$% | $60.0$% |\n | ReCon w/o Refinement Contemplation | $75.0$% | $63.9$% |\n | ReCon w/o Formulation Contemplation | $57.9$% | $57.9$% |\n | ReCon (authors) | $78.9$% | $73.7$% |\n\n[1] Fisher, Ronald Aylmer. \"A new test for 2×2 tables.\" Nature 156.3961 (1945).", "category": "Experimental_Exposition"}, {"question": "How effective is the proposed method when dealing with a dataset that is not exchangeable?", "answer": "The test's validity depends on data exchangeability, but datasets can often be adjusted slightly to retain test applicability. For example, non-exchangeability due to ascending IDs (e.g., in SQuAD and HumanEval) can be addressed by either removing these IDs or permuting the examples while keeping IDs constant. ", "category": "Method_Disambiguation"}, {"question": "What is the justification for choosing n=20 as the sample size in the paper's evaluation, and have authors conducted statistical significance tests to validate the results?", "answer": "The choice of 20 Avalon game rounds as a test size is substantial. Each game encompasses numerous dialogues, making this a considerable volume of data. A study recently accepted by EMNLP introduced a dataset containing 20 Avalon games, equivalent to the number of test rounds per method in the study. Additionally, the experimental costs amounted to around $20,000, emphasizing the scale of the 20 Avalon games. The 20 Avalon game rounds are adequate for demonstrating the efficacy of the ReCon framework and its components. Authors applied Barnard's test to the results in Figure 4 to determine if there is a statistically significant performance difference between the methods (ReCon and its ablated versions) and CoT. P-values less than 0.05, indicating statistical significance, are highlighted with an asterisk '*', showing that nearly all the methods demonstrate statistically significant superiority over CoT.", "category": "Motivation_Analysis"}, {"question": "What are the key differences between the ReCon method and the Tree of Thoughts (ToT) approach, particularly in terms of core motivation, algorithmic focus, and nature of reasoning trace?", "answer": "The key differences between ReCon and ToT are as follows: 1. Core Motivation: ToT is designed for tasks like the Game of 24, Creative Writing, and Mini Crosswords, which do not involve misinformation or deception and do not occur in communicative contexts. ReCon is developed to manage deceptions and misinformation in communicative environments, where there may not be a definitive 'correct' answer, emphasizing the nuances of communication. 2. Algorithmic Focus: ToT focuses on constructing complex reasoning trees with looking ahead or backtracking to address problems, while ReCon prioritizes Theory of Mind  (such as understanding and speculating about others' thoughts and identities) and communicative skills (such as expressing opinions while concealing its own identity when necessary) without relying on complex reasoning trees. 3. Nature of Reasoning Trace: ToT's reasoning trace consists of decomposed, intermediate thought steps aimed at problem resolution, whereas ReCon's reasoning trace involves original version of thoughts and spoken contents tailored for enhanced Theory of Mind and communicative skills.", "category": "Method_Comparison"}, {"question": "What is the reason for the differing performance between ReCon without refinement contemplation and ReCon without formulation contemplation when applied to Claude versus ChatGPT, as observed in Figure 4?", "answer": "The discrepancy in Figure 4 is likely due to the different abilities of LLMs to hide internal thoughts from their spoken outputs. Empirical analysis suggests that ChatGPT, particularly GPT-4, is more adept at concealing internal thoughts than Claude. As a result, refinement contemplation based on GPT-4 is more effective for hiding internal thoughts in spoken content, while formulation contemplation is more critical for ReCon when implemented with Claude. Thus, in Figure 4, the relative magnitude of the performance without refinement contemplation (the green bin) compared to the performance without formulation contemplation (the red bin) is different between ChatGPT and Claude.", "category": "Experimental_Analysis"}, {"question": "How does the performance of Chain-of-Thought (CoT) compare to ReCon when applied to larger language models, particularly in scenarios involving deceptive cognitive processes and complex logic?", "answer": "In scenarios involving deceptive cognitive processes and complex logic, such as in the Avalon game, ReCon is expected to outperform Chain-of-Thought (CoT). This is because an LLM agent using CoT exposes all its thoughts and lacks perspective-taking abilities, whereas an LLM agent with ReCon can separate internal thoughts from spoken content and possesses perspective-taking capabilities, leading to more comprehensive thinking.", "category": "Method_Comparison"}, {"question": "How would the consistency of the results change if human players, rather than Chain-of-Thought (CoT) methods, were used in the Avalon game?", "answer": "Current large language models (LLMs) using the ReCon method may not yet match the performance of expert human players due to inherent limitations and the lack of a comprehensive Avalon game dataset for fine-tuning. The performance gap between novice and experienced human players is substantial, making results less consistent compared to the more consistent performance of GPT-4. However, the paper demonstrates significant improvements using ReCon in the Avalon game compared to CoT methods. ", "category": "Experimental_Exposition"}, {"question": "Did the statistical tests applied to the results in Figure 4 show statistically significant performance differences between ReCon and CoT, except for performance on Claude?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the reliability of the results from 20 Avalon game rounds statistically tested and explained in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the 20 Avalon game rounds equivalent in volume to a complete dataset used in a study accepted by EMNLP?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the number of 20 test rounds considered sufficient for demonstrating the efficacy of the ReCon framework based on the volume of data and comparison to a similar study?", "answer": "True", "category": "Claim_Verification"}, {"question": "Do the results suggest that ReCon's methods are likely to remain effective for future LLMs that approach human intelligence?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the sample size of 20 Avalon game rounds considered insufficient for demonstrating the efficacy of the ReCon framework?", "answer": "False", "category": "Claim_Verification"}, {"question": "Were measures taken to ensure reproducibility, such as setting the GPT-4 evaluator's temperature to 0 and documenting the GPT-4 version?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the study include a suitable statistical significance test, such as Barnard's test, to support the performance claims of the ReCon framework?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the multi-dimensional evaluation in section 4.2 conducted exclusively with ChatGPT, potentially affecting reliability and reproducibility?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "TOE6N8dp4w", "venue": "ICLR 2024", "paper_decision": "Reject", "title": "Harnessing large-language models to generate private synthetic text", "authors": ["Alexey Kurakin", "Natalia Ponomareva", "Umar Syed", "Liam MacDermed", "Andreas Terzis"], "abstract": "Differentially private training algorithms like DP-SGD protect sensitive training data by ensuring that trained models do not reveal private information. An alternative approach, which this paper studies, is to use a sensitive dataset to generate synthetic data that is differentially private with respect to the original data, and then non-privately training a model on the synthetic data.  Doing so has several advantages: synthetic data can be reused for other tasks (including for hyper parameter tuning), retained indefinitely, and shared with third parties without sacrificing privacy. \n\nHowever, generating private synthetic data is much harder than training a private model. To improve performance on text data, recent work has utilized public data by starting with a pre-trained generative language model and privately fine-tuning it on sensitive data. This model can be used to sample a DP synthetic dataset. While this strategy seems straightforward, executing it has proven problematic. Previous approaches either show significant performance loss, or have, as we show, critical design flaws.\n \nIn this paper we demonstrate that a proper training objective along with tuning fewer parameters results in excellent DP synthetic data quality. Our approach is competitive with direct DP-training of downstream classifiers in terms of performance on downstream tasks. Further, we demonstrate that our DP synthetic data is not only useful for downstream classifier training, but also to tune those same models.", "keywords": "Privacy;synthetic data;large language models", "qa_pairs": [{"question": "What is the novelty of the paper given that it combines existing techniques rather than proposing new ones?", "answer": "The novelty lies in how the existing techniques are combined and the results achieved: (1) The paper is the first to highlight the importance of data deduplication for achieving true differential privacy (DP) guarantees, addressing data contamination issues in prior work. (2) It is the first to demonstrate that synthetic data can be effectively used for validation and hyperparameter tuning, not just as a replacement for the downstream training set. (3) It shows consistent high-quality synthetic data generation across multiple datasets using LoRa-tuning with DP, unlike prior work where results were inconsistent.", "category": "Method_Disambiguation"}, {"question": "What are the observed differences in accelerator memory requirements between full finetuning, LoRa, and prompt tuning?", "answer": "For full finetuning and LoRa, a similar batch size of 32 examples can fit into accelerator memory, whereas about 256 examples can fit for prompt tuning. This indicates that finetuning and LoRa have similar memory requirements, while prompt tuning has lower memory requirements. In all setups, an effective batch size of 1024 or more examples per step is used, necessitating gradient accumulation by splitting the batch into smaller chunks of 32 or 256 examples, computing gradients sequentially, and then aggregating them.", "category": "Experimental_Setup"}, {"question": "What are the novel contributions of this work, particularly in terms of combining existing techniques and achieving new results?", "answer": "The novel contributions of this work include: 1) Being the first to identify data deduplication as essential for achieving true DP-guarantees, addressing data contamination issues in prior work that invalidate DP claims. 2) Demonstrating for the first time that synthetic data can be effectively used as a validation set and for hyperparameter tuning, not just as a replacement for downstream model training sets. 3) Achieving high-quality and consistent synthetic data across multiple datasets using the proposed method of LoRa-tuning with DP, where prior work showed inconsistent results.", "category": "Method_Disambiguation"}, {"question": "Does the paper adequately address and clarify the contributions of prior work, and how does it differentiate its own contribution in the context of dataset contamination between LLM pre-training and synthetic data generation?", "answer": "The authors provided a detailed explanation of their comparison to (Yue et al) and other prior work, updated section 2 of the paper for clarity, and focused on the most important points. They acknowledge all prior work as important building blocks for DP-synthetic data but emphasize that none of the prior work addresses dataset contamination between LLM pre-training and synthetic data generation, as demonstrated in appendix D.", "category": "Method_Disambiguation"}, {"question": "What aspects of the proposed framework demonstrate novelty, given that its key components are already studied in prior work?", "answer": "The novelty of the proposed framework lies in how the components are combined and the results achieved: (1) It is the first to highlight that data deduplication is essential for achieving true DP-guarantees, addressing data contamination issues in prior work. (2) It is the first to use synthetic data effectively as a validation set and for hyperparameter tuning, not just as a replacement for downstream model training. (3) It demonstrates the production of high-quality synthetic data through LoRa-tuning with DP, with consistent results across multiple datasets.", "category": "Method_Disambiguation"}, {"question": "How does the paper's claim of significant performance loss in downstream algorithms due to DP synthetic text align with the findings of Yue et al. (2022)?", "answer": "While Yue et al. (2022) achieved good metrics on some datasets, they also reported a large drop in accuracy (~25% drop) of downstream classifiers on other datasets using DP synthetic data. The paper's reference to 'significant performance loss' pertains to this accuracy reduction, as detailed in Table 6 of Yue et al.'s work.", "category": "Experimental_Exposition"}, {"question": "What adjustments can be made to a non-exchangeable dataset to ensure the effectiveness of the method?", "answer": "Depending on the source of non-exchangeability, datasets can be adjusted to retain the test's applicability. For example, if the non-exchangeability is due to ascending IDs (e.g., in SQuAD and HumanEval), the data can be altered by removing these IDs or permuting the examples while keeping IDs constant. This adjustment ensures the method remains effective.", "category": "Experimental_Setup"}, {"question": "How does the current work differentiate its contribution from prior research on parameter-efficient tuning approaches like prompt tuning and LoRa tuning in the context of differential privacy (DP)?", "answer": "The current work differentiates its contribution by demonstrating that parameter-efficient tuning approaches, when combined with data deduplication and careful synthetic data generation, achieve consistent and high-quality results across multiple datasets. Additionally, it emphasizes addressing dataset contamination, which prior work has not adequately considered, as a critical factor in achieving true DP guarantees.", "category": "Experimental_Analysis"}, {"question": "Does the data de-duplication method detect duplications that occur in the middle of the samples, or does it only check the suffix of the samples?", "answer": "The de-duplication method detects all common substrings between two datasets, including substrings in the middle of the text. Authors use the same approach as https://arxiv.org/abs/2107.06499 and adapted their code https://github.com/google-research/deduplicate-text-datasets. It uses a suffix array-based algorithm to efficiently search for duplicates in large datasets, ensuring that all duplicates, not just suffixes, are identified.", "category": "Method_Mechanics"}, {"question": "What is the impact on performance when both the classifier generating DP synthetic data and the final downstream classifier have the same capacity and are exposed to the same public information?", "answer": "When both classifiers have the same capacity and use the same public information, DP on real data will likely be an upper bound on performance that any DP synthetic data with the same epsilon (eps) can achieve. In this scenario, using DP synthetic data may not be optimal for performance maximization. However, DP synthetic data offers other benefits, such as data sharing, evaluation, and hyperparameter tuning. Additionally, in cases where multiple downstream classifiers are trained using the same private data and revealed to the end user, DP synthetic data can be created once and reused for all models. This approach may perform better than training multiple models directly on DP real data, as the budget for DP on real data would need to be split among all classifiers.", "category": "Experimental_Analysis"}, {"question": "Why are there missing numbers in Table 1, and what steps were taken to address this issue?", "answer": "The missing numbers in Table 1 were due to incomplete experiments on AGNews at the time of ICLR submission, caused by high computational costs and limited resources. The authors omitted these experiments initially, expecting their results to be less significant compared to LoRa results. Subsequently, all three DP-finetuning experiments on AGNews were re-run with a smaller computational budget (10 epochs instead of 40), and the updated numbers were included in the paper. The accuracy of eps=1 full finetuning on AGNews also changed after this re-run.", "category": "Experimental_Analysis"}, {"question": "How does the study contribute to the field despite the use of well-known techniques like DP-SGD and parameter-efficient fine-tuning?", "answer": "The study's novelty lies in the unique combination of known techniques and the resulting achievements: 1) It is the first to emphasize the importance of data deduplication for true DP guarantees, addressing the data contamination issue in prior work. Technically, this would mean that DP-guarantee can no longer be claimed and epsilon values are infinity. 2) It demonstrates the effective use of synthetic data for validation and hyperparameter tuning, beyond just replacing downstream training sets. 3) It achieves high-quality and consistent synthetic data across multiple datasets using LoRa-tuning with DP, unlike the inconsistent results of previous methods.", "category": "Method_Mechanics"}, {"question": "What is the rationale behind choosing the hyperparameter value of 50 for the use of suffix arrays, and how does it relate to the removal of common sequences and the training process?", "answer": "The value of 50 tokens for the hyperparameter was chosen based on the suggestion from https://arxiv.org/abs/2107.06499. Sequences that are too short are likely to correspond to common words or phrases, so this approach removes any common sequence of 50 tokens or longer. For the training process, instances up to 512 tokens long are used, so the pretraining data that even partially matches 'private' instances is removed. This makes the deduplication stronger than simply removing exact matches or exact instances.", "category": "Experimental_Setup"}, {"question": "Why is the de-duplication method used in this study considered a stronger approach than simply removing datasets like IMDB, Yelp, and AGNews from The Pile?", "answer": "The de-duplication method is stronger because there is no straightforward way to remove datasets like IMDB, Yelp, and AGNews from The Pile. The Pile does not directly include examples from these datasets but contains web pages with content that overlaps with them. For example, IMDB includes sanitized movie reviews, while The Pile contains web pages from imdb.com with multiple reviews, often overlapping with IMDB. Since The Pile does not provide URL annotations, direct removal of 'imdb.com' domains is impossible. Instead, the de-duplication algorithm identifies and removes overlapping content by searching for common substrings between The Pile and the target datasets, ensuring a more precise removal of redundant data.", "category": "Method_Disambiguation"}, {"question": "How should the results in Table 3, which involve metrics such as RBO, Spearman, and Kendall, be interpreted, especially for readers unfamiliar with these metrics?", "answer": "The metrics in Table 3, such as RBO, Spearman, and Kendall, do not have absolute 'good' thresholds. Their interpretation depends on the context, such as the correlation level required for specific models. For example, a 70% correlation might be sufficient for some models but insufficient for others. The 'real data' row with eps inf serves as a useful comparison point. These metrics are best used for relative comparisons, as demonstrated in Table 3, where tuning on real data yields the best results (column Mean 25% accuracy), even though its correlation with downstream performance is not perfect (e.g., Spearman of 0.96 or RBO of 0.56). This discussion will be expanded in the text, space permitting.", "category": "Concept_Understanding"}, {"question": "What is the computational cost of LoRa fine-tuning compared to prompt tuning and full fine-tuning, and how does the cost of DP-fine tuning the 20M parameter LoRa part of the model relate to the forward pass through the 8B parameter pre-trained model?", "answer": "LoRa fine-tuning has a computational cost roughly 2x that of prompt tuning but is still much faster than full fine-tuning. The cost of DP-fine tuning the 20M parameter LoRa part of the model is primarily driven by the forward pass through the 8B parameter pre-trained model, which remains a significant factor in the overall computational expense.", "category": "Method_Comparison"}, {"question": "Is the computational cost of LoRa fine-tuning the same as that of full fine-tuning of the model?", "answer": "No, LoRa fine-tuning is much faster than full fine-tuning. The computational cost of LoRa fine-tuning is roughly 2x compared to prompt tuning in the experiments, making it still significantly faster than full fine-tuning. These details have been added to appendix M.", "category": "Method_Comparison"}, {"question": "What are the observed differences in memory requirements between full finetuning, LoRa, and prompt tuning for accelerator memory?", "answer": "For full finetuning and LoRa, a similar batch size of 32 examples can fit into accelerator memory, whereas for prompt tuning, about 256 examples can fit. This suggests that finetuning and LoRa have similar memory requirements, while prompt tuning has lower memory requirements. In all setups, an effective batch size of 1024 or more examples per step is used, thus authors have to use gradient accumulation technique to be able to train with such batch size. Authors split the entire batch into smaller chunks of either 32 or 256 examples, sequentially compute gradients for each chunk and then aggregate to compute final gradients for the entire batch.", "category": "Method_Comparison"}, {"question": "How do the authors justify their claim that their results indicate the possibility of obtaining good fidelity non-private synthetic data, contrary to the results reported in Yue et al. (2022) and Putta et al. (2023), given the similar performance gaps and model sizes used in prior work?", "answer": "The authors have refined their comparison to focus on the most impactful aspects of prior work and provided detailed explanations in a shared reply. They emphasize that while Yue et al. achieved good metrics on some datasets, they reported a large drop in accuracy (~25%) on others, which the authors refer to as 'significant performance loss' in their abstract.", "category": "Method_Mechanics"}, {"question": "How do the authors elaborate on the statement that the distribution of data conditioned on features like domain, sentiment, and category did not reflect the real data distribution in Yue et al. (2022), and what does 'augmenting the fine-tuning process' entail? Additionally, how do the authors address the potential issue of uniform label distribution in synthetic data when randomly selecting example labels during synthetic data generation?", "answer": "The authors have updated section 2 to clarify these points, focusing on the most impactful aspects of prior work. They acknowledge that Yue et al. (2022) used labels in the prefix, similar to their approach. However, they emphasize that their method ensures better fidelity of synthetic data by addressing data contamination issues and leveraging parameter-efficient fine-tuning techniques.", "category": "Method_Mechanics"}, {"question": "What was the methodology for hyperparameter search in the comparison between full fine-tuning and parameter efficient fine-tuning, and was the same comprehensive search applied to both approaches while still observing superior performance for parameter efficient fine-tuning?", "answer": "The hyperparameter search followed best practices from prior work (Li et al., 2021, and https://arxiv.org/abs/2303.00654). For all settings (full fine-tuning, prompt-tuning, and LoRA), a sweep was conducted over learning rate and clipping norm. Due to resource constraints, the clipping norm was tuned for a single value of epsilon and reused for others. For some full fine-tuning experiments, the granularity of the learning rate sweep was reduced (e.g., 1e-5, 1e-4, 1e-3, …) compared to parameter efficient tuning (1e-5, 3e-5, 1e-4, 3e-4, …). Despite training full fine-tuning for 40 epochs (compared to 10 epochs for LoRA and prompt tuning), full fine-tuning did not match the performance of parameter efficient tuning.", "category": "Experimental_Setup"}, {"question": "How does the paper address the issue of data contamination in prior work and ensure a fair comparison with existing methods in the context of DP-synthetic data generation?", "answer": "The paper acknowledges the importance of prior work in DP-synthetic data generation but highlights that their results should be viewed cautiously due to data contamination. Appendix D provides evidence of data contamination, such as the inclusion of an IMDB test set example in OpenWebText. Models that have seen 'private' data without proper clipping and noising have technically infinite epsilon guarantees. ", "category": "Method_Mechanics"}, {"question": "Did the re-running of the experiments on AGNews with fewer training epochs change the accuracy of eps=1 full finetuning?", "answer": "True", "category": "Claim_Verification"}, {"question": "Was the discussion on the technical details of training and sampling in Yue et al. removed from the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Did the paper originally misinterpret or downplay the contributions of prior work, leading to an overstated contribution?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the missing numbers in Table 1 due to incomplete experiments on AGNews at the time of ICLR submission?", "answer": "True", "category": "Claim_Verification"}, {"question": "Have the missing numbers in Table 1 been updated after re-running the DP-finetuning experiments on AGNews with a smaller computational budget?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does the paper claim that all prior work shows a significant loss in downstream algorithm performance with DP synthetic text, without acknowledging instances of good quality results?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the computational cost of LoRa fine-tuning equivalent to the full fine-tuning of the model?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the deduplication method only check for duplications at the suffix of the samples, without detecting those in the middle?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "GKau1ekOtH", "venue": "ICLR 2024", "paper_decision": "Withdraw", "title": "SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understanding", "authors": ["Haoxiang Wang", "Pavan Kumar Anasosalu Vasu", "Fartash Faghri", "Raviteja Vemulapalli", "Mehrdad Farajtabar", "Sachin Mehta", "Mohammad Rastegari", "Oncel Tuzel", "Hadi Pouransari"], "abstract": "The landscape of publicly available vision foundation models (VFMs), such as CLIP and Segment Anything Model (SAM), is expanding rapidly. VFMs are endowed with distinct capabilities stemming from their pre-training objectives. For instance, CLIP excels in semantic understanding, while SAM specializes in spatial understanding for segmentation. In this work, we introduce a simple recipe to efficiently merge VFMs into a unified model that assimilates their expertise. Our proposed method integrates multi-task learning, continual learning techniques, and teacher-student distillation. This strategy entails significantly less computational cost compared to traditional multi-task training from scratch. Additionally, it only demands a small fraction of the pre-training datasets that were initially used to train individual models. By applying our method to SAM and CLIP, we derive SAM-CLIP : a unified model that amalgamates the strengths of SAM and CLIP into a single backbone, making it apt for edge device applications. We show that SAM-CLIP learns richer visual representations, equipped with both localization and semantic features, suitable for a broad range of vision tasks. SAM-CLIP obtains improved performance on several head probing tasks when compared with SAM and CLIP. We further show that SAM-CLIP not only retains the foundational strengths of its precursor models but also introduces synergistic functionalities, most notably in zero-shot semantic segmentation, where SAM-CLIP establishes new state-of-the-art results on 5 benchmarks. It outperforms previous models that are specifically designed for this task by a large margin, including +6.8% and +5.9% mean IoU improvement on Pascal-VOC and COCO-Stuff datasets, respectively.", "keywords": "Foundation Model;Multi-Task Learning;Zero-Shot Learning;Semantic Segmentation", "qa_pairs": [{"question": "Is the proposed method limited to the sizes of released SAM models, and can it be applied to obtain models of arbitrary sizes?", "answer": "The proposed method is not limited to the sizes of released SAM models. SAM provides three ViT versions：ViT-B(ase), ViT-L(arge), and ViT-H(uge), and the method can be directly applied to ViT-L and ViT-H. Additionally, it can be extended to other model sizes, such as those in MobileSAM [1], for SAM-CLIP.