text string | source string |
|---|---|
A continue training strategy, where the model is first fine-tuned on our large-scale synthetic data, followed by fine-tuning on a mixture of existing smaller dataset and reweighted synthetic data, achieves superior performance on various offline evaluations, and 2.47% to 7.18% relative improvements on key metrics in pr... | https://arxiv.org/abs/2505.18488v1 |
20000 30000 40000 Cluster Size100101102103Number of Clusters (b) Figure 2: The statistics of the 20k clusters for 100 million documents. The mean with standard deviation of cluster sizes is 5225 ±2972. To synthesize the EC text pairs, we add two types of errors to the clean texts: grammar error, and typing error. The g... | https://arxiv.org/abs/2505.18488v1 |
performance by sequence accuracy, i.e., the exact match between corrected sentence and the target clean sentences, and shows that fine-tuning help while too many steps on small dataset may quickly degrade utility. Our evaluation on error correction matches previous observation on dialogue generation and summarization t... | https://arxiv.org/abs/2505.18488v1 |
trained on public C4 dataset, and model Sfis further fine-tuned from Spon user data in a production FL system [ 30,32]. Model Sfis a privacy-preserving model with formal DP guarantee ϵ <10, and captures the domain information from mobile application. We define a reweighting model parameterized by θ= (θf, θp, θb) as w(θ... | https://arxiv.org/abs/2505.18488v1 |
predicting live metrics across different settings. The absolute value of cross-validation and validation residuals are smaller than training ressiduals as training is the summation over all live metric samples in Eq. (2). After training, our reweighting model parameters are ( θf, θp, θb) = (40 .64,−30.44,−1.59), which ... | https://arxiv.org/abs/2505.18488v1 |
Training Original Data Eval (%) Synthetic Data Eval (%) Method Top-1 Top-1 (w) Top-3 Top-3 (w) Top-1 Top-1 (w) Top-3 Top-3 (w) Original [18] 68.74±0.3871.16±0.5479.34±0.0280.38±0.2671.35±0.3176.24±0.1182.96±0.3086.24±0.26 SynthLB 66.64±1.2868.25±1.26 77.4±0.91 77.93±0.72 87.5±0.12 90.29±0.0792.75±0.0594.30±0.21 Mix 70.... | https://arxiv.org/abs/2505.18488v1 |
achieves good reweighted metrics even if the uniform weighted metrics slightly degrades compared to mixing with the full synthetic data. (4) Continue training. As our synthetic data is large, we propose a continue training strategy: first fine-tune on the full synthetic dataset for 1000 steps (about one epoch), and the... | https://arxiv.org/abs/2505.18488v1 |
Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL http://github.com/google/jax . [2]Christopher Bryant, Mariano Felice, and Ted Briscoe. Automatic annotation and evaluation of err... | https://arxiv.org/abs/2505.18488v1 |
Data weighted training strategies for grammatical error correction. Transactions of the Association for Computational Linguistics , 8: 634–646, 2020. [17]Alisa Liu, Xiaochuang Han, Yizhong Wang, Yulia Tsvetkov, Yejin Choi, and Noah A Smith. Tuning language models by proxy. COLM , 2024. [18]Renjie Liu, Yanxiang Zhang, Y... | https://arxiv.org/abs/2505.18488v1 |
pages 629–639, Toronto, Canada, July 2023. [33]Da Yu, Peter Kairouz, Sewoong Oh, and Zheng Xu. Privacy-preserving instructions for aligning large language models. ICML , 2024. 12 [34]Xiang Yue, Huseyin A Inan, Xuechen Li, Girish Kumar, Julia McAnallen, Hoda Shajari, Huan Sun, David Levitan, and Robert Sim. Synthetic te... | https://arxiv.org/abs/2505.18488v1 |
arXiv:2505.18497v1 [cs.CL] 24 May 2025The Pragmatic Mind of Machines: Tracing the Emergence of Pragmatic Competence in Large Language Models Kefan Yu†*, Qingcheng Zeng†*, Weihao Xuan‡, Wanxin Li♯, Jingyi Wu†, Rob Voigt† †Northwestern University‡The University of Tokyo♯Zhejiang University qcz@u.northwestern.edu Abstract... | https://arxiv.org/abs/2505.18497v1 |