\n[1] Zhang, Chaoning, et al. \"Faster Segment Anything: Towards Lightweight SAM for Mobile Applications.\" arXiv preprint arXiv:2306.14289 (2023).\n\n", "category": "Method_Disambiguation"}, {"question": "How was the dataset scale for Merged-41M determined, including the decision to include specific datasets like CC3M, CC12M, ImageNet-21k, and YFCC-15M? Additionally, how was the distillation loss value ratio of 1:10 for the SAM and CLIP heads empirically determined?", "answer": "The dataset scale for Merged-41M was determined by following the approach of EVA-CLIP[1], which used CC3M, CC12M, and ImageNet-21k. YFCC-15M, a subset of the 400M data used to train OpenAI’s CLIP, was also included. The distillation loss value ratio of 1:10 for the SAM and CLIP heads was empirically determined by testing three options: 1:1, 1:10, and 1:100. The ratio of 1:10 was chosen as it provided the best trade-off between mitigating the forgetting of SAM’s ability and learning CLIP’s ability.\n[1] Fang, Yuxin, et al. \"EVA: Exploring the limits of masked visual representation learning at scale.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.", "category": "Experimental_Setup"}, {"question": "What specific advantages does SAM-CLIP offer for edge device applications in terms of storage, memory, and compute costs, and how does it compare to using separate SAM and CLIP models?", "answer": "SAM-CLIP merges SAM and CLIP into a single ViT-B model, requiring only one ViT model for tasks that previously needed two models. This reduces the storage, memory, and compute costs, as SAM-CLIP only requires one forward pass per image, compared to two forward passes when using separate models. Existing benchmarks on devices are valid, as SAM-CLIP's performance is identical to SAM with ViT-B on the same device.", "category": "Method_Comparison"}, {"question": "Can the proposed merging technique, which is based on an embedding distillation loss and an unlabeled image dataset, be generalized to other vision foundation models (VFMs) beyond CLIP and SAM?", "answer": "Yes, the proposed merging technique can be utilized beyond CLIP and SAM with any image encoder. For example, one can merge DINOv2[1] with SAM (or SAM-CLIP) by substituting the CLIP teacher model with a DINOv2 model. In this paper, authors focus on SAM and CLIP because they are both zero-shot models with different capabilities, and merging them provides a vision model with broader zero-shot abilities. In contrast, DINOv2 is a pre-trained image backbone (with strong representation learning ability) that does not operate in a zero-shot manner. However, merging with DINOv2 could enhance representation learning and potentially offer some benefits. For example, one can also merge CLIP and DINOv2 using authors' pipeline (e.g., using CLIP as the base model and then applying multi-task distillation with CLIP and DINOv2 as teacher models), and possibly obtain a merged model with stronger representation suitable for certain downstream tasks.\n[1] Oquab, Maxime, et al. \"Dinov2: Learning robust visual features without supervision.\" arXiv preprint arXiv:2304.07193 (2023).\n\n", "category": "Method_Mechanics"}, {"question": "How does the training cost of SAM-CLIP compare to the training costs of SAM and CLIP, and what are the efficiency advantages of SAM-CLIP over the model composition approach for semantic segmentation?", "answer": "SAM-CLIP is trained on a 41M image dataset, which is 20-30x smaller than the training sets of SAM and CLIP. For instance, SAM is trained on SA-1B, and authors use only a 5.7% subset from SA-1B for the SAM self-distillation of our model; the state-of-the-art CLIP model is trained on DataComp-1B (1.4B image-text pairs), and the data authors use for CLIP distillation (40M unlabeled images) represents less than 3% of DataComp-1B. Hence, the training cost of SAM-CLIP is quite small compared with the training of SAM and CLIP. Regarding deployment, it has been shown that composing SAM and CLIP for semantic segmentation is feasible by using SAM to generate all possible segmentation masks and then using CLIP to provide labels [3]. However, this approach requires loading two models simultaneously (2x memory footprint) and, for each image, needs a forward pass of the SAM backbone (under 1024 resolution) to generate K object segments, followed by a forward pass of the CLIP model for each segment to filter (overall K+1 passes). With SAM-CLIP, only one ViT model needs to be loaded (lower memory footprint), and a single forward pass of the ViT backbone is required for each image. Overall, the method offers significant efficiency advantages over the model composition approach in terms of memory and computational costs during inference.\n[3] \"Grounded Segment-Anything.\" https://github.com/IDEA-Research/Grounded-Segment-Anything (2023)", "category": "Method_Comparison"}, {"question": "How does the proposed SAM-CLIP method compare in terms of efficiency and practicality to a simple cascading approach of SAM and CLIP for semantic segmentation?", "answer": "The SAM-CLIP method offers significant efficiency advantages over the simple cascading approach. This approach requires loading two models simultaneously (2x memory footprint) and, for each image, needs a forward pass of the SAM backbone (under 1024 resolution) to generate K object segments, followed by a forward pass of the CLIP model for each segment to filter (overall K+1 passes). SAM-CLIP only requires loading one ViT model and a single forward pass per image, reducing both memory and computational costs during inference.", "category": "Method_Comparison"}, {"question": "What are the advantages of SAM-CLIP over SAM with a text prompt in terms of zero-shot capabilities and the overall motivation behind SAM-CLIP?", "answer": "SAM-CLIP can perform all zero-shot tasks of CLIP, including zero-shot classification and image-text retrieval, which neither the public nor the unreleased versions of SAM can do. The motivation behind SAM-CLIP is to demonstrate the merging of two foundational models with different zero-shot capabilities into a single model, rather than just improving SAM or CLIP individually.", "category": "Method_Comparison"}, {"question": "Does the SAM-CLIP model suffer from catastrophic forgetting, and how does its segmentation performance compare to the original SAM model?", "answer": "The SAM-CLIP model exhibits mild performance degradation compared to SAM in instance segmentation, as shown in the last two columns of Table 1 for COCO and LVIS. This indicates that while SAM-CLIP does experience some degree of forgetting, it is not catastrophic. Additionally, segmentation quality comparisons between SAM and SAM-CLIP are provided in Appendix B of the updated manuscript.", "category": "Method_Comparison"}, {"question": "What is the performance gap between SAM-CLIP and the original SAM in terms of segmentation quality, and how does SAM-CLIP compensate for this gap?", "answer": "SAM-CLIP has a small performance gap compared to the original SAM in terms of instance segmentation ability, with -0.3% and -1.8% mAP on COCO and LVIS datasets, respectively, as shown in Table 1. However, SAM-CLIP compensates for this gap by offering broader zero-shot capabilities (Fig 1) and improved representation quality (Fig 4). Additionally, qualitative examples in Fig. 5 of Appendix A.1 demonstrate that the instance segmentation outputs of SAM and SAM-CLIP are of similar quality.", "category": "Experimental_Exposition"}, {"question": "Have the authors provided the segmentation outputs of the original SAM using the same examples shown in Figure 3 for a qualitative comparison?", "answer": "Figure 6 in Appendix A.1 shows that SAM segments only a sub-part of the object class pinpointed by the point prompts, highlighting the importance of combining CLIP and SAM abilities in SAM-CLIP for semantic segmentation.", "category": "Experimental_Setup"}, {"question": "Are the predicted masks by SegCLIP in Table 2 refined by SAM, and can the authors provide the zero-shot semantic segmentation results without using geometric prompts?", "answer": "The performance of SAM-CLIP reported in Table 2 is obtained using only the CLIP-head, without any SAM-head refinement — for a fair comparison, the input images to SAM-CLIP, similar to those in the baseline methods (shown in Table 2), are resized to 448px. Figure 3 illustrates that combining the CLIP-head and SAM-head can further enhance performance, although the results in Table 2 do not utilize the SAM-head. The performance of SAM-CLIP with SAM-head enhancement is showcased in Table 5. For example, on the Pascal-VOC dataset, SegCLIP achieves 52.6 mIoU, while the paper's SAM-CLIP, using only the CLIP-head without any refinement, attains 60.6 mIoU (as shown in Table 2). If both CLIP and SAM heads are combined, the paper's SAM-CLIP's performance can further improve to 66.0 mIoU (as shown in Table 5).", "category": "Experimental_Exposition"}, {"question": "How does SAM-CLIP's approach to zero-shot instance segmentation differ from HQ-SAM's approach, and why does SAM-CLIP experience a performance drop on LVIS compared to HQ-SAM?", "answer": "SAM-CLIP integrates CLIP's abilities into the SAM model to create a multi-task zero-shot model, whereas HQ-SAM enhances SAM's original instance segmentation ability through fine-tuning on high-quality datasets. The performance drop in SAM-CLIP on LVIS is due to its focus on multi-tasking rather than improving instance segmentation.", "category": "Method_Comparison"}, {"question": "What are the model size, speed, and memory consumption of the proposed SAM-CLIP compared to SAM/CLIP, particularly for edge device applications?", "answer": "The proposed SAM-CLIP uses the SAM ViT-B version as its base model architecture. After merging, a single ViT-B image encoder can perform SAM tasks (prompt-based segmentation), CLIP tasks (prompt-based classification and retrieval), and new text-to-segmentation tasks with a single inference. Note that all these tasks can be done with a single inference of ViT-B and require loading only one image encoder. For CLIP-related tasks, SAM-CLIP includes an additional light-head (3 transformer layers), which increases the memory footprint by 25% compared to a standalone CLIP with ViT-B (12 transformer layers).", "category": "Method_Comparison"}, {"question": "Can the proposed method perform instance segmentation using semantics from CLIP instead of bounding boxes as prompts?", "answer": "Yes, the method can handle zero-shot semantic segmentation, which involves providing an image and a set of candidate object class names (e.g., “dog”, “cat”, “human”) for semantic segmentation. Quantitative results for this task are provided in Table 2.", "category": "Method_Disambiguation"}, {"question": "What are the technical details, methodology, and design choices involved in the cross-VFM distillation framework, and how do they balance the trade-off between SAM’s original capabilities and learning CLIP’s capabilities?", "answer": "The cross-VFM distillation framework involves a two-stage training process. In the first stage, only the CLIP-head is trainable, using a single CLIP distillation loss (cosine distance between embeddings) with image batches from $D_{CLIP}$. This stage trains for 20 epochs without early stopping to provide a warm start for the CLIP-head. In the second stage, the SAM-head and ViT backbone become trainable, and a multi-task objective is used: CLIP Distillation (Eq. 1) and SAM self-distillation (Eq. 2). The balance between the losses is determined by the coefficient $\\lambda$, which is chosen to optimize the trade-off between forgetting SAM’s original capabilities (e.g., instance segmentation) and learning CLIP’s capabilities (e.g., zero-shot classification). Each training step involves: (1) computing gradients from the CLIP Distillation loss using a batch of 2048 images from $D_{CLIP}$, (2) computing gradients from the SAM self-distillation loss using a batch of 32 images from $D_{SAM}$, and (3) applying one optimization step. Early stopping occurs after 16 epochs to mitigate severe forgetting of SAM’s capabilities.", "category": "Method_Mechanics"}, {"question": "Why was SAM-CLIP not benchmarked against SAM-based models for zero-shot semantic segmentation, and how does SAM-CLIP's performance compare to CLIP-based models?", "answer": "SAM-CLIP was not benchmarked against SAM-based models for zero-shot semantic segmentation because, as of the submission date, the state-of-the-art zero-shot semantic segmentation models were all CLIP-based, and no SAM-based models were available for comparison in this task. SAM-CLIP demonstrates superior zero-shot semantic segmentation performance compared to previous CLIP-based methods. Additionally, SAM-based models like Semantic-SAM and SEEM are trained on datasets with semantic segmentation annotations, which disqualifies them from being considered zero-shot models in the literature.", "category": "Method_Comparison"}, {"question": "How does SAM-CLIP perform on out-of-distribution datasets such as ImageNet-v2 and Places-365, and how does its performance compare to baseline CLIP models?", "answer": "SAM-CLIP's performance on ImageNet-v2 is comparable to state-of-the-art CLIP models, as shown in Table 1. Additionally, on the Places-365 dataset, which has a significantly different data distribution from ImageNet, SAM-CLIP achieves slightly better performance than baseline CLIP models with a ViT-B backbone, indicating its robustness under distribution shifts.", "category": "Method_Comparison"}, {"question": "How is the joint training between head probing and multi-task distillation performed, and how is the balance between these tasks maintained during training? Additionally, what metrics or criteria are used to determine the stopping points for training?", "answer": "The training process consists of two stages: i) CLIP-head probing and ii) joint training of the SAM-head, CLIP-head, and the ViT backbone using a multi-task distillation loss. In the first stage, only the CLIP-head is trainable, and it is trained using a single CLIP distillation loss (cosine distance between embeddings). At this stage, all image batches come from only $D_{CLIP}$. This stage does not involve early stopping (authors train for 20 epochs). The motivation for this step is to have a warm start for the CLIP-head in the next stage where authors also allow modifying the backbone, similar to [1]. In the second stage, the SAM-head and the ViT backbone become trainable, and authors have a multi-task objective: CLIP Distillation (Eq. 1) and SAM self-distillation (Eq. 2). The balance between the losses is determined by the coefficient $\\lambda$, which authors picked to optimize the aforementioned trade-off. Each training step for this stage is performed as follows: - Pick a batch of 2048 images from $D_{CLIP}$, run the forward pass, and compute gradients backward from the CLIP Distillation loss (note that only parameters of the CLIP-head and ViT backbone will get gradients after this step). - Pick a batch of 32 images from $D_{SAM}$, run the forward pass, and compute gradients backward from the SAM self-distillation loss (note that only parameters of the SAM-head and ViT backbone will get gradients after this step). - Apply one optimization step (note that at this point, the parameters of the ViT backbone have accumulated gradients from the above two steps). Authors early-stop after 16 epochs (out of a full training length of 20 epochs) as authors observed more severe forgetting (as measured by instance segmentation performance on COCO) after the 16th epoch.\n[1] Kumar, Ananya, et al. \"Fine-tuning can distort pretrained features and underperform out-of-distribution.\" arXiv preprint arXiv:2202.10054 (2022).\n", "category": "Experimental_Setup"}, {"question": "Does the SAM-CLIP model require loading two separate models and multiple forward passes for each image during inference, similar to the model composition approach?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the training dataset for SAM-CLIP significantly larger than the datasets used for the original SAM and CLIP models?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the proposed merging technique limited to only CLIP and SAM, without potential application to other vision encoders?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the performance of SAM-CLIP in Table 2 obtained using both CLIP-head and SAM-head refinement?", "answer": "False", "category": "Claim_Verification"}, {"question": "Are the predicted masks by SegCLIP in Table 2 refined by SAM?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the training cost of SAM-CLIP exceed the combined pre-training costs of SAM and CLIP?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "jh7wH2AzKK", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Augmented Language Models: a Survey", "authors": ["Grégoire Mialon", "Roberto Dessi", "Maria Lomeli", "Christoforos Nalmpantis", "Ramakanth Pasunuru", "Roberta Raileanu", "Baptiste Roziere", "Timo Schick", "Jane Yu", "Asli Celikyilmaz", "Edouard Grave", "Yann LeCun", "Thomas Scialom"], "abstract": "This survey reviews works in which language models (LMs) are augmented with reasoning skills and the ability to use tools. The former is defined as decomposing a potentially complex task into simpler subtasks while the latter consists in calling external modules such as a\ncode interpreter. LMs can leverage these augmentations separately or in combination via heuristics, or learn to do so from demonstrations. While adhering to a standard missing tokens prediction objective, such augmented LMs can use various, possibly non-parametric external modules to expand their context processing ability, thus departing from the pure language modeling paradigm. We therefore refer to them as Augmented Language Models (ALMs). The missing token objective allows ALMs to learn to reason, use tools, and even act, while still performing standard natural language tasks and even outperforming most regular LMs on several benchmarks. In this work, after reviewing current advance in ALMs, we conclude that this new research direction has the potential to address common limitations of traditional LMs such as interpretability, consistency, and scalability issues.", "keywords": "", "qa_pairs": [{"question": "What additional details were included in the updated ethical concerns section regarding the impact of ALMs?", "answer": "ALMs raise new ethical concerns. LM predictions based on tools may look more trustworthy and authoritative at a first glance, when in fact many of them will still be incorrect. Moreover, authors can expect this phenomenon to be amplified as LMs reason in quite a similar manner to humans (Dasgupta et al., 2022), making it even harder to detect mistakes. Hence, ALMs may be leveraged to generate more convincing fake news and conspiracy theories. Conversely, conditioning ALMs on 'safe' and trusted sources should lead to more factuality and less toxicity in their answers. While such ethical concerns apply to most tools, it is important to distinguish between passive and active tools. The former only collects external information and passes it to the LM’s context, while the latter allows it to act on the virtual or physical world without human validation in the loop. An example of active tool is letting the LM control a search engine. Hence, there exists a broad spectrum of possible harmful consequences of LM usage. Authors are moving from passive LMs that generate text in isolation of the external environment, towards ALMs that act in the real world. In this context, the aforementioned ethical concerns may resonate even further, as ALM will be connected to more and more tools and environments. Currently, training LLM results in a large number of gas emissions, leveraging tools may result in smaller hence cheaper models that are more environmentally friendly both at training, inference, and update. ALMs are therefore a promising avenue to decrease the environmental impact of LLMs. ALMs also have the potential to transform the LM landscape. Nowadays, state of the art LLMs have been the product of large research organizations, fine-tuning these models towards more versatility (via reasoning and learning to use different tools) and agency (via actions) may be within range of smaller entities and perhaps individuals. Authors may therefore see the burgeoning of a market of open, possibly specialized and even personalized ALM agents, learning to use new tools, handling various data modalities and interacting with each other. Authors may also see the emergence of a new breed of app store, dedicated (via appropriate documentation or API constraints for example) to be used by ALMs. Overall, this new ecosystem may further accelerate the current pace of LLM research and application to the real world, while also making its control even more uncertain.", "category": "Method_Disambiguation"}, {"question": "What are the specific ethical concerns associated with Active Language Models (ALMs), and how do these concerns differ from those related to traditional Language Models (LMs)?", "answer": "ALMs raise new ethical concerns as their predictions based on tools may appear more trustworthy and authoritative, yet many remain incorrect. This is exacerbated by their human-like reasoning, making mistakes harder to detect. ALMs can generate more convincing fake news and conspiracy theories but can also be conditioned on 'safe' sources for more factual and less toxic outputs. Ethical concerns differ based on passive (collecting information) and active (acting on the virtual/physical world without human validation) tools. For instance, an active tool might let an LM control a search engine, leading to a broader spectrum of harmful consequences. The shift from passive LMs to ALMs that interact with the real world amplifies these ethical concerns as ALMs connect to more tools and environments.", "category": "Method_Comparison"}, {"question": "What role do ALMs play in reducing the environmental impact of LLMs?", "answer": "ALMs contribute to reducing the environmental impact of LLMs by enabling the development of smaller, cheaper models that are more environmentally friendly during training, inference, and updates, thereby decreasing gas emissions associated with LLMs.", "category": "Method_Mechanics"}, {"question": "How has the paper expanded its discussion on the limitations and open questions, particularly regarding retrieval efficiency, costs of ALMs, scalability, combining reasoning and tool use, and offline RL?", "answer": "The paper has expanded its discussion in the subsection titled 'Efficient large-scale retrieval' in Section 3.2.1, which describes optimizations such as indexing structures (e.g., FAISS), distributed frameworks, approximate nearest neighbor algorithms, and compression techniques like product quantization. For the costs of ALMs, the paper includes a discussion on how tool use can be more energy-efficient than retraining large models and proposes a way to estimate costs based on price per request and time required. Scalability is addressed through discussions on approaches like Toolformer for self-annotating tool use and embedding-based methods for selecting tools. Combining reasoning and tool use is highlighted with the need for benchmarks to evaluate complex tool combinations and the alignment of reasoning steps with tool use. Offline RL is discussed in terms of its promise and challenges, such as suboptimality of initial policies and distributional shifts.", "category": "Method_Disambiguation"}, {"question": "Why is the use of tools in the context of ALMs considered intractable, and how does the modified equation p(c | x_1, ..., x_{t-1}) address this issue?", "answer": "The use of tools in ALMs is considered intractable because (1) marginalizing over all possible tool uses is insufficient, as the probability p(c | x_1, ..., x_{t-1}) for some tools cannot be modeled due to unpredictable outputs (e.g., web browsing with evolving content), and (2) in cases where tools can be called recursively, the number of tool uses can grow exponentially with the number of tools, leading to intractable complexity. The equation was modified to p(c | x_1, ..., x_{t-1}) to better reflect the context-dependent choice of tools, but this does not resolve the intractability issue.", "category": "Motivation_Analysis"}, {"question": "How does the paper define 'taking actions' for ALMs, and how does this differ from the actions performed by standard language models?", "answer": "The paper defines 'taking actions' for ALMs as 'calling a tool that modifies a state in a virtual or physical object.' This differs from standard language models, which generate tokens but are not considered to be taking actions in this context.", "category": "Concept_Understanding"}, {"question": "What approaches are proposed for efficient large-scale retrieval in retrieval-augmented models, particularly when the retrieval set is extremely large, exact inference is intractable, or on-device retrieval is constrained by GPU memory?", "answer": "The paper addresses these concerns in Section 3.2 by proposing optimizations such as building index structures for embeddings (e.g., using FAISS), leveraging distributed frameworks for multi-node, multi-GPU setups, and using approximate nearest neighbor algorithms for faster retrieval when exact search is intractable. Compression techniques like product quantization are also discussed to reduce memory footprints. Additionally, vector database frameworks with built-in optimizations can be used to implement retrieval as a service without requiring custom solutions.", "category": "Method_Mechanics"}, {"question": "What are the primary costs associated with using ALMs, and how do these costs compare to those of integrating all knowledge into a single dense model? What dimensions of these costs should be considered, and has prior work thoroughly accounted for them? If not, what framework is proposed for measuring these costs?", "answer": "The primary costs of ALMs include computational costs such as API call expenses and time requirements. These costs have not been comprehensively studied in prior work. A proposed framework for estimating these costs involves considering the price per request and the time needed to fulfill the request. Retrieval-augmented models can be more energy-efficient than retraining large models, as updating an external data store is less costly. Including cost as a loss term during training and inference can help ALMs optimize tool use, balancing internal reasoning and external tool calls.", "category": "Method_Disambiguation"}, {"question": "What are some potential avenues for the community to explore regarding the decision of when and which tools to use in the context of ALMs?", "answer": "Potential avenues include approaches like Toolformer, which uses self-annotation to determine useful tool uses, embedding-based methods for selecting tools, incorporating tool use costs into training to develop optimal strategies, and creating benchmarks to evaluate tool use capabilities for complex tasks requiring sequential or combined tool use.", "category": "Method_Mechanics"}]}
{"id": "Ee277P3AYC", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "CoCa: Contrastive Captioners are Image-Text Foundation Models", "authors": ["Jiahui Yu", "Zirui Wang", "Vijay Vasudevan", "Legg Yeung", "Mojtaba Seyedhosseini", "Yonghui Wu"], "abstract": "Exploring large-scale pretrained foundation models is of significant interest in computer vision because these models can be quickly transferred to many downstream tasks. This paper presents Contrastive Captioner (CoCa), a minimalist design to pretrain an image-text encoder-decoder foundation model jointly with contrastive loss and captioning loss, thereby subsuming model capabilities from contrastive approaches like CLIP and generative methods like SimVLM. In contrast to standard encoder-decoder transformers where all decoder layers attend to encoder outputs, CoCa omits cross-attention in the first half of decoder layers to encode unimodal text representations, and cascades the remaining decoder layers which cross-attend to the image encoder for multimodal image-text representations. We apply a contrastive loss between unimodal image and text embeddings, in addition to a captioning loss on the multimodal decoder outputs which predicts text tokens autoregressively. By sharing the same computational graph, the two training objectives are computed efficiently with minimal overhead. CoCa is pretrained end-to-end and from scratch on both web-scale alt-text data and annotated images by treating all labels simply as text, seamlessly unifying natural language supervision for representation learning. Empirically, CoCa achieves state-of-the-art performance with zero-shot transfer or minimal task-specific adaptation on a broad range of downstream tasks, spanning visual recognition (ImageNet, Kinetics-400/600/700, Moments-in-Time), crossmodal retrieval (MSCOCO, Flickr30K, MSR-VTT), multimodal understanding (VQA, SNLI-VE, NLVR2), and image captioning (MSCOCO, NoCaps). Notably on ImageNet classification, CoCa obtains 86.3% zero-shot top-1 accuracy, 90.6% with a frozen encoder and learned classification head, and 91.0% with a finetuned encoder.", "keywords": "", "qa_pairs": [{"question": "What is the architecture of the captioning-only model in terms of the number of unimodal and multimodal layers in the decoder, as referenced in Table 7b and the discussion paragraph?", "answer": "The captioning-only model, which uses the same architecture as CoCa, contains a total of 12 layers in the decoder, consisting of 6 unimodal layers and 6 multimodal layers. This configuration was chosen to ensure a fair comparison for computational efficiency.", "category": "Experimental_Setup"}, {"question": "What does 'encoding the captions of each video as target embeddings' in Section 3.3 refer to in terms of the retrieval pipeline?", "answer": "It refers to encoding each caption into a unimodal text embedding independently.", "category": "Method_Disambiguation"}, {"question": "What is the design and number of queries used in the attention pooler for the image classification task?", "answer": "The design used is 'feature_map --> attention pooler with n_cls queries --> softmax NLL loss', and it employs 256 queries, which is the same as during pretraining.", "category": "Experimental_Setup"}, {"question": "What type of classification head is used in other works such as SwinV2, ViT, and CoAttNet, as compared to CoCa in Figure 4?", "answer": "For other networks like SwinV2, ViT, and CoAttNet, standard linear classification heads are used, as obtained from their corresponding publications.", "category": "Method_Mechanics"}, {"question": "Why is an additional attentional pooler with 256 queries needed despite similar results being observed with 0 queries, as shown in Sec 4.3 and Table 7.f?", "answer": "An additional attentional pooler with 256 queries is needed because it restricts the number of tokens during finetuning on larger image resolutions, making the process computationally more efficient.", "category": "Experimental_Analysis"}, {"question": "What decoding method is used in the image captioning evaluation, and if beam search is used, what is the beam size?", "answer": "The image captioning evaluation uses beam search with a beam size of 4.", "category": "Method_Disambiguation"}]}
{"id": "b4tMhpN0JC", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "GIT: A Generative Image-to-text Transformer for Vision and Language", "authors": ["Jianfeng Wang", "Zhengyuan Yang", "Xiaowei Hu", "Linjie Li", "Kevin Lin", "Zhe Gan", "Zicheng Liu", "Ce Liu", "Lijuan Wang"], "abstract": "In this paper, we design and train a Generative Image-to-text Transformer, GIT, to unify vision-language tasks such as image/video captioning and question answering. While generative models provide a consistent network architecture between pre-training and fine-tuning, existing work typically contains complex structures (uni/multi-modal encoder/decoder) and depends on external modules such as object detectors/taggers and optical character recognition (OCR). In GIT, we simplify the architecture as one image encoder and one text decoder\nunder a single language modeling task. We also scale up the pre-training data and the model size to boost the model performance. Without bells and whistles, our GIT establishes new state of the arts on numerous challenging benchmarks with a large margin. For instance, our model surpasses the human performance for the first time on TextCaps (138.2 vs. 125.5 in CIDEr). Furthermore, we present a new scheme of generation-based image classification and scene text recognition, achieving decent performance on standard benchmarks.", "keywords": "", "qa_pairs": [{"question": "What are the performance differences when the visual encoder is initialized using different methods, including random initialization, supervised pretraining, self-supervised pretraining, and CLIP pretraining?", "answer": "The setting follows GIT-B. The image encoder is the base-sized version of ViT, which is initialized 1) from the CLIP model, 2) from the supervised pretraining (classification task on ImageNet), 3) from the self-supervised pretraining (MAE on ImageNet), or 4) randomly initialized. From the results, authors can clearly observe the higher performance with the CLIP pretrained weights. Compared with the supervised/self-supervised, authors note that the pretraining datasets for the image encoder are different here due to the availability of these weights. Although it is unclear whether the pretraining dataset is more important than the task or vice versa, authors choose the contrastive pretraining as the pretraining dataset is also easy to scale up. For randomly initialization, authors observe significant lower performance. The reason could be the small scale of the pretraining set (10M image-text pairs in the set-up of GIT-B ). A larger dataset may reduce the gap, but it may require longer training iterations. \n\n|                 | COCO  | TextCaps | VizWiz-Captions |\n|-----------------|-------|----------|-----------------|\n| CLIP            | 131.4 | 64.9     | 61.2            |\n| Supervised      | 122.1 | 47.0     | 58.3            |\n| self-supervised | 123.4 | 44.9     | 51.6            |\n| Random          | 89.0  | 36.5     | 38.1            |", "category": "Method_Comparison"}, {"question": "What are the main factors contributing to the performance improvement observed in the method compared to existing methods, and how have these factors been demonstrated in the experiments?", "answer": "The primary factors contributing to the performance improvement are large data scales and large model size. This is demonstrated in Section 4.6, where the paper shows significant performance gains from scaling both data and model on 9 benchmarks. Additionally, Table 2 provides a comparison of pretraining dataset scales and model parameters, illustrating that the base-sized model with 4M pretraining images achieves a competitive CIDEr score of 131.4, compared to VinVL's 129.3 with 6M pretraining images.", "category": "Method_Comparison"}, {"question": "Why does the proposed model use a randomly initialized text decoder instead of leveraging pretrained text decoders, and what insights can be provided on the observed behavior compared to existing work?", "answer": "The text decoder is randomly initialized based on ablation observations from MiniVLM. An ablation study comparing random and initialized transformers showed no improvement with initialization. This suggests that initialized weights cannot effectively understand image inputs, crucial for vision-language (VL) tasks. In contrast, models like Flamingo use fixed pretrained transformers where initialization is important, and VX2TEXT converts all modalities to textual names, making the input meaningful to the pretrained transformer. In the model, proper initialization is essential for equipping the model with respective capabilities, especially when models are not trained.\n| Initialization  | COCO  | TextCaps  | VizWiz-Captions |\n|-----------------|-------|-----------|-----------------|\n| Random          | 127.5 | 61.7      | 68.0            |\n| Initialized     | 126.9 | 59.7      | 67.0            |", "category": "Experimental_Analysis"}, {"question": "Is the difference in size between the image encoder and text decoder in this work compared to Flamingo correlated to the random initialization of the text decoder? What is the principle for deciding the size of each component, and what is the performance impact of using a relatively larger text decoder?", "answer": "The random initialization of the text decoder makes it more difficult to train effectively, whereas Flamingo freezes the decoder to avoid this issue. Experiment results in Table 10 show no improvement with larger decoders. The principle for deciding the size of each component is based on empirical results. If the decoder is trained with the whole network, it should be small; if it is fixed, it should be well initialized and as large as possible.", "category": "Experimental_Analysis"}, {"question": "What are the performance comparisons between the proposed approach and Flamingo in zero-shot, few-shot, and full fine-tuning scenarios for image captioning and ImageNet classification?", "answer": "Specifically, on the image captioning, authors evaluate the performance on Flickr with the Karpathy split. Authors do not evaluate it on COCO or nocaps because the COCO dataset are already in the pretraining dataset. The result is shown as follows, in which authors can observe 1) with few-shots of 16 or 32 shots, the approach is better than Flamingo, 2) the approach benefits from more shots from 16 to 32, while it is not necessary to be the case for Flamingo, which is the in-context learning without parameter updates. Last but not the least, with full fine-tuning, authors can also achieve new SoTA performance on this dataset.\n\n|                   | 0    | 16   |   32 | 290 (1%) | full |\n|-------------------|------|------|-------|----------|------|\n| Zhou et al (2020) | -    | -    | -     | -        | 68.5 |\n| Flamingo          | 67.2 | 78.9 | 75.4  |          |      |\n| GIT               | 49.6 | 78.0 | 80.5  | 86.6     | 98.5 |\n\nThe following shows the results of the zero/few-shot study on ImageNet. As the model is generative, which can predict any output, authors evaluate the performance in three metrics: 1) the predicted caption should be exactly matched to the ground-truth label name, named as equal; 2) the predicted caption should contain the ground-truth label name, named as in; 3) with the vocabulary as a prior, authors apply a trie structure to constrain the candidate tokens in each prediction step, such that the predicted caption is always within the vocabulary. As only the output space is limited from the whole token vocabulary to a subset, the inference cost is almost not changed, and is also constant to the category size. For Flamingo, each category label is fed into the network to compute the score and is thus linear to the vocabulary size. From the results, authors can see with the vocabulary as a prior, the generative model can be better than Flamingo with 1 shot or 5 shot per class. More discussions are added in the paper (Sec. 4.5, table 9).\n\n|               |       | zero-shot |              |       | 1 shot per class |              |       | 5 shot |           |\n|---------------|-------|-----------|--------------|-------|------------------|--------------|-------|--------|-----------|\n| Accuracy type | equal | in        | voc-prior    | equal | in               | voc-prior    | equal | in     | voc-prior |\n| Flamingo      |       |           |              |       |                  | 71.7         |       |        | 77.3      |\n| GIT           | 1.93  | 40.88     | 33.48        | 64.54 | 66.76            | 72.45        | 79.79 | 80.15  | 80.93     |", "category": "Method_Comparison"}, {"question": "What are the specific details of how videos are sampled and preprocessed during training and inference?", "answer": "During training, 6 frames are randomly sampled with equal intervals, and the same random crop is applied to these frames. During inference, 6 frames are uniformly sampled with a center crop. As with the image domain, only random crop is applied without any other augmentation, such as random flip or color distortion.", "category": "Experimental_Setup"}, {"question": "How does the model acquire OCR capability without specific OCR modules?", "answer": "The model acquires OCR capability primarily through the data, as the model structure consists of one image encoder and one text decoder without OCR-specific modules. The data contains 15% to 31% of associated texts with scene text, and the model learns the mapping from the image to the associated text, thereby acquiring OCR capability.", "category": "Method_Mechanics"}, {"question": "Why does the paper not include experiments on the TextOCR dataset, and what are the results of the fine-tuned GIT model on the TextOCR validation set?", "answer": "The paper does not include experiments on the TextOCR dataset because the test set is not publicly available, and the reference paper of TextOCR only reports results on the test set. However, the authors fine-tuned their pretrained GIT model on the TextOCR validation set with 10 epochs and a learning rate of 2e-5, achieving an accuracy of 81.27. In the paper of TextOCR, the reported best accuracy on the test set is 69.49. Assuming the validation and test sets follow a similar distribution, the authors believe their approach can achieve competitive or better performance on the TextOCR dataset.", "category": "Experimental_Exposition"}, {"question": "How does the amount of training data (pre-training and fine-tuning) and the number of parameters used by GIT compare to recent large-scale vision and language models?", "answer": "All methods share the same amount of fine-tuning data. Performance generally improves with more training data and more model parameters. For the base-sized model, GIT achieves a 131.4 CIDer score with 4M pretraining images, while VinVL achieves a 129.3 CIDer score with 6M pretraining images.", "category": "Method_Comparison"}, {"question": "What are the zero-shot and few-shot performance results of the proposed approach on image captioning and ImageNet classification tasks, and how do they compare to Flamingo?", "answer": "The zero-shot and few-shot performance results for image captioning were evaluated on the Flickr dataset with the Karpathy split, showing that the proposed approach outperforms Flamingo with 16 or 32 shots and benefits from more shots. On ImageNet classification, the generative model was evaluated using three metrics: exact match (equal), containing the ground-truth label (in), and using a vocabulary prior with a trie structure. The results indicate that the proposed model performs better than Flamingo with 1 or 5 shots per class when using the vocabulary prior. Detailed results and discussions are provided in the manuscript (Sec. 4.5, table 9).\n|                   | 0    | 16   |  32   | 290 (1%) | full |\n|-------------------|------|------|-------|----------|------|\n| Zhou et al (2020) | -    | -    | -     | -        | 68.5 |\n| Flamingo          | 67.2 | 78.9 | 75.4  |          |      |\n| GIT               | 49.6 | 78.0 | 80.5  | 86.6     | 98.5 |\n\n|               |       | zero-shot |              |       | 1 shot per class |              |       | 5 shot |           |\n|---------------|-------|-----------|--------------|-------|------------------|--------------|-------|--------|-----------|\n| Accuracy type | equal | in        | voc-prior    | equal | in               | voc-prior    | equal | in     | voc-prior |\n| Flamingo      |       |           |              |       |                  | 71.7         |       |        | 77.3      |\n| GIT           | 1.93  | 40.88     | 33.48        | 64.54 | 66.76            | 72.45        | 79.79 | 80.15  | 80.93     |", "category": "Method_Comparison"}, {"question": "What are the fundamental limitations of the model, particularly regarding its capability for few-shot learning?", "answer": "The model is designed for pretraining-and-finetuning scenarios, and it is unclear how to apply in-context learning to adapt to new tasks without parameter updates. Additionally, controlling the output style, such as consistently outputting or not outputting scene text, is challenging. However, recent findings indicate that the model's few-shot learning capability is strong, as discussed in A1 for the first question.", "category": "Experimental_Analysis"}, {"question": "Does the paper include experimental results on the recent TextOCR dataset's test set?", "answer": "False", "category": "Claim_Verification"}, {"question": "Did the authors fine-tune their pretrained GIT model on the TextOCR validation set and achieve an accuracy of 81.27?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the test set of TextOCR publicly available according to the official website?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "y1pPWFVfvR", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Multimodal Chain-of-Thought Reasoning in Language Models", "authors": ["Zhuosheng Zhang", "Aston Zhang", "Mu Li", "hai zhao", "George Karypis", "Alex Smola"], "abstract": "Large language models (LLMs) have shown impressive performance on complex reasoning by leveraging chain-of-thought (CoT) prompting to generate intermediate reasoning chains as the rationale to infer the answer. However, existing CoT studies have primarily focused on the language modality. We propose Multimodal-CoT that incorporates language (text) and vision (images) modalities into a two-stage framework that separates rationale generation and answer inference. In this way, answer inference can leverage better generated rationales that are based on multimodal information. Experimental results on ScienceQA and A-OKVQA benchmark datasets show the effectiveness of our proposed approach. With Multimodal-CoT, our model under 1 billion parameters achieves state-of-the-art performance on the ScienceQA benchmark. Our analysis indicates that Multimodal-CoT offers the advantages of mitigating hallucination and enhancing convergence speed. Code is publicly available at https://github.com/amazon-science/mm-cot.", "keywords": "", "qa_pairs": [{"question": "What are the novel and significant aspects of the two-stage prompting approach in this study?", "answer": "The novelty and significance of the two-stage prompting approach are highlighted in three aspects: (1) It is the first to study Chain-of-Thought (CoT) reasoning across different modalities. (2) It provides a unique and effective solution for multimodal CoT problems, distinct from text-only reasoning or Visual Question Answering (VQA) approaches, by enabling explicit intermediate reasoning with multimodal input through a frozen image encoder and two-stage framework. This framework mitigates hallucination and enhances convergence speed. (3) The two-stage approach analyzes why naive CoT fails in this context and demonstrates how incorporating vision features addresses the problem. The approach is effective across tasks and backbone models, and in-depth studies explore its benefits and limitations to support future research.", "category": "Method_Disambiguation"}, {"question": "What evaluation results have been added for the multimodal reasoning benchmark MMMU, and how does Multimodal-CoT perform on it?", "answer": "The authors have added experiment results on the MMMU benchmark, which evaluates multimodal models on college-level tasks across six core disciplines. Multimodal-CoT was tested on MMMU without additional training and achieved an accuracy of 28.7, outperforming larger models such as Kosmos-2 (1.6B), Fuyu (8B), OpenFlamingo-2 (9B), and MiniGPT4-Vicuna (13B). The results and discussions are included in Section 6.6.\n| Method | Size | Accuracy |\n| ---------- | ---------- | ---------------- |\n| Kosmos-2 | 1.6B | 24.4  |\n| Fuyu | 8B |  27.9 |\n| OpenFlamingo-2 | 9B | 28.7  |\n|  MiniGPT4-Vicuna | 13B | 26.8 |\n| Multimodal-CoT | 738M | 28.7 |\n| GPT-4V(ision) | - | 56.8 |\n| Gemini Ultra | - | 59.4 |\n[1] Yue, Xiang, et al. \"Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi.\" arXiv preprint arXiv:2311.16502 (2023).", "category": "Experimental_Exposition"}, {"question": "How are the results in Table 7 computed for ChatGPT when the input includes images?", "answer": "For computing the results with image input for ChatGPT in Table 7, only the question is passed to ChatGPT, without the image input. InstructBLIP is used to generate rationales for questions with paired images, while ChatGPT generates rationales for questions without paired images. These generated pseudo-rationales are combined as the target rationales for training (Multimodal-CoT w/ Generation), replacing the need for human-annotated reasoning chains (Multimodal-CoT w/ Annotation). This approach allows Multimodal-CoT to achieve better performance than using InstructBLIP or ChatGPT alone.", "category": "Method_Mechanics"}, {"question": "What are the novel aspects of the proposed method that distinguish it from previous studies?", "answer": "The proposed method is novel in three key aspects: (1) it is the first to study Chain-of-Thought (CoT) reasoning in different modalities, (2) it introduces a pioneering approach to the multimodal CoT problem by providing explicit intermediate reasoning with multimodal input, powered by a frozen image encoder and a two-stage framework, and (3) it analyzes why the naive application of CoT fails in this context and how incorporating vision features alleviates the problem. Additionally, the method has been shown to be effective across tasks and backbone models, and in-depth studies have been conducted to explore its benefits and limitations.", "category": "Method_Comparison"}, {"question": "How does Multimodal-CoT perform on the MMMU benchmark, which it was not trained on, compared to other larger models?", "answer": "Multimodal-CoT demonstrates effective generalization to the MMMU benchmark without additional training, achieving an accuracy of 28.7% which is comparable to or better than larger models such as Fuyu (8B, 27.9%) and OpenFlamingo-2 (9B, 28.7%), as detailed in Section 6.6. The performance of GPT-4V and Gemini Ultra is also reported for reference but is not considered a fair comparison due to their larger model scales and lack of training details.\n| Method | Size | Accuracy |\n| ---------- | ---------- | ---------------- |\n| Kosmos-2 | 1.6B | 24.4  | \n| Fuyu | 8B |  27.9 |\n| OpenFlamingo-2 | 9B | 28.7  | \n|  MiniGPT4-Vicuna | 13B | 26.8 |\n| Multimodal-CoT | 738M | 28.7 |\n| GPT-4V(ision) | - | 56.8 |\n| Gemini Ultra | - | 59.4 |", "category": "Method_Comparison"}, {"question": "How do different language models and visual backbones, including larger variants, affect the performance of the proposed approach?", "answer": "The analysis shows that the proposed approach is generally effective across different language models and visual backbones. For language models, FLAN-Alpaca achieves the highest accuracy (85.31), followed by FLAN-T5 (83.19) and UnifiedQA (82.55). For visual backbones, ViT performs best (85.31), followed by CLIP (84.27) and DETR (83.16). Additionally, intricate alignment strategies, such as the Image-grounded text encoder inspired by BLIP, contribute to better performance than direct answering, with the Unimodal encoder achieving the highest accuracy (85.31).\n| Language Backbone | Accuracy |\n| ---------- | ---------------- |\n| UnifiedQA |  82.55  | \n| FLAN-T5 | 83.19 |\n| FLAN-Alpaca |  85.31 |\n\n| Vision Backbone | Accuracy |\n| ---------- | ---------------- |\n| ViT |  85.31  | \n| CLIP | 84.27 |\n| DETR |  83.16 |\n\n| Method | Accuracy |\n| ---------- | ---------------- |\n| Direct Answering | 82.62 |\n| Unimodal encoder | 85.31 | \n| Image-grounded text encoder | 84.60  |", "category": "Experimental_Exposition"}]}
{"id": "hFALpTb4fR", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models", "authors": ["Long Lian", "Boyi Li", "Adam Yala", "Trevor Darrell"], "abstract": "Recent advancements in text-to-image diffusion models have yielded impressive results in generating realistic and diverse images. However, these models still struggle with complex prompts, such as those that involve numeracy and spatial reasoning. This work proposes to enhance prompt understanding capabilities in diffusion models. Our method leverages a pretrained large language model (LLM) for grounded generation in a novel two-stage process. In the first stage, the LLM generates a scene layout that comprises captioned bounding boxes from a given prompt describing the desired image. In the second stage, a novel controller guides an off-the-shelf diffusion model for layout-grounded image generation. Both stages utilize existing pretrained models without additional model parameter optimization. Our method significantly outperforms the base diffusion model and several strong baselines in accurately generating images according to prompts that require various capabilities, doubling the generation accuracy across four tasks on average. Furthermore, our method enables instruction-based multi-round scene specification and can handle prompts in languages not supported by the underlying diffusion model. We anticipate that our method will unleash users' creativity by accurately following more complex prompts. Our code, demo, and benchmark are available at: https://llm-grounded-diffusion.github.io", "keywords": "", "qa_pairs": [{"question": "What challenges does the paper's approach face in generalizing to diverse and complex scenarios with ambiguous or highly creative text prompts?", "answer": "The method may face challenges with extremely complicated prompts, where the bottleneck is often the diffusion models rather than the LLM, leading to misalignment between the generated layout and the final image. Improvements in generalization and robustness are expected with better LLMs and diffusion models. ", "category": "Experimental_Analysis"}, {"question": "How does the computational resource requirement of the proposed method impact its accessibility for users with less powerful hardware, and what measures can mitigate this issue?", "answer": "The proposed method is more computationally demanding due to its two-stage pipeline involving an LLM, which increases resource usage compared to single-stage text-to-image pipelines like Stable Diffusion. The main cost is in the LLM query part, requiring local model serving or API queries, which incur additional costs and latency. However, existing production pipelines like DALL-E 3 already use LLMs for prompt processing, suggesting potential reuse for layout generation and scalable deployment. Future releases of more efficient LLMs and serving methods are expected to enhance efficiency. Additionally, the method's applicability to various LLMs and diffusion models off-the-shelf expands its reach to users without the budget or hardware for annotated image collection and model fine-tuning.", "category": "Method_Mechanics"}, {"question": "What is the computational trade-off of the proposed method compared to existing methods, and is it more demanding due to its per-instance approach?", "answer": "The proposed method is more computationally demanding because it involves a two-stage pipeline that uses a large language model (LLM), making it more complex than single-stage text-to-image pipelines like Stable Diffusion. The primary computational cost comes from the LLM query part, which requires either serving a model locally or querying an API, both of which incur costs and latency. However, production pipelines such as DALL-E 3 already use an LLM for prompt processing, suggesting that the existing LLM could be reused for layout generation, enabling scalable deployment. Additionally, future advancements in computationally efficient LLMs and serving methods are expected to significantly improve the efficiency of the proposed method. Compared to methods requiring training or external human annotation, the proposed method can evolve with LLMs at minimal adaptation costs.", "category": "Method_Comparison"}, {"question": "What is the rationale for comparing the proposed method LMD with MultiDiffusion, Backward Guidance, and BoxDiff on the novel benchmark rather than on T2I-CompBench?", "answer": "The proposed method LMD follows a text-to-image pipeline, whereas MultiDiffusion, Backward Guidance, and BoxDiff are layout-to-image methods. These layout-to-image methods do not perform text-to-image generation and are not directly comparable with LMD. Without layouts as inputs, these methods fall back to vanilla Stable Diffusion. Therefore, comparisons were made against the base text-to-image diffusion model Stable Diffusion in Table 1. Ablations on various layout-to-image methods were performed in Table 2, comparing LMD (as both stage 1 and 2) with variants of LMD that use other methods as stage 2. The ablations were conducted on the novel benchmark due to this rationale. Additionally, ablations were also performed on T2I-CompBench, showing that LMD outperforms baseline methods and other ablations on all four tasks.\n\n| # | T2I-CompBench                                       | Color      | Shape      | Texture    | Spatial    |\n| - | --------------------------------------------------- | ---------- | ---------- | ---------- | ---------- |\n| 1 | SDv1                                                | 0.3765     | 0.3576     | 0.4156     | 0.1246     |\n| 2 | LMD (stage 1) + MultiDiffusion (**new results**)    | 0.4631     | 0.4497     | 0.4007     | 0.1604     |\n| 3 | LMD (stage 1) + Backward Guidance (**new results**) | 0.4877     | 0.5069     | 0.4643     | 0.2361     |\n| 4 | LMD (stage 1) + Boxdiff (**new results**)           | 0.4579     | 0.4967     | 0.4720     | 0.1965     |\n| 5 | LMD (stage 1 and 2)                                 | **0.5495** | **0.5462** | **0.5241** | **0.2570** |\n", "category": "Experimental_Analysis"}, {"question": "Does the prompt used in the text-to-layout stage introduce a bias towards generating elements with spatial distributions similar to the in-context examples?", "answer": "The authors tested this by presenting the LLM with only one in-context example and querying it with a prompt similar to the example. For instance, with the in-context example 'A panda in a forest without flowers', the LLM was queried to generate a layout for 'An apple'. The results, detailed in Appendix E and Fig. E1, show that the LLM generates layouts tailored to the objects in the query prompt rather than mimicking the layout boxes from the in-context examples. This qualitative analysis indicates that the LLM often generates natural layouts according to the prompts, reducing the need for heavy prompt engineering to avoid overly similar layouts.", "category": "Experimental_Analysis"}, {"question": "How does the proposed method address challenges in generating images from prompts involving hard-to-segment scenes, such as complex or abstract elements?", "answer": "The proposed method uses an intermediate representation for improved text-to-image generation. While 2D layout boxes are the primary intermediate representation, the authors suggest that generating spatially fine-grained scenes can be further enhanced by using other forms of intermediate representation, such as 2D masks or 3D bounding boxes for occlusion awareness. The layout-to-image stage incorporates both the intermediate representation layout and the user prompt text, ensuring that the method performs at least on par with baseline text-to-image diffusion models in challenging cases. ", "category": "Method_Mechanics"}, {"question": "How were the 11 responses utilized for 10 prompts in the human feedback evaluation in Section 4.2.3, and is this number sufficient to establish meaningful trends?", "answer": "Each of the 11 evaluators provided 10 responses, resulting in 110 responses for each of the two questions. The responses show small variations, particularly for the primary question on text-image alignment, revealing significant performance gaps between the method and the baseline SD (88% vs 10% favoring the method). Despite the relatively small number of responses, the observed trends are unlikely due to randomness. ", "category": "Method_Mechanics"}, {"question": "How does the proposed method address the potential bias introduced by using a few-shot prompt of examples, particularly in relation to generating layouts for objects not included in the in-context examples?", "answer": "The proposed method is not limited to objects in the in-context examples, as demonstrated by its ability to generate layouts for objects like a deer, bear, horse, and astronaut, none of which appear in the in-context examples. This indicates that the LLM can provide object-specific layouts even without examples as references. While in-context examples can improve accuracy for the objects mentioned, the method still achieves significant improvements for objects not in the in-context examples, demonstrating its applicability to open-set object classes. Clarifications have been added in Section 5.", "category": "Method_Mechanics"}, {"question": "What is the novelty of the layout-grounded diffusion model in the second stage of the proposed two-stage generation pipeline, as illustrated in Figure 5, and how does it differ from previous layout-to-image methods?", "answer": "The method is a two-stage generation pipeline with an LLM for text-to-layout generation (stage 1) and a layout-grounded diffusion model for layout-to-image generation (stage 2). Authors focus on addressing the concern about the novelty of the second stage illustrated in Fig. 5 of the paper. Indeed, there are previous layout-to-image methods that manipulate the attention maps to guide the diffusion models to generate images according to box layouts, which authors cited as related works. However, since the cross-attention map represents text-to-image similarities, such manipulation could only control 'which part of the image corresponds to which concept', but it could not control the number of object instances in the specified region. One prominent consequence is that these methods could not distinguish between two boxes of the same concept side-by-side touching each other and one large box that is the union of the smaller boxes. In contrast, the layout-grounded Stable Diffusion first generates a masked latent for each object instance, with the prompt specifically asking for one object (Sec 3.1). These masked latents are then composed together to serve as hints to guide the overall image generation (Sec 3.2), with the strength of hints adjustable to prevent artifacts in the overall generation. Since each masked latent only represents one object instance instead of a semantic concept, the simple yet effective method ensures that only one object will be generated in each box, allowing precise control at an instance level without additional training for the first time. To verify that it is advantageous to use layout-grounded Stable Diffusion in the layout-to-image stage of the pipeline, authors perform ablations on the pipeline in Tab. 2, swapping the stage 2 with other layout-to-image methods (e.g., Bar-Tal et al., Chen et al., and Xie et al.) while keeping stage 1 the same. The method surpasses these previous methods that only perform coarse-grained control on a semantic level. Therefore, the stage 2 method is simple and easy to understand, yet it enables instance-level control that previous training-free methods could not achieve, which indicates non-trivial novelty compared to previous works.", "category": "Method_Comparison"}, {"question": "What is the technical novelty of the proposed method, particularly in terms of the difference between semantics and instances, and how does the per-instance masked latent generation in Algorithm 1 differentiate it from existing methods?", "answer": "The method focuses on improving the prompt understanding ability of a text-to-image diffusion model, as explained in the beginning of Sec. 3. The text-to-image generation setting that the method follows is different from the layout-to-image setting that previous works such as MultiDiffusion (Bar-Tal et al.), Layout Guidance (Chen et al.), and BoxDiff (Xie et al.) focus on. These works require layouts as input and fall back to their baseline (Stable Diffusion) when used to perform text-to-image generation (i.e., with only a global caption but no boxes in the provided layout). This is also the reason that authors only compare with  true baseline Stable Diffusion in Tab. 1. However, these layout-to-image methods can be used as subroutines to create ablated variants of the methods, which authors discuss in the contribution 2 section below. Here authors summarize contributions. Contribution 1: overall text-to-image pipeline. The paper's first contribution is a two-stage generation pipeline that improves the prompt understanding ability of diffusion models in a text-to-image setting. In this proposed pipeline, authors use an LLM for text-to-layout generation (stage 1) and a layout-grounded diffusion model for layout-to-image generation (stage 2). While layout-to-image methods inherently require layouts as inputs, the idea of involving an LLM in the generation pipeline allows the improvement of prompt understanding ability in a text-to-image setting with no layouts provided, which authors believe should not be overlooked. The proposed pipeline is flexible in terms of the selection of the LLM and the layout-grounded diffusion method, which has been extensively validated by the ablation studies in the experiments section. Specifically, methods such as Bar-Tal et al., Chen et al., and Xie et al. can be used as the stage 2 of the pipeline, which authors ablated in Tab. 2, but these methods on their own do not perform text-to-image generation and thus are not directly comparable with the overall method. The work is among the first works that involve LLMs in a text-to-image generation pipeline. Recent concurrent works VisualChatGPT and GILL also leverage an LLM and a text-to-image generation model (Stable Diffusion). However, in comparison with the method, as shown in Fig. 8, these methods still suffer from problems that can be successfully solved by the proposed pipeline in Fig. 1. Contribution 2: the training-free layout-conditioned generation (stage 2). In addition, authors introduce a novel layout-to-image method called layout-grounded Stable Diffusion. This is used as the stage 2 of the pipeline to generate images grounded on bounding box layouts from the LLM. Compared to previous training-free layout-grounded generation methods, the method applies control on the instanceness rather than semantics. Specifically, previous methods perform layout control by only manipulating the cross-attention maps. Since the cross-attention map represents text-to-image similarities, such manipulation could only control 'which part of the image corresponds to which concept', but it could not control the number of object instances in the specified region or the arrangement of object instances in regions with the same semantics. One prominent consequence is that these methods could not distinguish between two boxes of the same concept side-by-side touching each other and one large box that is the union of the smaller boxes. In contrast, the layout-grounded Stable Diffusion first generates a masked latent for each object instance, with the prompt specifically asking for one object (Sec 3.1). These masked latents are then composed together to serve as hints to guide the overall image generation (Sec 3.2), with the strength of hints adjustable to prevent artifacts in the overall generation. Since each masked latent only represents one object instance instead of a semantic concept, the method ensures that only one object will be generated in each box, allowing precise control at an instance level without additional training for the first time. To verify that it is advantageous to use the paper's layout-grounded Stable Diffusion in the layout-to-image stage of the paper's pipeline, authors perform ablations on the paper's pipeline in Tab. 2, swapping the paper's stage 2 with other layout-to-image methods (e.g., Bar-Tal et al., Chen et al., and Xie et al.) while keeping stage 1 the same. The method surpasses these previous methods that only perform coarse-grained control on a semantic level. Contribution 3: instruction-based scene specification, broader language support, and a prompt understanding benchmark for text-to-image generative models. Authors discovered that the overall proposed method LMD enables instruction-based scene specification (Fig. 6) and allows broader language support for the prompt input (Sec 3.3) without any additional training, which are useful for downstream applications. Authors also propose a prompt understanding benchmark to quantitatively evaluate the ability of a text-to-image generation pipeline to understand prompts that involve certain concepts (e.g., generative numeracy and negation in the prompt). ", "category": "Method_Disambiguation"}, {"question": "Were 11 responses used to evaluate 10 prompts, resulting in some prompts being rated twice?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper suggest that the observed gaps in text-image alignment between the proposed method and baseline SD are likely due to randomness?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "yzkSU5zdwD", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Emergent Abilities of Large Language Models", "authors": ["Jason Wei", "Yi Tay", "Rishi Bommasani", "Colin Raffel", "Barret Zoph", "Sebastian Borgeaud", "Dani Yogatama", "Maarten Bosma", "Denny Zhou", "Donald Metzler", "Ed H. Chi", "Tatsunori Hashimoto", "Oriol Vinyals", "Percy Liang", "Jeff Dean", "William Fedus"], "abstract": "Scaling up language models has been shown to predictably improve performance and sample efficiency on a wide range of downstream tasks. This paper instead discusses an unpredictable phenomenon that we refer to as emergent abilities of large language models. We consider an ability to be emergent if it is not present in smaller models but is present in larger models. Thus, emergent abilities cannot be predicted simply by extrapolating the performance of smaller models. The existence of such emergence raises the question of whether additional scaling could potentially further expand the range of capabilities of language models.", "keywords": "", "qa_pairs": [{"question": "What additional analyses were conducted to further compare Gopher and Chinchilla models, and what insights were gained from these analyses?", "answer": "Additional analyses were conducted by stratifying MMLU results into four categories (Humanities, STEM, Social Science, and other) with three x-axes (training FLOPs, model params, and WikiText ppl). The most emergent categories were humanities and social science, while the least emergent category was STEM, with specific sub-topics like high school math, abstract algebra, college math, and high school physics being the least emergent. This suggests that language models struggle with tasks requiring multi-step abstract reasoning. These results are summarized in a new plot (Figure 8) and detailed in Appendix Section A.3.", "category": "Experimental_Analysis"}, {"question": "What other techniques, besides prompt-based testing, can be used to evaluate the performance of language models?", "answer": "In addition to prompting, language models can be evaluated through finetuning on specific tasks, as demonstrated by Nye et al. (2021) in 'Scratchpads for intermediate computation,' where finetuning enabled models to perform 8-digit and 9-digit addition at sufficient scale (shown in Figure 3C). Multi-task finetuning followed by zero-shot prompting is also a viable approach (Figure 3B). Another method is calibration, as shown in the recent paper 'Language models (mostly) know what they know' by Kadavath et al. (2022), where directly asking a model whether an answer is true/false outperformed standard calibration methods, particularly at the scale of the largest 52B language model (Figure 3D).", "category": "Method_Mechanics"}, {"question": "What additional performance tests, beyond prompt-based ones, were conducted to evaluate the language models?", "answer": "Besides prompt-based tests, the language models were evaluated through finetuning on specific tasks, as demonstrated in 'Scratchpads for intermediate computation' by Nye et al. 2021, showing emergent abilities in 8-digit and 9-digit addition with sufficient model scale (Figure 3C). Multi-task finetuning followed by zero-shot prompting was also used (Figure 3B). Additionally, calibration methods from Kadavath et al. 2022, which involve directly asking the model true/false questions, were employed and shown to outperform standard calibration methods, particularly in the largest 52B model (Figure 3D).", "category": "Method_Mechanics"}, {"question": "What are the similarities and differences in emerging capabilities between language models trained on English data and multilingual models, and what trends have been observed?", "answer": "Scaling up language models enables emergent performance in multilingual tasks, such as Hindi knowledge, Hindi toxicity, and Swahili–English proverbs for LaMDA. However, emergence varies across models; for example, GPT-3 outperforms human raters in Hindi QA, while LaMDA does not, and Persian QA emerges for PaLM 62B but not for LaMDA 137B. These differences may be due to variations in training data, such as PaLM 62B using 22% non-English data compared to LaMDA's 10%. Results in Subsection 5.6 and Figure 2(D)", "category": "Experimental_Analysis"}, {"question": "What are some promising directions for future work in the context of the paper?", "answer": "Promising directions for future work include: (1) Further model scaling to enable new abilities, despite computational and hardware challenges. (2) Improved model architecture and scaling, such as mixture-of-experts, adaptive computation, and memory-augmented models, to enhance performance beyond simple scaling. (3) Data scaling as an orthogonal approach to model scaling, requiring more data and training FLOPs but not additional model parameters (e.g., Chinchilla). (4) Better techniques for and understanding of prompting, such as chain-of-thought prompting, to elicit emergent abilities in models. (5) Exploration of frontier tasks that are currently not achievable by even the largest language models (e.g., dozens of tasks in BIG-Bench), with a focus on understanding why these abilities have not emerged and how to enable them.", "category": "Experimental_Analysis"}, {"question": "What are the potential new abilities that future larger models may have, and what are the potential research directions for achieving larger scale beyond current dense language models?", "answer": "Appendix E.4, which lists 45 BIG-Bench tasks where current few-shot prompted models, including PaLM, have not achieved better-than-random performance. These tasks are potential candidates for future emergence if more-capable models achieve better-than-random performance in a few-shot setting. Additionally, Section 5.6 discusses potential future research directions for training better models, such as improved model architectures and training, data quality and scaling, and better prompting techniques, rather than simply scaling, which is challenging due to computational costs and hardware and optimization constraints.", "category": "Experimental_Analysis"}, {"question": "How does the method remain effective when applied to a dataset that is not exchangeable?", "answer": "The method remains effective by slightly altering the dataset to address the source of non-exchangeability. For instance, if non-exchangeability arises from ascending IDs (e.g., in SQuAD and HumanEval), the dataset can be adjusted by removing these IDs or permuting the examples while keeping IDs constant. ", "category": "Method_Mechanics"}, {"question": "What are the factors and theoretical explanations behind the emergence of abilities such as few-shot prompting, and how do different architectures contribute to these abilities?", "answer": "The emergence of abilities like few-shot prompting is influenced by various factors beyond scale, including data properties and example frequencies. For instance, few-shot learning emerges when pretraining documents have long-range coherence, requiring the language model to infer a latent document-level concept to generate coherent next tokens (Xie et al., 2021). In-context learning emerges due to burstiness and having large numbers of rare classes (Chan et al., 2022). Computational psycholinguistics studies emergent behavior in smaller neural networks, such as syntactic rule-learning, where models acquire abilities like subject-verb agreement at specific thresholds of training data (Zhang et al., 2021; Abend et al., 2017). While these studies are on smaller models and different tasks, they suggest that certain data properties or architectures (e.g., transformers, LSTMs) can unlock emergent behavior.", "category": "Experimental_Analysis"}, {"question": "What evidence supports the use of training FLOPs as a metric in Figure 2, and how does the inclusion of Figure 10 with model parameters on the x-axis address potential limitations in interpreting the results?", "answer": "Training FLOPs and model parameters are typically increased simultaneously in large language models, so both metrics are plotted in Figures 2 and 3. Figure 10, which replicates Figure 2 with model parameters on the x-axis, is included to clarify that models usually increase both parameters and training FLOPs together. This addresses the limitation of interpreting results solely based on training FLOPs by providing an alternative view. Additionally, the Chinchilla paper, which shows better performance with decreased model parameters and increased data size, supports the use of training FLOPs as a metric. The caption of Figure 2 also clarifies that each point represents a separate model, not intermediate training steps, to avoid misleading interpretations.", "category": "Experimental_Analysis"}, {"question": "Why was model performance compared based on training FLOPs rather than model size, given the same training resources?", "answer": "Model performance was compared based on training FLOPs rather than model size due to the high computational cost of pre-training large LMs, which prevents perfectly controlled comparisons ablating each factor. LMs are scaled on three axes: training FLOPs, model parameters, and data size. Figures 2 and 3 show comparisons in both training FLOPs and model parameters. Figure 10 and 11 focus on varying model parameters. The Chinchilla paper, which has the same training FLOPs as Gopher but decreases model parameters while increasing data size, motivated this approach. The paper emphasizes viewing emergence as a function of multiple variables. Figure 2 clarifies that each point represents a different model, not the same model at intermediate training steps. Comparing models with both training FLOPs and data scale held constant would be biased due to potential over-training of small models or under-training of large models, and the high cost of training large models makes such comparisons impractical.", "category": "Experimental_Analysis"}]}
{"id": "eGLdVRvvfQ", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "DEUP: Direct Epistemic Uncertainty Prediction", "authors": ["Salem Lahlou", "Moksh Jain", "Hadi Nekoei", "Victor Ion Butoi", "Paul Bertin", "Jarrid Rector-Brooks", "Maksym Korablyov", "Yoshua Bengio"], "abstract": "Epistemic Uncertainty is a measure of the lack of knowledge of a learner which diminishes with more evidence. While existing work focuses on using the variance of the Bayesian posterior due to parameter uncertainty as a measure of epistemic uncertainty, we argue that this does not capture the part of lack of knowledge induced by model misspecification. We discuss how the excess risk, which is the gap between the generalization error of a predictor and the Bayes predictor, is a sound measure of epistemic uncertainty which captures the effect of model misspecification. We thus propose a principled framework for directly estimating the excess risk by learning a secondary predictor for the generalization error and subtracting an estimate of aleatoric uncertainty, i.e., intrinsic unpredictability. We discuss the merits of this novel measure of epistemic uncertainty, and highlight how it differs from variance-based measures of epistemic uncertainty and addresses its major pitfall. Our framework, Direct Epistemic Uncertainty Prediction (DEUP) is particularly interesting in interactive learning environments, where the learner is allowed to acquire novel examples in each round. Through a wide set of experiments, we illustrate how existing methods in sequential model optimization can be improved with epistemic uncertainty estimates from DEUP, and how DEUP can be used to drive exploration in reinforcement learning. We also evaluate the quality of uncertainty estimates from DEUP for probabilistic image classification and predicting synergies of drug combinations.", "keywords": "", "qa_pairs": [{"question": "What is the rationale behind the 'fit-the-fit' approach in DEUP, and how does it address the limitations of requiring OOD data or validation sets?", "answer": "The 'fit-the-fit' approach in DEUP does not assume access to OOD data, as highlighted in the referenced work 'Density of States Estimation for Out-of-Distribution Detection'. While Figure 1 might suggest the need for a validation set, this is only a pedagogical example to introduce DEUP. The focus is on interactive settings where OOD data is 'free', meaning it is obtained at a certain timestep before the main predictor is trained on it. DEUP does not require OOD data, withheld training data, or validation data in the settings of interest. OOD detection is just one of many tasks that benefit from accurate uncertainty estimation. DEUP is evaluated in active learning settings, where high epistemic uncertainty drives exploration. Even if the fit is perfect on a given dataset, the learner can still encounter poor fits on new examples, prompting further investigation and model improvement, as demonstrated in the SMO experiments.", "category": "Method_Mechanics"}, {"question": "Why was the work 'Density of States Estimation for Out-of-Distribution Detection' not referenced, and how does DEUP differ from this approach in terms of OOD data assumptions and application scope?", "answer": "The work 'Density of States Estimation for Out-of-Distribution Detection' was not referenced because the authors were not aware of it, but they will add it. DEUP does not assume access to OOD data, and its focus is on interactive settings where OOD data is 'free' (data obtained before the main predictor is trained on it). Unlike the referenced work, OOD detection is just one of many tasks that benefit from accurate uncertainty estimation, and DEUP is evaluated in active learning settings where high epistemic uncertainty drives exploration.", "category": "Method_Comparison"}, {"question": "Why is it claimed that DEUP can be adapted to use additional features, and how does this relate to the main model's access to these features?", "answer": "The claim about DEUP's adaptability to additional features was misleading. In principle, extra features could be used as inputs to the main model, making discrepancy-based measures of EU implicitly rely on them. The intended point was that excess risk depends on both x and the training dataset, allowing features conveying information about both to be used in the excess risk predictor, as explained from Equation 11 to Equation 12.", "category": "Motivation_Analysis"}, {"question": "Why does DEUP not address the potential for systematic discrepancies uncovered by fit-the-fit to be passed upstream to improve the main model, and how does it differ from Bayesian model criticism and Masegosa's predictive spin?", "answer": "DEUP does not directly address the shortcomings of discrepancy-based measures of EU, except for their assumption that model uncertainty is fixed and can be ignored. DEUP argues that such measures are incomplete and suggests using predicted excess risk as a measure of EU in crucial decision-making settings (e.g., interactive settings), as it captures both model uncertainty and approximation uncertainty. This is detailed in Section 2 of the paper.", "category": "Method_Comparison"}, {"question": "Why does the hypothesis space H' in Figure 2 (right) not shrink or stay the same as more data is used for training, contrary to the Bayesian perspective?", "answer": "H and H' represent the set of hypotheses considered before training, i.e., the prior in a Bayesian sense. With more data, the optimal capacity increases, allowing better generalization by having a larger prior hypothesis space (H' > H). This is due to a trade-off between bias and variance, where more data implies greater capacity and smaller bias. Additionally, the effective capacity in neural networks, influenced by SGD training and early stopping, increases with more training iterations, leading to a larger effective capacity with more data.", "category": "Experimental_Analysis"}, {"question": "How does the definition of epistemic uncertainty as approximation uncertainty + model uncertainty/bias align with the traditional view that epistemic uncertainty is reducible by gaining access to more data, given that model bias is typically assumed to be fixed?", "answer": "In traditional analyses, model bias is assumed to be fixed, but in practical models like neural networks, model bias is not fixed and is reducible. This reduction is influenced by factors such as learning algorithms and techniques like early stopping, as argued in Figure 2 and Section 2.", "category": "Experimental_Analysis"}, {"question": "How is the Spearman's Rank Correlation Coefficient (SRCC) computed between in-distribution (ID) and out-of-distribution (OOD) data, and how does the OOD generalization error between CIFAR-10 and SVHN relate to this computation?", "answer": "The SRCC is computed by calculating the true errors on OOD and unseen ID examples using the predictions from the main predictor and the ground truth labels, and then correlating these errors with the predicted uncertainties. OOD examples have different ranks based on the errors made by the main predictor. The OOD generalization error between CIFAR-10 and SVHN refers to the negative log-likelihood (NLL) loss of the classifier trained on CIFAR-10 when evaluated on SVHN examples, where the 'true' labels of OOD examples correspond to the zero vector, unlike the one-hot vectors of ID examples. This metric is used to establish how well the uncertainty estimates correlate with the errors made by the predictor.", "category": "Method_Mechanics"}, {"question": "Why do Sequential Model Optimization (SMO) methods, such as GP-EI or GP-UCB, perform worse than random acquisition in the one-dimensional example in Figure 3, and how was the length scale for GP experiments determined?", "answer": "Random acquisition is a strong baseline in low-dimensional settings, particularly in the one-dimensional example in Figure 3, because GP-EI and GP-UCB tend to get stuck in local maxima, which were intentionally designed into the function. In the GP experiments in Figure 3, the length scale was optimized at every acquisition step by maximizing the log-likelihood of the available data.", "category": "Experimental_Setup"}, {"question": "Why is it appropriate to formulate the proposed estimator as part of epistemic uncertainty in the context of deep neural networks, despite the traditional view that misspecification in a fixed hypothesis space is irreducible and not part of epistemic uncertainty?", "answer": "Authors agree that in a classical parametric setting, authors consider the hypothesis space $\\mathcal{H}$ fixed, and in that sense the corresponding misspecification is irreducible, thus not something one would want to include as part of epistemic uncertainty. However, consider the modern setting of (typically large) deep neural networks. As more data is collected (e.g. after each round of active learning), authors are allowed to choose a larger neural network (and this may be the optimal thing to do in terms of validation error) or to train the old one for longer (e.g., using validation error early stopping will allow for a larger number of training epochs because the number of training examples has increased, making the optimal effective capacity larger). Effective capacity is not only determined by the size of the network but also by the training time (more time yields greater effective capacity, lower training error). This is why authors propose to quantify epistemic uncertainty in terms of excess risk, i.e., the part of the error that *can be reduced* if more data (i.e. knowledge) is acquired or more compute is available (to extract more knowledge from the data). ", "category": "Experimental_Analysis"}, {"question": "Why is it incorrect to rephrase the bias due to the model being restricted to space H as epistemic uncertainty, and how does authors's paper address this distinction?", "answer": "It is incorrect because the bias is irreducibly conditioned on the space H and is not the same as regular epistemic uncertainty. Authors' paper focuses on learners that can change their hypothesis space $\\mathcal{H}$, and for these learners, incorporating the bias into any measure of epistemic uncertainty (EU) is crucial, as demonstrated in the experiments and related works. Authors will clarify this distinction in the introduction and Section 2.", "category": "Experimental_Analysis"}]}
{"id": "iO4LZibEqW", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Holistic Evaluation of Language Models", "authors": ["Percy Liang", "Rishi Bommasani", "Tony Lee", "Dimitris Tsipras", "Dilara Soylu", "Michihiro Yasunaga", "Yian Zhang", "Deepak Narayanan", "Yuhuai Wu", "Ananya Kumar", "Benjamin Newman", "Binhang Yuan", "Bobby Yan", "Ce Zhang", "Christian Cosgrove", "Christopher D Manning", "Christopher Re", "Diana Acosta-Navas", "Drew A. Hudson", "Eric Zelikman", "Esin Durmus", "Faisal Ladhak", "Frieda Rong", "Hongyu Ren", "Huaxiu Yao", "Jue WANG", "Keshav Santhanam", "Laurel Orr", "Lucia Zheng", "Mert Yuksekgonul", "Mirac Suzgun", "Nathan Kim", "Neel Guha", "Niladri S. Chatterji", "Omar Khattab", "Peter Henderson", "Qian Huang", "Ryan Andrew Chi", "Sang Michael Xie", "Shibani Santurkar", "Surya Ganguli", "Tatsunori Hashimoto", "Thomas Icard", "Tianyi Zhang", "Vishrav Chaudhary", "William Wang", "Xuechen Li", "Yifan Mai", "Yuhui Zhang", "Yuta Koreeda"], "abstract": "Language models (LMs) are becoming the foundation for almost all major language technologies, but their capabilities, limitations, and risks are not well understood. We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models. First, we taxonomize the vast space of potential scenarios (i.e. use cases) and metrics (i.e. desiderata) that are of interest for LMs. Then we select a broad subset based on coverage and feasibility, noting what’s missing or underrepresented (e.g. question answering for neglected English dialects, metrics for trustworthiness). Second, we adopt a multi-metric approach: We measure 7 metrics (accuracy, calibration, robustness, fairness, bias, toxicity, and efficiency) for each of 16 core scenarios to the extent possible (87.5% of the time), ensuring that metrics beyond accuracy don’t fall to the wayside, and that trade-offs across models and metrics are clearly exposed. We also perform 7 targeted evaluations, based on 26 targeted scenarios, to more deeply analyze specific aspects (e.g. knowledge, reasoning, memorization/copyright, disinformation). Third, we conduct a large-scale evaluation of 30 prominent language models (spanning open, limited-access, and closed models) on all 42 scenarios, including 21 scenarios that were not previously used in mainstream LM evaluation. Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common. We improve this to 96.0%: now all 30 models have been densely benchmarked on a set of core scenarios and metrics under standardized conditions. Our evaluation surfaces 25 top-level findings concerning the interplay between different scenarios, metrics, and models. For full transparency, we release all raw model prompts and completions publicly for further analysis, as well as a general modular toolkit for easily adding new scenarios, models, metrics, and prompting strategies. We intend for HELM to be a living benchmark for the community, continuously updated with new scenarios, metrics, and models.", "keywords": "", "qa_pairs": [{"question": "What is the basis for the claim that the carbon intensity estimates are 'of the right order of magnitude', given the significant differences in energy sources across regions?", "answer": "The claim 'right order of magnitude' refers to the estimation error, not misspecification. While regional carbon intensity values (e.g., France's nuclear power) may differ from specific cluster intensities (e.g., Jean-Zay cluster used to train BLOOM), the deviation is expected to be small within regions. Additionally, finer-grained public data is unavailable, and no emissions estimates are provided if the cluster location is unknown.", "category": "Experimental_Analysis"}, {"question": "What are the inter-rater agreement values for the annotations, and what quality control mechanisms and onboarding procedures were implemented for the crowd workers?", "answer": "The inter-rater agreement values were computed using Krippendorff’s $\\alpha$ for each annotation item. For **Reiteration**, the values are: support thesis: 0.593, style: 0.576. For **Wedging**, the values are: address audience: 0.536, support goal: 0.489, style: 0.399, divisive: 0.257, toxic: 0.510. The quality control procedure is described on page 71, and the annotation guidelines are provided in 8.5.1.", "category": "Experimental_Setup"}, {"question": "How can the execution and maintenance of the benchmark be made more cost-effective while still ensuring comprehensive evaluation of all models, scenarios, and metrics?", "answer": "Future work could reduce the evaluation cost by identifying the 'principal components' of HELM, framing the evaluation as a constrained optimization problem to minimize cost while maximizing the ability to generate results predictive of the full gold-standard HELM evaluation. These possibilities are outlined in Appendix H, which introduces a priority system with analogies to the knapsack problem and core set selections.", "category": "Method_Mechanics"}, {"question": "What is the distinction between 'fairness' and 'bias & stereotypes' in the context of Sections 4.6 and 4.7, and why are they not merged as sub-metrics of the 'fairness' category?", "answer": "Fairness refers to disparities in the presence of desired outcomes, specifically the performance of the model when compared to the ground truth answer, and is operationalized for classification scenarios. Bias refers to disparities in the absence of desired outcomes, specifically the performance of the model without consideration for the ground truth answer, and is operationalized for generative scenarios. These distinctions are based on well-established notions in the literature, and thus they are not merged as sub-metrics of the 'fairness' category.", "category": "Method_Disambiguation"}, {"question": "Is the phrase 'right order of magnitude' used to indicate that the estimation error in carbon intensity within regions is expected to be small?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is an estimate of emissions provided when the location of the data center cluster is unknown?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the statement 'should be of right order of magnitude' refer to the exact carbon intensity differences between data centers powered by different energy sources?", "answer": "False", "category": "Claim_Verification"}, {"question": "Was the inter rater agreement (Krippendorff's alpha) provided for each annotation item in the study?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the quality control mechanisms and onboarding procedures for crowd workers described in the paper?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are finer-grained public data on specific carbon intensities of data centers mentioned as available in the paper?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "CD9Snc73AW", "venue": "TMLR 2024", "paper_decision": "Accept with minor revision", "title": "Improving and generalizing flow-based generative models with minibatch optimal transport", "authors": ["Alexander Tong", "Kilian FATRAS", "Nikolay Malkin", "Guillaume Huguet", "Yanlei Zhang", "Jarrid Rector-Brooks", "Guy Wolf", "Yoshua Bengio"], "abstract": "Continuous normalizing flows (CNFs) are an attractive generative modeling technique, but they have been held back by limitations in their simulation-based maximum likelihood training. We introduce the generalized conditional flow matching (CFM) technique, a family of simulation-free training objectives for CNFs. CFM features a stable regression objective like that used to train the stochastic flow in diffusion models but enjoys the efficient inference of deterministic flow models. In contrast to both diffusion models and prior CNF training algorithms, CFM does not require the source distribution to be Gaussian or require evaluation of its density. A variant of our objective is optimal transport CFM (OT-CFM), which creates simpler flows that are more stable to train and lead to faster inference, as evaluated in our experiments. Furthermore, we show that when the true OT plan is available, our OT-CFM method approximates dynamic OT. Training CNFs with CFM improves results on a variety of conditional and unconditional generation tasks, such as inferring single cell dynamics, unsupervised image translation, and Schrödinger bridge inference. The Python code is available at https://github.com/atong01/conditional-flow-matching.", "keywords": "", "qa_pairs": [{"question": "Why does SB-CFM underperform compared to DSB in Table 4 for higher-dimensional data, and under what conditions is SB-CFM expected to outperform DSB?", "answer": "In Table 4, SB-CFM underperforms compared to DSB because a better Schrödinger bridge does not necessarily correlate with better performance on the task, which involves extrapolating out of distribution on how observed cells evolve. This suggests that the Schrödinger bridge with σ=1 (the training objective) does not match biology as well as the optimal transport interpolant (σ=0) or random pairing. SB-CFM is expected to perform better on generative modeling tasks, particularly in terms of speed of training (due to its simulation-free nature) and speed of inference (due to using an ODE). This is demonstrated by an experiment on CIFAR-10, where SB-CFM achieves comparable results (3.7 FID) to an improved version of DSB (SB-FBSDE, 3.01 FID) while being faster (16 A100 hours), and with a much lower memory footprint (no need for caching trajectories), and only requires a single rather than 2 models.", "category": "Experimental_Analysis"}, {"question": "Why were 100 steps used for simulation in Table 5 instead of fewer steps, and how does OT-CFM compare to traditional diffusion models in terms of sample quality with the same number of steps?", "answer": "The inclusion of 100 steps was to ensure thoroughness in the simulation. For comparison, results from [Lipman et al. 2023] show an FID of 7.48 for DDPM and 6.35 for FM, indicating that while flow matching may not achieve state-of-the-art performance in absolute terms, augmenting Flow Matching with optimal transport ideas can improve performance and approximate dynamic OT in some settings.", "category": "Method_Comparison"}, {"question": "What are the real-life applications and rationale for training the drift using a neural network with access to a static OT/OT-regularized coupling?", "answer": "The real-life applications include (1) generative modeling of images and (2) modeling of time series in single-cell. The rationale is that access to the OT-regularized coupling improves state-of-the-art methods in these tasks with little extra cost.", "category": "Motivation_Analysis"}, {"question": "What is the utility of approximating the dynamic optimal transport (OT) by a neural network when the discrete static OT plan between training samples is already available?", "answer": "While the discrete dynamic OT map can be computed from the OT plan using McCann interpolants, it cannot produce novel trajectories because the OT plan is computed between fixed source and target distributions. The static OT plan can only generate real target samples from the source empirical distribution. In generative modeling, the goal is to generate novel samples from an approximation to the data distribution. This has been used in the single-cell setting for example in Cell-OT [Bunne et al. 2023] which trained a neural network from small datasets to approximate the optimal transport pushforward which can be subsequently applied to push forward new source datasets. This approach has been used in settings like Cell-OT and is analogous to diffusion models, where neural networks are used to amortize the computation of the true score, inducing smoothness and enabling the generation of high-quality samples. Similarly, in this work, a neural network is used to amortize the optimal transport ODE or the probability flow ODE of the Schrödinger bridge, allowing for fast generation of new samples with induced smoothness from both the neural network and the minibatch approximation to the static OT.", "category": "Motivation_Analysis"}, {"question": "What is the practical utility of using minibatch optimal transport (OT) in OT-CFM and SB-CFM, and how does it compare to the complete OT plan?", "answer": "In OT-CFM, the OT plan is computed between a source and a target minibatch using an incomplete OT plan estimator with one pair of source and target minibatch couple (Eq.7 in [1] with k=1). This construction is equivalent to drawing samples from the minibatch OT plan in expectation, and it converges linearly to the complete minibatch OT plan as the number of source and target minibatch couples grows. When entropic OT is used in SB-CFM, the minibatch entropic OT plan is computed in expectation as well. Using one pair of minibatch couples is common in deep learning to achieve state-of-the-art performance[2,3]. The question of how close minibatch OT is to OT remains an open question, though the distance between minibatch OT and OT as a cost has been studied in [4].\n[1] Learning with minibatch Wasserstein : asymptotic and gradient properties, Fatras et al, AISTATS 2020.\n\n[2] Learning Generative Models with Sinkhorn Divergences, Genevay et al., AISTATS 2018.\n\n[3] DeepJDOT: Deep Joint Distribution Optimal Transport for Unsupervised Domain Adaptation, Damadoran et al, ECCV 2018.\n\n[4] Optimal transport: Fast probabilistic approximation with exact solvers, Sommerfeld et al., JMLR 2019.", "category": "Method_Comparison"}, {"question": "How does the use of Brownian motion as the reference process in authors' Gaussian to Gaussian experiment align with the DSB paper's use of Ornstein Uhlenbeck in synthetic data experiments, and why is this configuration chosen?", "answer": "While the DSB paper uses Ornstein Uhlenbeck (OU) in synthetic data experiments, it employs Brownian motion (a special case of OU) as the reference process in the MNIST → MNIST experiment. Authors adopted this configuration for Gaussian to Gaussian experiment to align with the DSB approach in this specific setting. Authors will clarify this in the text.", "category": "Motivation_Analysis"}, {"question": "Can any diffusion process, including one from SB, be converted into a flow using the probability flow ODE, and can these solvers be used?", "answer": "Yes, any pair of SDEs can be converted into an ODE by averaging the forward SDE drift and negative backward SDE drift. At convergence, when forward and backward SDEs are reverses of each other, this equals the probability flow ODE. However, since the DSB is trained with a fixed discretization, in practice it performs best when integrated with this same discretization.", "category": "Experimental_Setup"}, {"question": "What is the justification for the convergence of flow matching when using different minibatch couplings or samples from couplings?", "answer": "The justification lies in the fact that as long as the marginals of the couplings match the initial and terminal data distributions, the specific coupling used for CFM does not affect the outcome. Independent couplings (minibatch size 1) and arbitrary couplings (e.g., OT or entropic OT) between pairs of minibatches sampled uniformly from the data both lead to correct marginals. At intermediate time points, the objective being a least-squares regression with a stochastic target has a unique optimum, equal to the expectation of this target. As the loss decreases to its minimum, the learned vector field approaches this expectation, which is the drift of an ODE transforming the initial distribution to the terminal one. ", "category": "Method_Mechanics"}, {"question": "What are the key differences between the proposed method and Chen et al. (2022) for generative modeling, particularly in terms of training procedures and performance?", "answer": "Chen et al. (2022) assumes continuous time but relies on discretization and simulation steps during training, as it trains on stochastic trajectories sampled from the model. Additionally, their approach includes a denoising score matching pretraining stage, which runs for 200k iterations and accounts for a substantial portion of the training steps. Better performance can be achieved by continuing this pretraining until convergence, as shown in Song et al. Furthermore, generation from diffusion models as SDEs yields better results than integrating equivalent ODEs, with recent work suggesting that time discretization degrades the DDIM near the noise endpoint.", "category": "Method_Comparison"}, {"question": "Is the code for all baselines publicly available either in the torchcfm package or the original package?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "AFDcYJKhND", "venue": "TMLR 2024", "paper_decision": "Accept as is", "title": "Scaling Autoregressive Models for Content-Rich Text-to-Image Generation", "authors": ["Jiahui Yu", "Yuanzhong Xu", "Jing Yu Koh", "Thang Luong", "Gunjan Baid", "Zirui Wang", "Vijay Vasudevan", "Alexander Ku", "Yinfei Yang", "Burcu Karagol Ayan", "Ben Hutchinson", "Wei Han", "Zarana Parekh", "Xin Li", "Han Zhang", "Jason Michael Baldridge", "Yonghui Wu"], "abstract": "We present the Pathways Autoregressive Text-to-Image (Parti) model, which generates high-fidelity photorealistic images and supports content-rich synthesis involving complex compositions and world knowledge. Parti treats text-to-image generation as a sequence-to-sequence modeling problem, akin to machine translation, with sequences of image tokens as the target outputs rather than text tokens in another language. This strategy can naturally tap into the rich body of prior work on large language models, which have seen continued advances in capabilities and performance through scaling data and model sizes. Our approach is simple: First, Parti uses a Transformer-based image tokenizer, ViT-VQGAN, to encode images as sequences of discrete tokens. Second, we achieve consistent quality improvements by scaling the encoder-decoder Transformer model up to 20B parameters, with a new state-of-the-art zero-shot FID score of 7.23 and finetuned FID score of 3.22 on MS-COCO. Our detailed analysis on Localized Narratives as well as PartiPrompts (P2), a new holistic benchmark of over 1600 English prompts, demonstrate the effectiveness of Parti across a wide variety of categories and difficulty aspects. We also explore and highlight limitations of our models in order to define and exemplify key areas of focus for further improvements.", "keywords": "", "qa_pairs": [{"question": "How is the conv-shaped masked sparse attention implemented in the decoder, and what are its primary benefits in terms of speed, memory, and image generation quality compared to full attention?", "answer": "The conv-shaped masked sparse attention is implemented with local sparse attention featuring row, column, and conv-window visibility, as detailed in DALL-E's Appendix B.1 Figure 11. The benefits of sparse attention in terms of speed and memory are highly dependent on the implementation and hardware. For training Parti models, masked full-attention (with conv-shape masking) did not provide speed or memory gains. However, during sampling/decoding, conv-shape masked sparse attention resulted in a ~2x speedup in latency for the 20B model.", "category": "Method_Mechanics"}, {"question": "What is the specific prompt used for unconditioned training G(z) in the context of CF-guidance as described in Section 2.4?", "answer": "For CF guidance training, no specific text prompt is used for unconditioned inputs. Instead, all text information is masked out by setting text ids to zeros and text paddings to ones, resulting in a fixed learned vector from the transformer encoder output, which serves as the unconditioned input during CF-guidance sampling.", "category": "Method_Mechanics"}, {"question": "Can the Parti model be used as a source of knowledge for other tasks in the NLP domain, such as GPT-3?", "answer": "The Parti model’s encoder is capable of performing on GLUE benchmarks, as shown in Appendix G Table 9, in comparison to BERT and other multimodal models. Parti is a 'language model' approach, and its results aim to motivate work on unifying text and multimodal models and developing better benchmarks to evaluate models' capabilities, such as visually-grounded text understanding.", "category": "Method_Mechanics"}, {"question": "What are the empirical reasons for using a small ViT-VQGAN encoder tokenizer and a large decoder, rather than using the large model for both encoding and decoding?", "answer": "The choice of using a small ViT-VQGAN encoder tokenizer and a large decoder is based on empirical reasons: (1) Increasing the decoder size improves FID (the primary quality metric of image tokenizer) more effectively than increasing the encoder size under compute constraints. For example, a small-size encoder and huge-size decoder achieve an FID of 1.16, which is better than other configurations. Base-size encoder and base-size decoder achieves FID of 1.55 on ImageNet validation set with 30k examples, large-size encoder and large-size decoder achieves FID of 1.42 (2) A smaller encoder speeds up the training of the sequence-to-sequence model, as only the encoder path is activated during training, while the decoder path is used during inference/sampling.\n|  name | num_layers | num_heads | model_dims | mlp_dims |\n|:-----:|:----------:|:---------:|:----------:|:--------:|\n| small |      8     |     12    |     512    |   2048   |\n|  base |     12     |     12    |     768    |   3072   |\n| large |     24     |     16    |    1024    |   4096   |\n|  huge |     32     |     16    |    1280    |   5120   |\n", "category": "Motivation_Analysis"}, {"question": "Why was CoCa not used for all tasks, given its superior performance in image captioning and zero-shot retrieval compared to SimVLM and ALIGN?", "answer": "CoCa was not used due to practical timeline limitations. The captioning data was prepared and the Parti models were trained before CoCa became available. Retraining Parti models would require significant computation and time. Additionally, while CoCa performs better than SimVLM in image captioning, the performance gap (e.g., CIDEr score) is not substantial, so the difference between SimVLM and CoCa does not significantly impact the results presented in this work.", "category": "Motivation_Analysis"}, {"question": "What are the CLIP-R-Precision scores for the retrieval baseline on the MS-COCO and LN-COCO datasets?", "answer": "The CLIP-R-Precision scores for the retrieval baseline on both MS-COCO and LN-COCO datasets are around 98.6-99%. Consistent results are observed across different subsets of 30k examples from mscoco2014, which is expected due to the 'uniquely identifying' embedding capability of the CLIP/ALIGN model in text-to-image retrieval.", "category": "Experimental_Setup"}, {"question": "Given the absence of ground truth images for P2 prompts, what alternative evaluation methods besides human evaluation have been used to assess the quality of text-to-image generation methods?", "answer": "In the absence of ground truth images for P2 prompts, the evaluation of text-to-image generation methods includes multiple metrics such as FID, finetuned FID, next-token-prediction loss, DALL-Eval, Human Side-by-Side, and retrieval baseline. These methods aim to provide a comprehensive assessment and inspire the development of better, standardized, and automated evaluation benchmarks.", "category": "Method_Mechanics"}]}
{"id": "wLg9JrwFvL", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Normalized/Clipped SGD with Perturbation for Differentially Private Non-Convex Optimization", "authors": ["Runze Li", "Dahun Kim", "Bir Bhanu", "Weicheng Kuo"], "abstract": "By ensuring differential privacy in the learning algorithms, one can rigorously mitigate the risk of large models  memorizing sensitive training data. In this paper, we study two algorithms for this purpose, i.e., DP-SGD and DP-NSGD, which first clip or normalize \\textit{per-sample} gradients to bound the sensitivity and then add noise to obfuscate the exact information. We analyze the convergence behavior of these two algorithms in the non-convex empirical risk minimization setting with two common assumptions and achieve a rate $\\mathcal{O}\\left(\\sqrt[4]{\\frac{d\\log(1/\\delta)}{N^2\\epsilon^2}}\\right)$ of the gradient norm for a $d$-dimensional model, $N$ samples and $(\\epsilon,\\delta)$-DP, which improves over previous bounds under much weaker assumptions.  Specifically, we introduce a regularizing factor in DP-NSGD and show that it is crucial in the convergence proof and subtly controls the bias and noise trade-off. Our proof deliberately handles the per-sample gradient clipping and normalization  that are specified for the private setting.   Empirically, we demonstrate that these two algorithms achieve similar best accuracy while DP-NSGD is  comparatively easier to tune than DP-SGD.", "keywords": "", "qa_pairs": [{"question": "How does the proposed method improve upon existing algorithms in terms of handling bias and convergence, as shown in Table 1?", "answer": "The proposed DP-NSGD does not aim to eliminate bias from gradient clipping/normalization but provides a theoretical analysis of how this bias affects convergence and model utility. It exhibits comparable convergence analysis with DP-SGD.", "category": "Method_Disambiguation"}, {"question": "Is it accurate to state that as training progresses, the gradient norm becomes small and clipping has no effect, given that some data points may still have large gradients?", "answer": "While most data points become well-fitted and have small gradient magnitudes as training progresses, some data points may still not be fitted and exhibit large gradients. Gradient normalization (DP-NSGD) amplifies the impact of samples with small gradients, leading to different training curves compared to gradient clipping (DP-SGD). ", "category": "Experimental_Analysis"}, {"question": "What evidence supports the claim that the amplification effect of DP-NSGD on the significance of instances with small gradients is more pronounced than that of DP-SGD, and how does this relate to gradient clipping in differentially private optimization?", "answer": "A recent study by He et al. (2023), titled 'Exploring the Limits of Differentially Private Deep Learning with Group-Wise Clipping,' uses an adaptive quantile estimation method for setting the clipping threshold. This method aims for a target quantile of 0.85, implying that 85% of gradients remain unclipped at each iteration. Their experiments on large language model fine-tuning tasks show that during the latter stages of differentially private training, sample gradients exhibit two distinct patterns: well-fitted data gradients become very small, while not-yet-fitted data gradients (around 15% of the total samples) remain significantly larger. Based on these findings, the authors refine their claim to state that 'While both DP-NSGD and DP-SGD tend to amplify the significance of instances with small gradients, the amplification effect of DP-NSGD is generally more pronounced than that of DP-SGD.' This adjustment provides a more nuanced understanding of gradient clipping effects in the context of differentially private optimization techniques, though these observations may not directly apply to image classification scenarios.", "category": "Experimental_Analysis"}, {"question": "Are all data points well-fitted with small gradient magnitudes as training progresses?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the statement on Page 10 imply that gradient clipping has no effect at all as training progresses?", "answer": "False", "category": "Claim_Verification"}, {"question": "Has the term 'good samples' been changed to 'well-fitted samples'?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "5HsBuYYx4i", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Progressive-Hint Prompting Improves Reasoning in Large Language Models", "authors": ["Zeyu Han", "Chao Gao", "Jinyang Liu", "Jeff Zhang", "Sai Qian Zhang"], "abstract": "The performance of Large Language Models (LLMs) in reasoning tasks depends heavily on prompt design, with Chain-of-Thought (CoT) and self-consistency being critical methods that enhance this ability. However, these methods do not fully exploit the answers generated by the LLM to guide subsequent responses. This paper proposes a new prompting method, named Progressive-Hint Prompting (PHP), that enables automatic multiple interactions between users and LLMs by using previously generated answers as hints to progressively guide toward the correct answers. PHP is orthogonal to CoT and self-consistency, making it easy to combine with state-of-the-art techniques to further improve performance. We conducted extensive and comprehensive experiments on seven benchmarks. The results show that PHP significantly improves accuracy while remaining highly efficient. For instance, with text-davinci-003, we observed a 4.2% improvement on GSM8K with greedy decoding compared to Complex CoT, and a 46.17% reduction in sample paths with self-consistency. With GPT-4 and PHP, we achieve state-of-the-art performances on SVAMP (89.1% $\\rightarrow$ 91.9%), GSM8K (92% $\\rightarrow$ 95.5%), AQuA (76.4% $\\rightarrow$ 79.9%) and MATH (50.3% $\\rightarrow$ 53.9%).", "keywords": "", "qa_pairs": [{"question": "How does the proposed PHP method differ from existing methods like Chain-of-Thought and Self-Consistency in terms of utilizing previous round generation to improve performance?", "answer": "The proposed PHP method improves performance by utilizing the previous round generation, unlike Chain-of-Thought which details the reasoning path and Self-Consistency which samples multiple paths and takes a majority vote. PHP enhances performance both with a single reasoning path (compared to CoT) and with multiple paths (compared to Self-Consistency), as shown in Figure 3, with the same cost. Extensive experiments demonstrate that LLMs can effectively use provided hints, as evidenced by significant performance improvements when replacing the base prompt Standard with Complex CoT or CoT in Table 3. Additionally, PHP shows unique behaviors in handling missing or multiple hints,不同于 GPT-3.5-Turbo and text-davinci-003.", "category": "Method_Comparison"}, {"question": "Does the consistency of answers in two consecutive rounds imply that the LLM has reached a definitively correct answer, and what evidence supports the effectiveness of the stop criterion and adaptive sampling?", "answer": "The consistency of answers in two consecutive rounds does not definitively mean the LLM has reached a correct answer, as LLMs can return different answers even with the same input due to network architectures. The stop criterion of 'two rounds with the same answers' is used to reduce computational cost. Evidence from Self-Consistency and Adaptive-Consistency shows that adaptive sampling and multiple sampling with majority voting can improve performance, supporting the effectiveness of the approach.", "category": "Experimental_Analysis"}, {"question": "What is the correlation between the PHP Design Principle and the implemented design methods, as shown in Table 1?", "answer": "The PHP Design Principle considers various situations of hints, including two potential scenarios: 1) The hints are the same as the correct answer, ensuring the model can still get the correct answer when the hint is correct; and 2) The hints are not the same as the correct answer, ensuring the model can jump out of the incorrect answer. These scenarios are reflected in the implemented design methods, as shown in Table 1. * The hints are the same as the correct answer: to be sure that the model can still get the correct answer when the hint is correct. As shown in Table 1, the given hint can be 6, while the correction answer is 6. * The hints are not the same as the correct answer: to be sure that the model can jump out of the incorrect answer. As shown in Table 1, the given hint can be 10, 8, while the correction answer is 6.", "category": "Experimental_Analysis"}, {"question": "How do the interactions between LLMs and prompts in the PHP framework explain the seemingly contradictory observations that more powerful LLMs require fewer interactions while more powerful prompts require more interactions?", "answer": "The interaction number for text-davinci-003 is lower than text-davinci-002 because a more powerful model has a higher probability of providing the correct answer, thus needing fewer interactions. For prompts, the Standard prompt results in fewer interactions than CoT, and CoT fewer than Complex CoT. This is because the quality of the prompt influences the model's ability to learn patterns. Complex CoT outperforms other prompts in instilling a pattern that corrects initial incorrect answers, leading to more interactions but better performance. Within the proposed PHP framework, the established pattern is as follows: 1) if the initial answer is correct, maintain the same correct response in the subsequent round; 2) if the initial answer is incorrect, strive to provide the correct answer in the next round. The Standard prompt lacks this effectiveness, resulting in fewer interactions and less dynamic responses.", "category": "Experimental_Analysis"}, {"question": "What does a large number of rounds indicate about the dataset and the LLM's performance?", "answer": "A large number of rounds indicates that the response answer often changes after each round, suggesting that the dataset is relatively difficult. This is supported by the iteration distribution shown in Table 10, which demonstrates that more challenging datasets like AQuA have a broader range of interaction numbers, indicating the LLM's uncertainty when facing difficult problems. In contrast, simpler datasets like Addsub are predominantly resolved with just 2 interactions, showing that interaction numbers can reliably indicate dataset difficulty.", "category": "Experimental_Exposition"}, {"question": "What are the advantages of the proposed PHP method when compared to tool-based methods for addressing mathematical problems?", "answer": "The advantages of the PHP method include its simplicity and adaptability to different tasks without requiring external tools, as highlighted in the updated paper.", "category": "Method_Comparison"}, {"question": "How does the experimental setup in Appendix A.3 demonstrate the advantages of adaptive sampling and the use of hints in Progressive-Hint Prompting (PHP)?", "answer": "The experimental setup in Appendix A.3 includes three configurations: CoT+N/A (chain-of-thought without adaptive sampling, with only one round of interaction), CoT + CoT (chain-of-thought with adaptive sampling, stopping if two subsequent answers are the same), and CoT + PHP-CoT (Progressive-Hint Prompting, where CoT is used in the first round and PHP-CoT with hints in subsequent rounds, stopping if two subsequent answers are the same). In the first round, authors employ CoT to get the answer. In the subsequent round, authors employ PHP-CoT and the previous answer’s hint to get an answer. If the two subsequent answers are the same, then authors stop. This is used to check the performance gains of hint usage. Therefore, the Appendix 3 experimental setup has proved the advantages of adaptive sampling (terminated when the two subsequent answers are the same). The whole method is called progressive-hint prompting (PHP), as authors progressively add hints and employ multiple rounds to get final answers.", "category": "Experimental_Setup"}, {"question": "What is the rationale behind stopping each sample independently based on that one sample's turns in Progressive-Hint Prompting (PHP), and how does it compare to stopping based on partial agreement across the overall basket of samples?", "answer": "In Progressive-Hint Prompting (PHP), the method progressively adds hints and employs multiple rounds to get final answers. In subsequent rounds, PHP-CoT and the previous answer’s hint are used to obtain an answer. If the two subsequent answers are the same, the process stops. This adaptive sampling approach has demonstrated performance gains, as shown in Appendix A.3. ", "category": "Motivation_Analysis"}, {"question": "Has it been demonstrated by methods like Self-Consistency and Adaptive-Consistency that adaptive sampling can improve performance even if the answer does not change in consecutive rounds?", "answer": "True", "category": "Claim_Verification"}, {"question": "Is the proposed method (PHP) limited to arithmetic problems, without applicability to more complex reasoning tasks like Commonsense reasoning?", "answer": "False", "category": "Claim_Verification"}, {"question": "Is the stop criterion of 'two rounds with the same answers' used to definitively conclude the correctness of the answer?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the LLM achieving the same answer in two consecutive rounds necessarily mean it has reached a fixed answer that will not change with more information?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "ETH1GDbSPw", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Imagen Video: High Definition Video Generation with Diffusion Models", "authors": ["Zeyu Han", "Chao Gao", "Jinyang Liu", "Jeff Zhang", "Sai Qian Zhang"], "abstract": "We present Imagen Video, a text-conditional video generation system based on a cascade of video diffusion models. Given a text prompt, Imagen Video generates high definition videos using a base video generation model and a sequence of interleaved spatial and temporal video super-resolution models. We describe how we scale up the system as a high definition text-to-video model including design decisions such as the choice of fully-convolutional temporal and spatial super-resolution models at certain resolutions, and the choice of the v-parameterization of diffusion models. In addition, we confirm and transfer findings from previous work on diffusion-based image generation to the video generation setting. Finally, we apply progressive distillation to our video models with classifier-free guidance for fast, high quality sampling. We find Imagen Video not only capable of generating videos of high fidelity, but also having a high degree of controllability and world knowledge, including the ability to generate diverse videos and text animations in various artistic styles and with 3D object understanding.", "keywords": "", "qa_pairs": [{"question": "How does the proposed spatial superresolution model compare in terms of artifacts to other existing superresolution methods and frame interpolation methods?", "answer": "The SSR models do not suffer from obvious artifacts that occur when performing video super-resolution by running image super-resolution on each frame separately, which assumes independence over time. Authors' setup predicts multiple output frames conditioned on multiple input frames, avoiding this independence assumption. Authors will also compare the models to existing video super-resolution methods by running them on the low-resolution generated videos.", "category": "Method_Comparison"}, {"question": "What are the possible reasons for the observed improvement in sample quality with larger video models, contrary to findings in existing literature for text-to-image models?", "answer": "The observed improvement in sample quality with larger video models, as measured by FVD and CLIP, is likely due to the networks being underparameterized for the video generative modeling task. Video generative modeling is more complex than image generative modeling, and the current scale of networks for video tasks has not yet reached a sufficient number of parameters, unlike those for image tasks.", "category": "Experimental_Analysis"}, {"question": "In what specific ways was dynamic thresholding insufficient in the initial experiments, particularly in addressing saturation issues?", "answer": "Dynamic thresholding was insufficient because it was challenging to identify a threshold that effectively addressed saturation issues while maintaining sample quality and avoiding artifacts.", "category": "Experimental_Analysis"}, {"question": "What are the failure cases of Imagen Video, and what are the underlying reasons for these failures?", "answer": "The failure cases of Imagen Video include the inability to generate long sequences of actions given by long text prompts, largely due to the limitations of the training data and the length limit of the generated videos, as well as difficulties in achieving samples of extreme photorealism and highly precise details in actions and movements of objects. These issues can potentially be resolved by further scaling of model sizes and data.", "category": "Experimental_Analysis"}]}
{"id": "umFYHBDCcW", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Can We Solve 3D Vision Tasks Starting from A 2D Vision Transformer?", "authors": ["Runze Li", "Dahun Kim", "Bir Bhanu", "Weicheng Kuo"], "abstract": "Vision Transformers (ViTs) have proven to be effective, in solving 2D image understanding tasks by training over large-scale image datasets; and meanwhile as a somehow separate track, in modeling the 3D visual world too such as voxels or point clouds. However, with the growing hope that transformers can become the ``universal'' modeling tool for heterogeneous data, ViTs for 2D and 3D tasks have so far adopted vastly different architecture designs that are hardly transferable. That invites an (over-)ambitious question: can we close the gap between the 2D and 3D ViT architectures? As a piloting study, this paper demonstrates the appealing promise to understand the 3D visual world, using a standard 2D ViT architecture, with only minimal customization at the input and output levels without redesigning the pipeline. To build a 3D ViT from its 2D sibling, we ``inflate'' the patch embedding and token sequence, accompanied with new positional encoding mechanisms designed to match the 3D data geometry. The resultant ``minimalist'' 3D ViT, named \\textbf{Simple3D-Former}, performs surprisingly robustly on popular 3D tasks such as object classification, point cloud segmentation and indoor scene detection, compared to highly customized 3D-specific designs. It can hence act as a strong baseline for new 3D ViTs. Moreover, we note that pursuing a unified 2D-3D ViT design has practical relevance besides just scientific curiosity. Specifically, we demonstrate that Simple3D-Former naturally is able to exploit the wealth of pre-trained weights from large-scale realistic 2D images (e.g., ImageNet), which can be plugged into enhancing the 3D task performance ``for free''. ", "keywords": "", "qa_pairs": [{"question": "How has the scope of the paper been adjusted to address the concern about overclaiming the capabilities of Simple3D-Former?", "answer": "The scope has been refined to specifically discuss the tasks evaluated in the manuscript, including 3D object classifications, 3D part segmentations, 3D indoor scene segmentations, and 3D indoor scene detections. The overclaimed tone in the Introduction and Conclusion has been modified to align with these evaluated tasks.", "category": "Method_Mechanics"}, {"question": "Why is the discussion of Neural Rendering and NeRF not included in the manuscript regarding Simple3D-Former's applications?", "answer": "The novel view synthesis task, which involves Neural Radiance Field, is distinct from the common 3D tasks addressed by Simple3D-Former. Neural Radiance Field requires per-instance optimization and differs in nature from the tasks requiring large-scale 3D data analysis for commonality. Additionally, novel-view synthesis is a more advanced topic involving per-scene optimization and generative priors, which is beyond the current scope of the manuscript.\nexample:\n\nXu, Dejia, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. 'SinNeRF: Training Neural Radiance Fields on Complex Scenes from a Single Image.' arXiv preprint arXiv:2204.00928 (2022).", "category": "Experimental_Analysis"}, {"question": "What are the key architectural differences between Simple3D-Former and Point Transformer, and how do these differences affect their performance on datasets like ModelNet40, ShapeNetPartSeg, and S3DIS?", "answer": "Simple3D-Former is a tokenizer-transformer architecture where the tokenization scheme and downstreaming head are detached from transformer layers, using a composition of transition down layers to embed point cloud sequences into tokenized sequences. It applies positional embedding only once during tokenization. In contrast, Point Transformer adds transformer layers after each transition down and transition up block and applies positional embedding layer-by-layer. The experimental results suggest that the universal, flexible design of Simple3D-Former may not offer a performance advantage over Point Transformer under specific point cloud modality designs. The method can be adapted with different structures, such as PointNeXT [1]. [1] Qian, Guocheng, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. 'PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies.' Advances in Neural Information Processing Systems (2022).", "category": "Method_Comparison"}, {"question": "How does the Simple3D-Former architecture differ from the point-transformer in terms of design flexibility and tokenization scheme for downstream tasks like point-cloud detection and segmentation?", "answer": "The Simple3D-Former is designed as a universal, flexible architecture that can adapt to various downstream tasks. It differs from the point-transformer by emphasizing a tokenizer-transformer architecture where the tokenization scheme is decoupled from the transformer layers, allowing for flexibility in replacing components such as using PointNeXT [1] as an encoder or decoder. Additionally, the positional embedding in Simple3D-Former is applied only once during tokenization, unlike point-transformer, which applies it layer-by-layer. These distinctions are clarified in the Figure 5.\n[1] Qian, Guocheng, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Abed Al Kader Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. 'PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies.' Advances in Neural Information Processing Systems (2022).", "category": "Method_Comparison"}, {"question": "What are the key differences between Simple3D-Former and Point-Transformer in terms of architecture and positional embedding?", "answer": "Simple3D-Former is a tokenizer-transformer architecture with the tokenization scheme and downstream head decoupled from the transformer layers, while Point-Transformer integrates transformer layers directly into the Transition Down and Transition Up blocks. Additionally, Simple3D-Former applies positional embedding only once during tokenization, whereas Point-Transformer applies it layer-by-layer.", "category": "Method_Comparison"}, {"question": "What was the motivation behind designing Simple3D-Former, and why did its performance not surpass existing approaches?", "answer": "The motivation behind designing Simple3D-Former was to demonstrate that 3D tasks with tokenized 3D input can be fit into a self-attention module (i.e., transformers), which is originally applied to 2D data, and to show that pretrained transformers can be applied to 3D tasks with minimal extension by leveraging 2D pretraining and teacher tasks. The goal was not to surpass state-of-the-art performance but to highlight these principles. Performance improvements could potentially be achieved by using different tokenization or transformer backbones.", "category": "Motivation_Analysis"}, {"question": "How does the Simple3D-Former's output sensitivity to input orientations vary when using the group embedding scheme in the canonical space, as demonstrated on the ShapeNet dataset?", "answer": "The Simple3D-Former's output sensitivity to input orientations varies based on the view direction when using the group embedding scheme. In the ShapeNet dataset, the X-dim projection is more powerful compared to other directions. An additional ablation study shows that the group embedding scheme works differently based on the viewing direction, with the YZ plane containing consistent information as shown in Table 4. The 2D+1D group embedding reduces the gap between different view directions compared to naive projection, which averages data along the third dimension. Current model performance varies slightly across different view directions: XYZ (87.41%), YZX (87.65%), and ZXY (87.32%).\n| View Direction      | OA. (%) |\n| ----------- | ----------- |\n| XYZ (XY plane)      | 87.41       |\n| YZX (YZ plane)   | 87.65        |\n| ZXY (ZX plane)   | 87.32 |\n", "category": "Experimental_Exposition"}, {"question": "What approach is recommended for applying the algorithm to real-world applications involving point cloud data with unknown 3D poses?", "answer": "For point cloud data with unknown 3D poses, the network is trained with data augmentation in scaling, rotation, and transition to enhance robustness. Additionally, the scheme can be integrated with previous works on upright orientation learning in both voxel and point cloud modalities to actively learn the poses.", "category": "Method_Mechanics"}]}
{"id": "lPw8Xdzw5f", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Chaos Theory and Adversarial Robustness", "authors": ["Hari Sowrirajan", "Jingbo Yang", "Andrew Y. Ng", "Pranav Rajpurkar"], "abstract": "Neural networks, being susceptible to adversarial attacks, should face a strict level of scrutiny before being deployed in critical or adversarial applications. This paper uses ideas from Chaos Theory to explain, analyze, and quantify the degree to which neural networks are susceptible to or robust against adversarial attacks. To this end, we present a new metric, the \"susceptibility ratio,\" given by $\\hat \\Psi(h, \\theta)$, which captures how greatly a model's output will be changed by perturbations to a given input.\n\nOur results show that susceptibility to attack grows significantly with the depth of the model, which has safety implications for the design of neural networks for production environments. We provide experimental evidence of the relationship between $\\hat \\Psi$ and the post-attack accuracy of classification models, as well as a discussion of its application to tasks lacking hard decision boundaries. We also demonstrate how to quickly and easily approximate the certified robustness radii for extremely large models, which until now has been computationally infeasible to calculate directly.", "keywords": "", "qa_pairs": [{"question": "What is the difference between the Chaos system and algorithmic stability, and is it possible to connect algorithmic stability with neural networks?", "answer": "Algorithmic stability refers to the stability of the learning algorithm in response to changes in the learning problem, rather than the stability of inferences made by an already trained model. The relationship between algorithmic stability and neural networks is discussed in an added paragraph in section 2.2.", "category": "Concept_Understanding"}, {"question": "What is the novelty of the work in the context of related studies on the impact of architecture on adversarial robustness, particularly in terms of theoretical justification and application?", "answer": "The novelty of the work lies in being the first to connect Chaos Theoretic sensitive dependence (the 'Butterfly Effect') to adversarial susceptibility. While related works like Huang et al. 2021 and Wu et al. 2021 provide theoretical justification, their analyses arise from entirely different fields. This connection, though currently with limited practical or experimental implications, is novel and opens avenues for further study.", "category": "Method_Disambiguation"}]}
{"id": "Ufc5cWhHko", "venue": "TMLR 2024", "paper_decision": "Withdraw", "title": "RECLIP: Resource-efficient CLIP by Training with Small Images", "authors": ["Runze Li", "Dahun Kim", "Bir Bhanu", "Weicheng Kuo"], "abstract": "We present RECLIP (Resource-efficient CLIP), a simple method that minimizes computational resource footprint for CLIP (Contrastive Language Image Pretraining). Inspired by the notion of coarse-to-fine in computer vision, we leverage small images to learn from large-scale language supervision efficiently, and finetune the model with high-resolution data in the end. Since the complexity of the vision transformer heavily depends on input image size, our approach significantly reduces the training resource requirements both in theory and in practice. Using the same batch size and training epoch, RECLIP achieves highly competitive zero-shot classification and image-text retrieval accuracy with 6 to 8× less computational resources and 7 to 9× fewer FLOPs than the base- line. Compared to the state-of-the-art contrastive learning methods, RECLIP demonstrates 5 to 59× training resource savings while maintaining highly competitive zero-shot classification and retrieval performance. Finally, RECLIP matches the state of the art in transfer learning to open-vocabulary detection tasks, achieving 32 APr on LVIS. We hope this work will pave the path for the broader research community to explore language supervised pretraining in resource-friendly settings.", "keywords": "", "qa_pairs": [{"question": "How does the TPU padding feature impact the computational efficiency of RECLIP-64 compared to RECLIP-80, and what measures are taken to minimize overhead?", "answer": "The authors ensure that the feature dimension is a multiple of 128 (i.e., 4096) and the batch size is a multiple of 1024 (i.e., 16384) to avoid padding on the sequence dimension and minimize overhead. In practice, RECLIP-64 provides a ~40% speedup compared to RECLIP-80 during the first low-res training stage, indicating that the shorter sequence of RECLIP-64 is beneficial for computational efficiency.", "category": "Method_Mechanics"}, {"question": "Why does RECLIP-112 perform worse than RECLIP-80 in some downstream tasks after high-resolution finetuning, despite having a higher image resolution?", "answer": "The authors agree that higher-resolution images should generally lead to improved results. During low-res training, RECLIP-112 outperforms RECLIP-80, as expected. However, after high-res finetuning, all RECLIP variants achieve very similar retrieval/classification performance, and the ranking among the variants does not perfectly match the ordering of image size. This is partly attributed to the effectiveness of high-res finetuning in recovering the representation quality loss from low-res training. At this time, there is no concrete explanation for the mixed ordering after high-res finetuning, and it remains an empirical question whether different hyper-parameters, data, or infrastructure might yield the expected ordering. ", "category": "Experimental_Analysis"}, {"question": "Has RECLIP been evaluated on object detection tasks, specifically for open-vocabulary detection, and what were the results?", "answer": "Authors conduct evaluation on the LVIS dataset [1] by using RECLIP for open vocabulary detection. Authors take a recent SOTA approach RO-ViT [2] as the baseline and apply RECLIP-80 to pre-train the model (RECLIP-RO-ViT). Authors train only on the LVIS base categories (frequent & common) and test on both the base and novel (rare) categories following the protocol of ViLD [3]. The results are in the table below. RECLIP-RO-ViT achieves 32.0 Mask APr (AP on rare categories) [1] and matches state of the art RO-ViT (32.1). This is surprising and encouraging because detection task typically requires even higher resolution e.g. 1024 than classification task to recognize the small objects, which can be especially challenging for RECLIP due to the low-res information loss. In addition, RECLIP-RO-ViT outperforms RO-ViT by 0.7 on all-category AP, showing that its representation is also suitable for standard detection on the base categories. These detection results suggest that RECLIP representation is versatile for a broader range of object/pixel-level tasks. \n| Vit based method | Pretrained model | Detector backbone | APr | AP |\n| ---------------- | ---------------- | ---------------- | --- | ---|\n| RO-ViT [2] | ViT-L/16 | ViT-L/16 | 32.1 | 34.0 |\n| RECLIP-RO-ViT (Authors) | ViT-L/16 | ViT-L/16 | 32.0 | 34.7 |\n\n[1]. Gupta et al. LVIS: A Dataset for Large Vocabulary Instance Segmentation. CVPR 2019.\n[2]. Kim et al. Region-Aware Pretraining for Open-Vocabulary Object Detection with Vision Transformers. CVPR 2023.\n[3] Gu et al. Open-Vocabulary Open Detection via Vision and Language Knowledge Distillation. ICLR 2022.\n\n\n\n\n", "category": "Experimental_Exposition"}, {"question": "What is the rationale for RECLIP using learnable positional embeddings instead of fixed sine-cosine embeddings, and have the performance differences between these approaches been evaluated?", "answer": "RECLIP uses learnable positional embeddings in line with standard practices in vision transformers  [Dosovitskiy et al 2021] and CLIP [Radford et al 2021]. A comparison between trainable and fixed SinCos positional embeddings on image-text retrieval shows that the trainable version improves Recall@1 metrics by 0.6 points on average. ", "category": "Motivation_Analysis"}, {"question": "Can authors provide a brief description of the neural network architecture used in this paper, considering that some readers may not be familiar with the previous papers?", "answer": "The neural network architecture used in this paper employs the ViT-Large backbone as the image encoder. The ViT-Large is a vision transformer comprising 24 multi-head self-attention layers, each with 16 heads, and a width dimension of 1024. Detailed descriptions of both the vision transformers and the text tower architecture have been included in the *Network architecture* section of Section 3.2.", "category": "Experimental_Setup"}, {"question": "How many images are required for high-resolution image finetuning, and do these images have the same content as those used for low-resolution training?", "answer": "Approximately 800 million images are needed for high-resolution finetuning. The images used for high-resolution training are the same as those for low-resolution training, but without downsampling to retain rich visual details. The total number of images for both low-resolution training and high-resolution finetuning matches the baseline method.", "category": "Experimental_Setup"}, {"question": "What are the key differences in capacity and performance between TPU-v3 and GPU for model training, and how does TPU-v3 excel in neural network workloads?", "answer": "GPUs are general-purpose processors that support many applications and require access to register or shared memory for calculations. TPUs, however, are specialized for neural network workloads and handle massive matrix operations at fast speeds. TPU v3 achieves high-performance training across multiple workloads; for example, using 16 TPU v3 chips for training ResNet-50 is 19% faster than using 16 V100 GPUs in a DGX-2 machine. TPU v3 also has 16GB high-bandwidth memory per core, comparable to V100, making it suitable for synchronous large-scale training.", "category": "Method_Comparison"}, {"question": "How does the performance of the masking strategy compare to the resizing strategy across different computational budgets when controlling for all other factors?", "answer": "An apple-to-apple comparison was conducted between masking and resizing strategies with matching computational budgets. All factors other than masking vs resizing were controlled to be the same, including batch size, data, training recipe, and the number of iterations. The masking ratios were set to match the sequence lengths between resizing and masking (e.g., Mask-112 masks 75% tokens to match RECLIP-112：assuming the baseline using full image size 224x224). The results show a clear advantage for resizing over masking, with RECLIP-112 outperforming Mask-112 by +2.9%, and the gap increasing to +5.4% for RECLIP-64 compared to Mask-64. \n| X | Mask-X | RECLIP-X |\n| ------- | -------- | -------- |\n| 112 | 72.9 | 75.8 (+**2.9**)|  \n| 80 | 71.3 | 75.8 (+**3.5**) | \n| 64 | 69.5 | 74.9 (+**5.4**) | ", "category": "Method_Comparison"}, {"question": "Can authors provide more insightful discussions or theoretical analyses to enrich the content of the proposed method, which appears to be a simple engineering trick despite its effectiveness?", "answer": "The method is inspired by the idea of coarse-to-fine in computer vision literature. Authors leverage small images to efficiently learn large-scale image-language correspondence, and then finetune the model with high-resolution data to help it learn the important details. Authors believe the effectiveness and computation savings of RECLIP can be valuable to the community. The existing complexity analysis is based on the self-attention operations in transformers. However, empirically the complexity of a transformer may not be dominated by the self-attention layers, because the feed forward layers also play an important role. GPT-3 [1] paper have provided computation analysis of their language models, where the computation cost is assumed to scale as $O(N)$, linear with the sequence length $N$ due to the feed forward layers. Thus, here authors discussed the lower-bound of the complexity $C_{lb}$ of RECLIP following the same assumption of GPT-3 [1]. Authors have $$ C_{lb} = O(BN) = O\\left( \\frac{BHW}{pr^2} \\right)$$. During the training, the B, H, W, p are normally constant, so the above equation can be simplified as:$C = O\\left(\\frac{1}{r^2}\\right)$. Compared to Eq. 5, authors observe that the computation savings in practice may be somewhere between $O\\left(\\frac{1}{r^2}\\right)$ and $O\\left(\\frac{1}{r^4}\\right)$. This suggests that changing \\(r\\) is still very effective for compute savings. Authors have added the new analysis accordingly in *Training complexity* section of Section 3.2.\n[1]. Brown et al. Language Models are Few-Shot Learners. NeurIPS, 2020", "category": "Method_Disambiguation"}, {"question": "Are approximately 800 million images used specifically for high-resolution image finetuning?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "t0L4xG4aGC", "venue": "TMLR 2024", "paper_decision": "Reject", "title": "Encoding Hierarchical Information in Neural Networks \\\\helps in Subpopulation Shift", "authors": ["Runze Li", "Dahun Kim", "Bir Bhanu", "Weicheng Kuo"], "abstract": "Over the past decade, deep neural networks have proven to be adept in image classification tasks, often surpassing humans in terms of accuracy. However, standard neural networks often fail to understand the concept of hierarchical structures and dependencies among different classes for vision related tasks. Humans on the other hand, seem to intuitively learn categories conceptually, progressively growing from understanding high-level concepts down to granular levels of categories. One of the issues arising from the inability of neural networks to encode such dependencies within its learned structure is that of subpopulation shift -- where models are queried with novel unseen classes taken from a shifted population of the training set categories. Since the neural network treats each class as independent from all others, it struggles to categorize shifting populations that are dependent at higher levels of the hierarchy. In this work, we study the aforementioned problems through the lens of a novel conditional supervised training framework. We tackle subpopulation shift by a structured learning procedure that incorporates hierarchical information conditionally through labels. Furthermore, we introduce a notion of graphical distance to model the catastrophic effect of mispredictions. We show that learning in this structured hierarchical manner results in networks that are more robust against subpopulation shifts, with an improvement up to ~3 % in terms of accuracy and up to 11 % in terms of hierarchical distance over standard models on subpopulation shift benchmarks.", "keywords": "", "qa_pairs": [{"question": "How does the claimed contribution of conditional training by filtering (Section 1, first bullet point, pg. 3) differ from prior works on cascading models, filtering inputs, and anytime inference for image classification, particularly in terms of addressing subpopulation shift?", "answer": "The claimed contribution of conditional training by filtering differs from prior works on cascading models, filtering inputs, and anytime inference for image classification in its application to subpopulation shift. While prior works on anytime inference aim for more efficient inference pipelines, the method uses conditional training via filtering to propagate conditional probabilities that aid in making better decisions under subpopulation shift. The formulation itself is not novel, but its usage in this context is, and this distinction will be clarified in the contributions section of the updated draft.", "category": "Method_Comparison"}, {"question": "How does the introduction of misprediction impact in authors' work differ from existing research on hierarchical error and cost-sensitive classification, particularly in the context of Bertinetto et al. 2020 and the original ImageNet papers?", "answer": "The work aims to provide a holistic assessment showing that with hierarchies, the network makes fewer errors and the errors made are 'closer' to the ground truth, measured by graphical traversal. Authors will clarify that authors use known strategies to demonstrate this behavior. Unlike Bertinetto et al. 2020, authors employ misprediction impact only as an evaluation metric to assess the impact of hierarchies on subpopulation shift, not as an optimization target during training.", "category": "Method_Comparison"}, {"question": "What is the connection between source accuracy (without shift, i.e., the 's-s' condition) and subpopulation shift accuracy, and is it necessary to specifically train for subpopulation shift, or will generally better classifiers resolve the issue?", "answer": "The connection between source accuracy (S-S) and subpopulation shift accuracy (S-T) varies across datasets. While improvements in S-S and S-T accuracies track each other roughly in custom datasets and LIVING-17 datasets, this trend does not hold for non-LIVING-26 and ENTITY-30 datasets. The nature of the shifted subpopulation and the specific classes where the shift occurs influence this relationship. For example, shifting one source to three targets in LIVING17-A, B, and C yields different results. ", "category": "Motivation_Analysis"}, {"question": "How does the proposed method address the assumption that earlier layer representations in the ResNet18 architecture correspond to higher levels in the label hierarchy, given that earlier layers typically learn low-level features?", "answer": "The earliest head in the proposed method is placed after 3 out of 4 ResNet blocks in the ResNet18 architecture, specifically at the 14th layer out of 18. This position provides the network with sufficient representational capacity to learn high-level class-relevant features. The approach is inspired by early exit conditions in neural networks, where simpler examples are inferred upon in earlier layers[1]. In this framework, 'simpler' classes at higher levels of the hierarchy are defined by two factors: a) early levels correspond to visually distinct categories, and b) fewer classes to distinguish at early layers make the task simpler compared to later layers. The conditional training framework aims to enforce this hierarchy in both the learned features and the inference process to improve performance under subpopulation shift.\n [1] BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks.", "category": "Method_Mechanics"}, {"question": "How was the performance of the hierarchical network measured during validation, and what datasets were used for training and validation? Additionally, did the model have access to the target distribution during validation, and would using a different backbone, such as larger ResNets, still yield good results?", "answer": "The performance was measured using the validation set of the seen distribution for training and validation. The model did not have access to the target distribution during validation. The aim was to assess if hierarchical distributions help in subpopulation shift, so shifted distributions were used only as a test set. Final results are reported on the validation set of the unseen distribution, as per BREEDS. The same rationale applies to not optimizing the catastrophic coefficient as part of the loss function, using it only as an evaluation metric.", "category": "Experimental_Setup"}, {"question": "What steps have been taken to include the standard deviation in the experimental results, given that the differences between methods are small and require significance assessment?", "answer": "The experiments were re-run for five iterations, and the draft has been updated to include both the mean and standard deviation of all five runs. The main text now reflects the new means, while the standard deviations are documented in the appendix.", "category": "Experimental_Setup"}, {"question": "Why are the improvements in catastrophic coefficient and accuracy when comparing the proposed method to BCNN-18 and other baselines considered small, and how do these improvements relate to the overall goal of incorporating hierarchy?", "answer": "The improvements are considered small because the catastrophic coefficient changes are minimal, such as 2.7 vs 2.43. Authors have updated the presentation to show absolute changes for clarity. The goal is to use the catastrophic coefficient as an evaluation metric to observe the effect of incorporating hierarchy, not to optimize for it. Despite the slight improvements, the method demonstrates that even minimal changes and without exposing the evaluation metric to the training methodology, incorporating hierarchies still yields some benefits in addressing subpopulation shift issues.", "category": "Motivation_Analysis"}, {"question": "What evidence demonstrates that the additional effort required to collect multi-level hierarchical labels is justified, and how can the proposed algorithm be extended to scenarios where such labels are not readily available?", "answer": "In many cases, datasets are built using an underlying hierarchy, such as WordNet for ImageNet, where labels are available for free. For cases where label hierarchies are not known, independent works like [1] can be used to build hierarchies at an additional cost. This has been added to the relevant prior works section. Additionally, few works[2] on subpopulation shifts show large improvements without access to the shifted dataset for semi-supervised training. Authors have applied the idea behind the original [3] paper. The small improvement in the work highlights that hierarchies contain information beneficial for addressing subpopulation shifts without exposure to the shifted subpopulations.\n[1] Fast and Balanced: Efficient Label Tree Learning for Large Scale Object Recognition.\n\n[2] A Theory for Label Propagation for Subpopulation Shift\n\n[3] FixMatch : Simplifying Semi-Supervised Learning with Consistency and Confidence", "category": "Method_Mechanics"}, {"question": "Do the models have access to subpopulation labels, and how does Figure 2 relate to the availability of these labels?", "answer": "The models do not have access to subpopulation labels; these labels are only used for evaluation. Figure 2 illustrates three levels of population (superclass, class, and subclass) for easier understanding, but the subpopulation level below 'subclass' is not shown in the figure as the network does not access labels at that level. The final classification is 'Dog' vs 'Cat,' with subpopulation breeds shifting between seen and target cases.", "category": "Method_Disambiguation"}, {"question": "Are subpopulation labels used solely for evaluation purposes and not shown to the network during training?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the experiments initially run for three iterations and later extended to five iterations, with updated mean and standard deviation values reported?", "answer": "True", "category": "Claim_Verification"}, {"question": "Does Figure 2 accurately represent the network's access to labels at the subclass level?", "answer": "False", "category": "Claim_Verification"}, {"question": "Do the models have access to subpopulation labels during training, as suggested by the description on page 6?", "answer": "False", "category": "Claim_Verification"}]}
{"id": "6dYBP3BIwx", "venue": "NEURIPS 2024", "paper_decision": "Accept (poster)", "title": "CogVLM: Visual Expert for Pretrained Language Models", "authors": ["Weihan Wang", "Qingsong Lv", "Wenmeng Yu", "Wenyi Hong", "Ji Qi", "Yan Wang", "Junhui Ji", "Zhuoyi Yang", "Lei Zhao", "Song XiXuan", "Jiazheng Xu", "Keqin Chen", "Bin Xu", "Juanzi Li", "Yuxiao Dong", "Ming Ding", "Jie Tang"], "abstract": "We introduce CogVLM, a powerful open-source visual language foundation model. Different from the popular \\emph{shallow alignment} method which maps image features into the input space of language model, CogVLM bridges the gap between the frozen pretrained language model and image encoder by a trainable visual expert module in the attention and FFN layers. As a result, CogVLM enables a deep fusion of vision language features without sacrificing any performance on NLP tasks. CogVLM-17B achieves state-of-the-art performance on 17 classic cross-modal benchmarks, including 1) image captioning datasets: NoCaps, Flicker30k, 2) VQA datasets: OKVQA, TextVQA, OCRVQA, ScienceQA, 3) LVLM benchmarks: MM-Vet, MMBench, SEED-Bench, LLaVABench, POPE, MMMU, MathVista, 4) visual grounding datasets: RefCOCO, RefCOCO+, RefCOCOg, Visual7W. Codes and checkpoints are available at Github.", "keywords": "Multimodal Learning;Vision and Language;Representation Learning;Large Language Model", "qa_pairs": [{"question": "What alternative solutions were explored to reduce the increased GPU memory usage caused by the Visual Expert in CogVLM, and what were the findings regarding their effectiveness and trade-offs?", "answer": "Authors acknowledge that the Visual Expert introduces additional parameters, increasing static GPU memory usage during training and inference. This represents a trade-off between performance and computational cost. Notably, CogVLM outperforms models like LLaVA-Next, which use much larger language models, demonstrating the efficiency of the approach. Similar parameter-heavy approaches have been employed in NLP models, such as DeepSeek-v2[1] (236B total parameters, 21B activated). The authors explored the LoRA expert method with a LoRA rank of 128 as an alternative to reduce parameter count. However, they found that this method required 3.7 times more steps to reach the same loss level and had nearly identical per-step computation time, indicating limited expressiveness and reduced training efficiency. Despite the increased memory usage, CogVLM demonstrated efficiency by outperforming models like LLaVA-Next, and the additional memory overhead was acceptable for language models up to 32B parameters. \n[1] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model", "category": "Method_Mechanics"}, {"question": "How does involving pure language data in the SFT stage, as done in [1], impact the retention of LLM abilities in CogVLM, and can these abilities be retained through this method?\n[1] InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model", "answer": "The approach improves the model structure, while InternLM-XComposer2 focuses on data-side enhancements. These methods are not mutually exclusive. The latest experiments show that the best multimodal fusion is achieved by using Visual Expert modules, unfreezing LLM parameters, and simultaneously training on multimodal and pure text data.", "category": "Experimental_Exposition"}, {"question": "What is the impact of CogVLM's positional encoding scheme on addressing the 'remote attenuation effect' in RoPE, and what are the comparative experimental results?", "answer": "CogVLM's positional encoding scheme addresses the 'remote attenuation effect' in RoPE by preventing the query from overfocusing on nearby image tokens. Comparative experiments show that at 224x224 resolution, both methods achieve the same pre-training loss; at 490x490 resolution, the method achieves 5% lower loss; and on the DocVQA task (1120x1120 resolution), the method improves accuracy from 47.7% to 49.1%. Concurrent work Mousi[2] observed a similar phenomenon, validating the approach.\n[2] MouSi: Poly-Visual-Expert Vision-Language Models", "category": "Experimental_Exposition"}, {"question": "What is the definition of contamination used in this work, and how does it differ from traditional heuristic-based definitions?", "answer": "In this work, contamination is defined as the problem of detecting statistical dependence between the test data and model parameters, differing from traditional heuristic-based definitions like n-gram overlap. This approach allows for provable guarantees of contamination in cases of verbatim contamination, where the full test set is embedded in the pretraining data. To illustrate the relevance of this setting, authors note that a search of The Pile, a large open-source language modeling dataset, yielded numerous instances of small real-world datasets embedded with examples appearing in-order. As one example, the following is an excerpt from a dataset for an annotation tool made by Explosion, the creators of spaCy, a popular natural language processing framework, found in The Pile: ``` {\"text\":\"Uber’s Lesson: Silicon Valley’s Start-Up Machine Needs Fixing\",\"meta\":{\"source\":\"The New York Times\"}} {\"text\":\"Pearl Automation, Founded by Apple Veterans, Shuts Down\",\"meta\":{\"source\":\"The New York Times\"}} {\"text\":\"How Silicon Valley Pushed Coding Into American Classrooms\",\"meta\":{\"source\":\"The New York Times\"}} Source: https://github.com/explosion/prodigy-recipes/tree/fc06f6a6d93bc477e98cf0d8357c39322e4f5a6a ``` What the work shows is that by exploiting exchangeability in this setting, authors are able to provide guarantees on the false positive rate of the test.", "category": "Concept_Understanding"}, {"question": "What are the results of the CogVLM model on the MME benchmark, and how do they compare to other state-of-the-art models?", "answer": "The CogVLM model was evaluated on the MME benchmark using LLaMA-3 and GLM3 as base models. The results are as follows: CogVLM-LLaMA3-32B scored 2043.54 on MME, with detailed scores across categories like OCR (132.50), artwork (150.75), and perception (1548.18). CogVLM-GLM3-32B scored 2211.76 on MME, with scores such as OCR (155.00), artwork (149.75), and perception (1655.33). \n| Model | MME | OCR | artwork | celebrity | color | count | existence | landmark | position | posters | scene | perception | \n|---|-----|-----|---------|-----------|-------|-------|-----------|----------|----------|---------|-------|------------|\n|CogVLM-LLaMA3-32B| 2043.54 | 132.50 | 150.75 | 172.06 | 168.33 | 153.33 | 195.00 | 174.50 | 103.33 | 147.62 | 150.75 | 1548.18 | \n|CogVLM-GLM3-32B| 2211.76 | 155.00 | 149.75 | 155.29 | 165.00 | 175.00 | 200.00 | 182.75 | 131.67 | 185.37 | 155.50 | 1655.33 | \n\nModel | code_reasoning | commonsense_reasoning | numerical_calculation | text_translation | reasoning |\n|--       |--------|-----|------|-----|----|\n|CogVLM-LLaMA3-32B |70.00 | 132.86 | 107.50 | 185.00 | 495.36 |\n|CogVLM-GLM3-32B| 120.00 | 151.43 | 92.50 | 192.50 | 556.43 |", "category": "Experimental_Exposition"}, {"question": "How does the method's focus on improving the model structure compare to approaches like VILA and InternLM-XComposer2 that focus on data improvements, and can text data be effectively combined with multi-modal data?", "answer": "The method primarily focuses on improving the model structure, while approaches like VILA and InternLM-XComposer2 make improvements on the data side. These methods are not mutually exclusive. The latest experiments show that the best multimodal fusion is achieved by using Visual Expert modules, unfreezing LLM parameters, and simultaneously training on multimodal and pure text data. ", "category": "Method_Comparison"}, {"question": "Why does freezing the parameters of the large language model (LLM) while training the visual encoder and connector underperform compared to fine-tuning the LLM parameters in multimodal tasks?", "answer": "Freezing the parameters of the large language model (LLM) while training the visual encoder and connector often underperforms on many benchmarks, as seen in Table 2 of the paper. Additionally, freezing LLM parameters can lead to issues with instruction following and multimodal fusion. Fine-tuning the language model parameters is crucial for optimal performance in multimodal tasks, which is why the approach allows for updating these parameters.", "category": "Experimental_Analysis"}, {"question": "What is the trade-off between parameter efficiency and computational cost in CogVLM, and how does it compare to other models like LLaVA-Next and DeepSeek-v2?", "answer": "CogVLM introduces additional parameters through the Visual Expert, increasing static GPU memory usage, which represents a trade-off between performance and computational cost. Despite this, CogVLM outperforms models like LLaVA-Next that use larger language models, demonstrating its efficiency. Similar parameter-heavy approaches, such as DeepSeek-v2 (236B total parameters, 21B activated), have been successful in NLP. While the LoRA expert method reduces parameter count, it requires 3.7 times more steps to reach the same loss level and has nearly identical per-step computation time compared to CogVLM's current method. In experiments with language models up to 32B parameters, the additional memory overhead remained acceptable. Authors plan to continue exploring parameter reduction techniques for larger models.", "category": "Method_Comparison"}, {"question": "Why did the authors choose to use a new set of QKV and FFN instead of implementing LoRA (Low-Rank Adaptation) for image tokens, given the potential parameter reduction?", "answer": "The Visual Expert introduces additional parameters, which increases static GPU memory usage but enables performance gains, allowing CogVLM to outperform models like LLaVA-Next. Similar approaches have been employed in NLP models, such as DeepSeek-v2 [1], which has 236B total parameters but only 21B activated parameters. While exploring LoRA with a rank of 128, it was found to be less expressive, requiring 3.7 times more steps to reach the same loss level and having similar per-step computation time as the current method. This indicates that LoRA's parameter reduction comes at the cost of training efficiency. For language models up to 32B parameters, the additional memory overhead was acceptable.\n[1] DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model\n", "category": "Motivation_Analysis"}, {"question": "How does the proposed positional encoding scheme compare to the original RoPE strategy in terms of addressing the 'remote attenuation effect' and its impact on pre-training loss and accuracy at different resolutions?", "answer": "The proposed positional encoding scheme addresses the 'remote attenuation effect' in RoPE, preventing overfocus on nearby image tokens. Comparative experiments show that at 224x224 resolution, both methods achieve the same pre-training loss, but at 490x490 resolution, the proposed method achieves 5% lower loss. On the DocVQA task (1120x1120 resolution), the proposed method improves accuracy from 47.7% to 49.1%, with similar observations in concurrent work Mousi [2], validating the approach.\n[2] MouSi: Poly-Visual-Expert Vision-Language Models", "category": "Method_Comparison"}, {"question": "What is the rationale for training a separate grounding model instead of integrating grounding capabilities into the chat model, and how does this approach impact performance on specific tasks?", "answer": "The main reason for using two separate models is to handle ambiguous prompts that require coordinate outputs, such as 'Where is the dog in the image?'. This approach allows for clearer instructions for grounding tasks without affecting the chat model's performance. Experimentally, training a unified model resulted in a 1.5 point decrease in the Nocaps dataset score (from 128.3 to 126.8) and a 0.8 point increase in the VQAv2 task score (from 82.3 to 83.1).", "category": "Motivation_Analysis"}, {"question": "Did the authors evaluate CogVLM using LLaMA-3 as the base model on the MME benchmark and provide specific performance metrics?", "answer": "True", "category": "Claim_Verification"}, {"question": "Are the results of the MME benchmark included in the original version of the paper for CogVLM models?", "answer": "False", "category": "Claim_Verification"}, {"question": "Does the paper discuss the limitations of the study, particularly regarding the definition and detection of contamination?", "answer": "True", "category": "Claim_Verification"}, {"question": "Were the parameters of the visual encoder (ViT) tuned during the training process?", "answer": "True", "category": "Claim_Verification"}]}
{"id": "MPJ3oXtTZl", "venue": "NEURIPS 2024", "paper_decision": "Accept (poster)", "title": "G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering", "authors": ["Xiaoxin He", "Yijun Tian", "Yifei Sun", "Nitesh V Chawla", "Thomas Laurent", "Yann LeCun", "Xavier Bresson", "Bryan Hooi"], "abstract": "Given a graph with textual attributes, we enable users to `chat with their graph': that is, to ask questions about the graph using a conversational interface. In response to a user's questions, our method provides textual replies and highlights the relevant parts of the graph. While existing works integrate large language models (LLMs) and graph neural networks (GNNs) in various ways, they mostly focus on either conventional graph tasks (such as node, edge, and graph classification), or on answering simple graph queries on small or synthetic graphs. In contrast, we develop a flexible question-answering framework targeting real-world textual graphs, applicable to multiple applications including scene graph understanding, common sense reasoning, and knowledge graph reasoning. Toward this goal, we first develop a Graph Question Answering (GraphQA) benchmark with data collected from different tasks. Then, we propose our \\textit{G-Retriever} method, introducing the first retrieval-augmented generation (RAG) approach for general textual graphs, which can be fine-tuned to enhance graph understanding via soft prompting. To resist hallucination and to allow for textual graphs that greatly exceed the LLM's context window size, \\textit{G-Retriever} performs RAG over a graph by formulating this task as a Prize-Collecting Steiner Tree optimization problem. Empirical evaluations show that our method outperforms baselines on textual graph tasks from multiple domains, scales well with larger graph sizes, and mitigates hallucination.~\\footnote{Our codes and datasets are available at: \\url{https://github.com/XiaoxinHe/G-Retriever}}", "keywords": "Retrieval Augmented Generation;Graph Question Answering;Graph Neural Network;Large Language Model;Textual Graphs", "qa_pairs": [{"question": "How is the quality of the retrieved subgraphs evaluated, and how does the proposed method compare to other retrieval methods?", "answer": "The quality of the retrieved subgraphs is evaluated by checking if the label is contained within the retrieved subgraph, which is considered a successful retrieval. The retrieval success rate is calculated for the proposed method and the KAPING method on the WebQSP dataset. The proposed PCST-based subgraph retrieval achieves a Hit@1 accuracy of 67.87%, outperforming KAPING's top-k triple retrieval, which achieves 60.81%. Additionally, the PCST-based retrieval outperforms other baselines such as top-k nodes plus their neighbors and shortest path retrieval, achieving an accuracy of 66.17% on the WebQSP dataset.", "category": "Method_Comparison"}, {"question": "Why was the G-Retriever framework applied to the ExplaGraphs dataset, given its similar performance to GraphToken on small graphs?", "answer": "The application of G-Retriever to the ExplaGraphs dataset aimed to demonstrate the framework's flexibility and effectiveness across various graph sizes. While G-Retriever and GraphToken show similar performance on small graphs like ExplaGraphs, G-Retriever significantly outperforms GraphToken on other small-scale datasets like SceneGraphs, achieving 0.8131 accuracy compared to 0.4903. Additionally, G-Retriever offers extra benefits such as a flexible conversational interface, which adds value beyond accuracy.\n|     | # node: min, max, mean ± std | # edge: min, max, mean ± std|\n| -------- | ------- | ------- |\n| ExplaGraphs  | 4, 9, 5.17 ± 1.18    | 3, 8, 4.25 ± 1.23 |\n| SceneGraphs | 1, 126, 19.13 ± 8.28     | 0, 1657, 68.44 ± 65.77 |\n| WebQSP    | 3, 2000, 1371.18 ± 566.91    | 2, 10818, 4253.27 ± 2238.12 |", "category": "Motivation_Analysis"}, {"question": "What additional baseline methods were included to compare the effectiveness of PCST, and what were the results of these comparisons on the WebQSP dataset?", "answer": "To compare the effectiveness of PCST, authors included the following baseline methods: KAPING [1], which retrieves top-k triples and adds them to the input question; top-k nodes plus their one-hop neighbors; and shortest path retrieval. For all methods, k was set to 5, and llama2-7b-chat was used as the LLM. The results showed that PCST-based retrieval achieved an accuracy of 66.17, outperforming the baseline methods which had accuracies of 52.64 for KAPING, 49.82 for top-k nodes plus neighbors, and 55.20 for shortest path retrieval.\n\n| Method                               | Hit@1 |\n|--------------------------------------|--------------------|\n| PCST retrieval                        | 66.17              |\n| top-k triples retrieval (KAPING)      | 52.64              |\n| top-k nodes plus its neighbors        | 49.82              |\n| shortest path retrieval               | 55.20              |\n[1] Knowledge-Augmented Language Model Prompting for Zero-Shot Knowledge Graph Question Answering, 2023.", "category": "Method_Comparison"}, {"question": "What were the results of replacing the tuned generation model with fixed LLMs like GPT-4 and llama2-7b-chat on the WebQSP dataset?", "answer": "Additional experiments were conducted by replacing the tuned generation model with fixed LLMs on the WebQSP dataset. Specifically, the PCST retrieval was applied to obtain the subgraph, which was then converted into text and fed into the fixed LLMs along with the query. The open-source llama2-7b-chat achieved a Hit@1 score of 66.17, while the closed-source GPT-4 achieved a Hit@1 score of 67.87. G-Retriever outperformed both with a Hit@1 score of 70.49, demonstrating the effectiveness of the tuned generation model.\n|  Method                   | Hit@1 |\n|---------------------|--------------------|\n| llama2-7b-chat      | 66.17              |\n| GPT-4o              | 67.87              |\n| G-Retriever | 70.49              |\n", "category": "Experimental_Exposition"}, {"question": "How is the context of nodes and edges, such as adjacent nodes, incorporated into the encoding process?", "answer": "The context is not directly included in the initial encoding step, which uses a pre-trained language model to encode text attributes into embeddings. However, the context is incorporated later in the process through a graph neural network (GNN) component in the generation step. The GNN aggregates information from adjacent nodes and updates node and edge representations based on the graph context, ensuring that contextual information is considered in the overall framework.", "category": "Experimental_Setup"}, {"question": "What does the term 'chat with their graph' signify, and how does the G-Retriever model facilitate this interaction?", "answer": "The term 'chat with their graph' refers to users interacting with a graph via a conversational interface. Users can input a textual graph and ask natural language queries, which G-Retriever responds to in natural language. For instance, if a user inputs a mindmap-like explanation graph and requests an argument essay, G-Retriever can generate the essay. This enhances human-AI interaction, making the model more intuitive and aligned with human expectations.", "category": "Concept_Understanding"}, {"question": "What are the advantages of using the PCST algorithm to retrieve subgraphs?", "answer": "The PCST algorithm offers two main advantages for subgraph retrieval: 1. Context-Relevance: Selecting nodes and edges in isolation may overlook neighborhood information. In contrast, PCST-based retrieval is guaranteed to return a connected subgraph, capturing the graph context during the retrieval process. This approach retrieves not only high-relevance nodes or edges but also 'bridge' elements that connect these with contextually significant nodes or edges, which are crucial for generating a comprehensive response. 2. Size Management: Compared to the shortest path method, PCST retrieval provides greater control over the size of the retrieved subgraph. By adjusting the prizes and costs on nodes and edges, users can fine-tune the subgraph's extent. In contrast, the shortest path approach lacks the ability to control the distance between the top-k nodes, which can lead to disconnected subgraphs or the inclusion of unnecessarily long paths.", "category": "Method_Disambiguation"}, {"question": "How is the quality and effectiveness of the retrieved subgraphs quantified and compared to other methods?", "answer": "The quality of the retrieval method is quantified by examining whether the label is contained within the retrieved subgraph, which is considered a successful retrieval. The retrieval success rate of the method is calculated and compared to KAPING on the WebQSP dataset, showing Hit@1 accuracies of 60.81% for KAPING and 67.87% for the proposed PCST-based subgraph retrieval. Additionally, PCST-based retrieval outperforms other baselines, achieving an accuracy of 66.17% on the WebQSP dataset, demonstrating its effectiveness.", "category": "Method_Mechanics"}, {"question": "What are the inference times and accuracy of G-Retriever and its variant without retrieval during the inference stage, and what factors contribute to the speedup observed in G-Retriever?", "answer": "The inference times for G-Retriever and its variant without retrieval are 9.55 minutes and 21.01 minutes, respectively, with accuracies of 70.49% and 63.84%. The speedup in G-Retriever is due to the efficient PCST subgraph construction, which takes only 0.29 seconds for the largest subgraph, and the significant reduction in graph size by PCST retrieval, eliminating up to 99% of nodes.\n| Method | Time in minutes | Accuracy with Hit@1 |\n |------------------------------|----------|--------|\n | G-Retriever | 9.55 | 70.49 |\n | G-Retriever (w/o retrieval) | 21.01 | 63.84 | ", "category": "Experimental_Exposition"}, {"question": "What are the key experimental details and evaluation metrics used in the hallucination evaluation in section 6.4, and how will the authors address concerns about transparency and bias?", "answer": "In section 6.4, the hallucination evaluation involves manual examination due to the flexible and varied output formatting of LLM responses, which makes automatic evaluation difficult. The determination of hallucinations is based on objective criteria, such as calculating fractions of valid nodes and edges. For example, in Table 1, in the response from 'LLM w/ Graph Prompt Tuning,' it is stated: 'The animal in the bushes is a deer. Nodes: * Deer (node 1), * Bushes (node 2); Edges: * Deer → Bushes (edge 1) * Deer → Grass (edge 2) * Bushes → Grass (edge 3).' The fraction of 'valid nodes' is 1/2 = 0.5, as 'Bushes' exists in the scene graph while 'Deer' does not. The fraction of 'valid edges' is 0/3 = 0, since none of the mentioned edges exist in the scene graph. As there are hallucinated nodes and edges, this is not a 'fully valid graph.'  To ensure transparency and mitigate bias, the authors will open-source the model outputs used for manual checks, allowing for public scrutiny and verification.", "category": "Experimental_Setup"}, {"question": "What is the role of text attributes in the G-Retriever system and how does their absence impact the performance in different stages as shown in the ablation study?", "answer": "The text attributes in G-Retriever are crucial in three areas: (1) Retrieval, where they help select top-k nodes and edges based on semantic similarity to the query embedding; (2) GNN Input, where they are fed into the GNN to refine embeddings and capture complex relationships; and (3) LLM Input, where the retrieved subgraph is textualized for the LLM to generate answers. The ablation study shows that without node attributes, performance drops to 66.58 in retrieval, 68.85 in GNN input, and 56.32 in LLM input. Without edge attributes, performance drops to 58.37 in retrieval, 67.87 in GNN input, and 68.24 in LLM input.\n| Features | Without Node | Without Edge | \n|----------|--------------|--------------| \n| Retrieval | 66.58 | 58.37 | \n| GNN Input | 68.85 | 67.87 | \n| LLM Input | 56.32 | 68.24 |", "category": "Experimental_Exposition"}, {"question": "What is the motivation for selecting Prize-Collecting Steiner Tree (PCST) for subgraph construction in Section 5.3, and how does 'optimally sized and relevant subgraphs' align with the objectives of PCST? Additionally, how does PCST-based retrieval compare to baseline methods?", "answer": "Subgraph retrieval is formulated as a Prize-Collecting Steiner Tree (PCST) optimization problem because it aligns well with the objective of finding a connected subgraph containing the most relevant nodes and edges. The goals of PCST, which include maximizing node values while minimizing edge costs, match this objective. While PCST is not universally acknowledged as optimal, empirical validation has demonstrated its effectiveness. Comparisons with baseline methods such as top-k triples retrieval (KAPING), top-k nodes plus their neighbors, and shortest path retrieval show that PCST-based retrieval outperforms these methods, achieving a Hit@1 accuracy of 66.17 on the WebQSP dataset. Additionally, PCST retrieval provides greater control over the size of the retrieved subgraph and ensures contextual relevance by returning connected subgraphs.\n| Method | Hit@1 |\n |--------------------------------------|--------------------|\n | PCST retrieval | 66.17 | \n| top-k triples retrieval (KAPING) | 52.64 |\n | top-k nodes plus its neighbors | 49.82 |\n | shortest path retrieval | 55.20 | ", "category": "Motivation_Analysis"}, {"question": "Can authors provide a detailed explanation of the evaluation methodology used to assess hallucinations, including the experiment design, evaluation metrics, baseline, and results?", "answer": "The evaluation of hallucinations involves the following: - **Experiment Design:** The LLM was instructed to answer graph-related questions and list the nodes or edges in the explanation graph that support its answers. Since there are no standard answers, 100 responses generated by both the method (G-Retriever) and a baseline method (LLM with graph prompt tuning) were manually examined to verify the existence of referenced nodes and edges in the graph. - **Evaluation Metrics:** Faithfulness was assessed using three metrics: the fraction of valid nodes (Valid Nodes), the fraction of valid edges (Valid Edges), and the fraction of times the entire set of nodes and edges cited was valid (Fully Valid Graphs). - **Baseline:** MiniGPT-4 was adapted to graph contexts as the baseline, using a frozen LLM with a trainable GNN that encodes graph data as a soft prompt (LLM+Graph Prompt Tuning). - **Results:** G-Retriever outperformed the baseline, achieving 77% validity in nodes, 76% in edges, and 62% overall validity in node-edge sets, compared to the baseline's 31% valid nodes, 12% valid edges, and 8% fully valid node-edge sets. These results demonstrate G-Retriever's effectiveness in reducing hallucinations. The detailed evaluation is provided in Appendix F.", "category": "Method_Disambiguation"}]}
{"id": "nRp0XhTf61", "venue": "NEURIPS 2024", "paper_decision": "Accept (poster)", "title": "InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD", "authors": ["Xiaoyi Dong", "Pan Zhang", "Yuhang Zang", "Yuhang Cao", "Bin Wang", "Linke Ouyang", "Songyang Zhang", "Haodong Duan", "Wenwei Zhang", "Yining Li", "Hang Yan", "Yang Gao", "Zhe Chen", "xinyue zhang", "Wei Li", "Li Jingwen", "Wenhai Wang", "Kai Chen", "Conghui He", "Xingcheng Zhang", "Jifeng Dai", "Yu Qiao", "Dahua Lin", "Jiaqi Wang"], "abstract": "The Large Vision-Language Model (LVLM) field has seen significant advancements, yet its progression has been hindered by challenges in comprehending fine-grained visual content due to limited resolution. Recent efforts have aimed to enhance the high-resolution understanding capabilities of LVLMs, yet they remain capped at approximately 1500 $\\times$ 1500 pixels and constrained to a relatively narrow resolution range. This paper represents InternLM-XComposer2-4KHD, a groundbreaking exploration into elevating LVLM resolution capabilities up to 4K HD (3840 × 1600) and beyond. Concurrently, considering the ultra-high resolution may not be necessary in all scenarios, it supports a wide range of diverse resolutions from 336 pixels to 4K standard, significantly broadening its scope of applicability. Specifically, this research advances the patch division paradigm by introducing a novel extension: dynamic resolution with automatic patch configuration. It maintains the training image aspect ratios while automatically varying patch counts and configuring layouts based on a pre-trained Vision Transformer (ViT) (336 $\\times$ 336), leading to dynamic training resolution from 336 pixels to 4K standard. Our research demonstrates that scaling training resolution up to 4K HD leads to consistent performance enhancements without hitting the ceiling of potential improvements. InternLM-XComposer2-4KHD shows superb capability that matches or even surpasses GPT-4V and Gemini Pro in 10 of the 16 benchmarks.", "keywords": "Large Vision Language Model (LVLM)", "qa_pairs": [{"question": "Why is it reasonable to use a large number of visual tokens for a single image when increasing the resolution, and what evidence supports this approach?", "answer": "Using a large number of visual tokens is justified for high-resolution images because the LVLM model effectively leverages these tokens to capture fine details, as evidenced by significant performance gains on high-resolution benchmarks such as DocVQA and InfographicVQA. Additionally, the study aims to develop a universal solution for LVLMs to handle varying resolutions (336 pixels to 4K), and further token compression could lead to unavoidable information loss. Figure 4 in the paper demonstrates that more visual tokens consistently improve performance on tasks requiring fine-detailed understanding.", "category": "Method_Disambiguation"}, {"question": "Would increasing the size of the large language model (LLM) parameters lead to better performance in the study, and what evidence supports this claim?", "answer": "Yes, increasing the size of the LLM parameters leads to better performance. The authors conducted experiments on a smaller model, InternLM2-1.8B, and observed a clear performance improvement as the LLM parameters increased. They believe that further increasing the LLM size would result in even better performance.\n| Model    | LLM Size | DocVQA  | InfoVQA | TextVQA | ChartQA | MMBench | MME     | SEED  |\n|----------|------|---------|---------|---------|---------|---------|---------|-------|\n| IXC-4KHD | 1.8B | 81.7    | 54.9    | 70.1    | 72.4    | 70.6    | 1,973.0 | 70.5  |\n| IXC-4KHD | 7B   | 90.0    | 68.6    | 77.2    | 81.0    | 80.2    | 2,204.9 | 74.7  |", "category": "Experimental_Exposition"}, {"question": "How does the proposed model generalize to multimodal video evaluation datasets like Video-MME, and what are the performance results on this and other video benchmarks?", "answer": "The proposed model was tested by concatenating video frames into a long image and finetuning with SFT data from Video-ChatGPT [1]. On the MME-Video benchmark, it performs comparably to GPT-4V, achieving a score of 54.2 without subtitles. On the MVBench and MLVU benchmarks, it significantly outperforms GPT-4V, with scores of 56.7 and 65.1 respectively.\n[1] Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models\n\n| Model    | MME-Video w/o subtitle | MVBench | MLVU  |\n|----------|------------------------|---------|-------|\n| GPT-4V   | 56.0                   | 43.5    | 49.2  |\n| IXC-4KHD | 54.2                   | 56.7    | 65.1  |", "category": "Experimental_Exposition"}, {"question": "What are the specific aspects of novelty in the proposed method compared to existing approaches like Monkey, LLaVA-NeXT, and InternVL-1.5?", "answer": "The proposed method introduces several novel aspects compared to existing approaches: 1. It is the first to address challenges in handling variability in image feature patch layouts, enabling effective training with dynamic high-resolution images. 2. It includes a dynamic resolution and automatic path configuration strategy, allowing seamless scaling from 336 pixels to 4K resolution, which exceeds the capabilities of previous methods. In contrast, existing methods for high-resolution understanding are limited to either predefined high-resolution settings or a narrow range of resolutions. For instance, Monkey and LLaVA-Next employ sliding windows and image splitting, respectively, but are restricted to segmenting images into a maximum of 6 and 4 patches, respectively. This limitation hinders their ability to handle images with extreme aspect ratios, commonly found in infographics. The IXC2-4KHD, on the other hand, supports up to 55 local image patches, making it capable of tackling these challenging cases. 3. It provides a thorough analysis of high-resolution strategy design, examining previously underexplored aspects such as the impact of global image, 'newline' token, and training/inference resolution. Through extensive experimentation, authors have gleaned valuable insights, including: - The LVLM struggles to comprehend the 2D structure of a simply flattened image, but incorporating a global image and 'newline' tokens could solve this problem with clear performance improvement. - For tasks requiring high-resolution images, the model demonstrates improved performance with larger images, showing no signs of saturation even at 4KHD resolution. This suggests potential for further scaling and improved results. - Conversely, for tasks insensitive to resolution, model performance plateaus as image size increases. Additionally, InternVL-1.5 is a concurrent, unpublished work, and LLaVA-Next is a GitHub blog post, not a rigorously evaluated academic publication.", "category": "Method_Comparison"}, {"question": "How does the proposed high-resolution solution compare to other solutions when trained with identical architecture, training strategy, and dataset?", "answer": "The proposed high-resolution solution, IXC-LLaVA, was trained with identical architecture, training strategy, and dataset but adopted the high-resolution strategy from LLaVA-Next. It was compared with LLaVA-Next 0530 official results and IXC2-4KHD under HD-9/16 setting. IXC-LLaVA achieved promising performance across all six benchmarks, leveraging additional training data and advanced IXC architecture design. However, it was outperformed by IXC2-4KHD HD-9, which utilized fewer image tokens yet yielded better results. This comparison highlights the efficiency and effectiveness of the proposed high-resolution strategy over the LLaVA-Next approach.\n| Model Name          | HD Strategy | Image Tokens | Max Resolution | DocVQA | TextVQA | ChartQA | MMBench |\n| ------------------- | ----------- | ------------ | -------------- | ------ | ------- | ------- | ------- |\n| LLaVA-Next (LLaMA3) | LLaVA-Next  | 2880         | 672x672        | 78.2   | 52.0    | 69.5    | 72.1    |\n| IXC-LLaVA           | LLaVA-Next  | 2880         | 672x672        | 78.9   | 71.5    | 74.1    | 77.6    |\n| IXC-4KHD            | HD9         | 1440         | 1008x1008      | 79.4   | 73.8    | 78.2    | 79.5    |\n| IXC-4KHD            | HD16        | 2448         | 1344x1344      | 84.9   | 75.7    | 80.1    | 80.2    |\n", "category": "Method_Comparison"}, {"question": "What amount of data is used for training in the proposed method, and how does this compare to the data usage of the previous SoTA LLaVA-NeXT?", "answer": "The proposed method uses almost 3 million data points for pre-training and 1.2 million data points for the SFT stage. In comparison, the previous SoTA LLaVA-NeXT uses less than 2 million data points.\n\n| General Semantic Alignment    | ShareGPT4V-PT (1.2M), COCO (50K), Nocaps (20K), TextCaps (50K), SBU (200K), LAION400M (1.2M), CC 3M (0.3M) |\n| ----------------------------- | ---------------------------------------------------------------------------------------------------------- |\n| World Knowledge Alignment     | Concept Data (500K)                                                                                        |\n| Vision Capability Enhancement | WanJuan (400K), Flicker (30K), MMC-Inst (100K), RCTW-17 (8K), CTW (2K), LSVT (40K), ReCTs (20K), ArT (6K)  |\n\n\n\n| Caption            | ShareGPT4V (100K), COCO (50K), Nocaps (5K)                                        |\n| ------------------ | --------------------------------------------------------------------------------- |\n| General QA         | VQAv2 (100K), GQA (100K), OK-VQA (10K), VD (100K), RD (2K), VSR (5K)              |\n| Science QA         | AI2D (15K), SQA (10K), TQA (6K), IconQA (10K)                                     |\n| Chart QA           | DVQA (30K), ChartQA (7K), ChartQA-AUG (21K)                                       |\n| Math QA            | MathQA (25K), Geometry3K (2K), TabMWP (20K), CLEVR-MATH (10K), Super (10K)        |\n| World Knowledge QA | A-OKVQA (17K), KVQA (5K), ViQuAE (3K)                                             |\n| OCR QA             | TextVQA (20K), OCR-VQA (10K), ST-VQA (26K)                                        |\n| HD-OCR QA          | InfoVQA (24K), DocVQA (40K)                                                       |\n| Conversation       | LLaVA-150k (100K), LVIS-Instruct4V (100K), ShareGPT-en\\&zh + InternLM-Chat (100K) |\n", "category": "Method_Comparison"}, {"question": "What strategies are employed in the paper to address the computational challenges associated with training high-resolution images, and how do these strategies impact the training efficiency?", "answer": "1. Enabling LVLM to understand high-resolution images remains a challenging and open problem in the field. The primary goal of the paper is to explore a general and effective strategy for high-resolution image understanding LVLMs, and further study how the image resolution influences LVLM performance. 2. Efficiency is an important under-explored topic, but in most cases, there is a trade-off between performance and efficiency. In this work, authors focus on pushing the performance upper bound without introducing additional variables, and authors would leave the study on efficiency in future work. 3. The proposed weighted multi-dataloader sampling strategy (Please see Appendix B.1) is designed to optimize the training process. By enabling each GPU to process batches with either high-resolution images only or low-resolution images only, authors reduce the need for padding tokens to handle unbalanced sample lengths caused by varying image sizes. This approach speeds up the training process, making it more efficient and scalable.", "category": "Method_Mechanics"}, {"question": "What is the inference efficiency of the model when processing high-resolution inputs, and how does the computational cost scale with the input resolution?", "answer": "The inference efficiency of the model is analyzed in two stages: prefix encoding and token decoding. The prefix encoding time increases linearly with the number of prefix tokens, while the per-token decoding speed remains constant due to optimizations like kv-cache and flash-attention. When generating 2048 tokens, the total inference time is nearly identical across different HD settings (HD9 to HD55), as the encoding time is relatively small compared to decoding time. The results suggest that the inference efficiency is acceptable, with prefix encoding times ranging from 0.2845 to 1.1272 seconds for HD9 to HD55, respectively, and per-token decoding speeds consistently around 0.0981 to 0.0983 seconds. \n| HD   | Image Tokens | Prefix Encoding Time | Per-Token Decoding Speed | Time to Generate 2048 Tokens |\n| ---- | ------------ | -------------------- | ------------------------ | ---------------------------- |\n| HD9  | 1440         | 0.2845               | 0.0982                   | 201.4                        |\n| HD16 | 2448         | 0.3966               | 0.0983                   | 201.7                        |\n| HD25 | 3744         | 0.5513               | 0.0981                   | 201.5                        |\n| HD55 | 8064         | 1.1272               | 0.0983                   | 202.4                        |\n", "category": "Experimental_Exposition"}, {"question": "What specific examples illustrate the variability in patch layouts discussed in Line 52, and how does this variability impact the model's understanding of 2D structures?", "answer": "The variability in patch layouts is illustrated with Fig.2 in the PDF, showing that distinct images (a and b) become identical one-dimensional inputs (Fig. 2(c)) for the LLM. This variability, along with diverse aspect ratios, can confuse object positioning. To address this, the LVLM uses Global-Local Format and Patch Layout Indicator to accurately comprehend the 2D structure, with the global image providing an overview and the indicator conveying the precise 2D structure.", "category": "Experimental_Analysis"}, {"question": "What does the phrase 'even when' in Line 56 imply about the model's performance at resolutions beyond 4K, and how is this supported by experimental results?", "answer": "The phrase 'even when' in Line 56 implies that the model's performance is not saturated at higher resolutions like HD55, as shown in Fig. 4 and Table 6. This is supported by quantitative results on a subset of InfoVQA with resolutions exceeding 1680x3696 pixels, where the model trained with 4KHD settings still benefits from larger inference resolutions, demonstrating performance improvements from 70.41 to 71.64 under HD55 and HD65 settings.\n\n| InfoVQA HD55 subset | 55    | 65       |\n| ------------------- | ----- | -------- |\n| HD55                | 70.41 | 71.64    |\n", "category": "Experimental_Exposition"}, {"question": "Do traditional methods trained at 4K resolution fail to gain additional performance at lower resolutions, and how does this observation support the study's goal?", "answer": "Traditional methods trained at 4K resolution do not gain additional performance at lower resolutions, as shown by nearly identical performance of MME and MMBench between HD9 and HD25 in Fig. 4. This indicates performance saturation when the training resolution exceeds the image's original resolution. However, performance improves with increasing training resolution when it is smaller than the image's original resolution, as seen in DocVQA and InfoVQA. These observations support the study's goal of developing a strategy for LVLM to understand images across various resolutions.", "category": "Method_Disambiguation"}, {"question": "How does the model perform on complex visual content such as hand-written mathematical formulas, circuit diagrams, or chemical structures, and what evidence supports its capability in handling these tasks?", "answer": "The model's performance on complex visual content was evaluated using a subset of high-resolution images (with resolutions exceeding 1000x1000 pixels) from the MMMU validation set, which accounts for approximately 10.5% of the dataset. IXC2-4KHD showed a notable improvement over IXC2, with scores of 52.63 and 46.32 respectively, indicating its ability to handle dense OCR tasks and complex, college-level tasks requiring high-resolution input. Exemplary cases demonstrating this capability are included in the Fig.3.\n| Model     | MMMU-val HD Subset |\n| --------- | ------------------ |\n| IXC2      | 46.32              |\n| IXC2-4KHD | 52.63              |\n", "category": "Experimental_Setup"}, {"question": "Is the amount of data used for training clearly specified in the paper, addressing the confusion regarding data usage comparison with SoTA LLaVA-NeXT?", "answer": "True", "category": "Claim_Verification"}]}