other playful and implicitly affirmative. We prompt LLMs to infer the pragmatic intention behind each continuation and identify the contextual conditions under which one would be preferred over the other. ing body of research has begun to explore whether these models exhibit sensitivity to pragmatic cues. Recent studie... | https://arxiv.org/abs/2505.18497v1 |
models approached human-level accuracy and exhibited similar error patterns. Building on this line of in- quiry, Sravanthi et al. (2024) introduced a com- prehensive benchmark designed to capture more nuanced aspects of pragmatic reasoning beyond conventional multiple-choice formats. Extending this evaluation paradigm ... | https://arxiv.org/abs/2505.18497v1 |
tent, politeness, and implicature. Specifically, we use GPT4o (OpenAI et al., 2024) to generate a reference set of conversations with alternatives ac- companied by human-verified explanations of prag- matic intent (Figure 2A). 3.1 First-round Data Generation In the initial round of data generation, we build on the scen... | https://arxiv.org/abs/2505.18497v1 |
conversation: 1. Conversation structure: ... 2. Pragmatic diversity: … 3. Sentence -Level Intentions: … Generationcontext initial_turn response_1 response_1_intent response_2 response_2_intent 650 datapointsContext: X and Y are colleagues who are leaving work on a Friday at the same time.B. Evaluation 1300 samplesiniti... | https://arxiv.org/abs/2505.18497v1 |
up. response_1_intent This response does not flout any maxim. It aims to directly ac- knowledge the intention to go to the gym while adding a self- reflective comment about recent habits. A speaker might prefer this reply over response_2 be- cause it is straightforward, hon- est, and avoids sarcasm, making the interact... | https://arxiv.org/abs/2505.18497v1 |
of pragmatic competence across training stages. 5 Results 5.1 General Results We present our overall findings from both the 10- point scoring and pairwise win rate comparisons, focusing on how pragmatic competence develops across model training stages. 10-Point Scoring. As shown in Figure 3, models generally achieve hi... | https://arxiv.org/abs/2505.18497v1 |
between LLaMA-3.1 7B and 70B models, which share the same pretraining corpus. In our head-to- head comparison, the 70B model achieves a sub- stantially higher win rate (66%) than the 7B model (34%), providing evidence that increased model capacity also plays an important role in improving pragmatic competence. Similarl... | https://arxiv.org/abs/2505.18497v1 |
intent. In parallel, we also observe a shift to- ward more sociopragmatic competence —the abil- ity to recognize social roles, politeness strategies, and contextual appropriateness—suggesting that the DPO stage broadens the scope of pragmatic strategies beyond purely cognitive interpretations. The full comparison resul... | https://arxiv.org/abs/2505.18497v1 |
63.8 Qwen3-1.7B-Base 67.3 Qwen3-4B-Base 69.5 Table 3: Accuracy on the Goldilocks implicature rea- soning task. Top section shows original results reported by Ruis et al. (2023); bottom section reports our own evaluation of modern base models using the same ex- perimental setup. Beyond Pretraining: Refining Pragmatic Co... | https://arxiv.org/abs/2505.18497v1 |
Computational Linguistics . Mingyue Jian and N. Siddharth. 2024. Are llms good pragmatic speakers? Preprint , arXiv:2411.01562. Cameron R. Jones and Benjamin K. Bergen. 2025. Large language models pass the turing test. Preprint , arXiv:2503.23674. Lauri Karttunen. 1974. Presupposition and linguistic context. Robert Kir... | https://arxiv.org/abs/2505.18497v1 |
by llms. In Proceedings of the 37th International Con- ference on Neural Information Processing Systems , NIPS ’23, Red Hook, NY , USA. Curran Associates Inc. Jerrold M Sadock. 1978. On testing for conversational implicature. In Pragmatics , pages 281–297. Brill. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad... | https://arxiv.org/abs/2505.18497v1 |
GPU. We obtain all open-source models from HuggingFace. The full prompt used in generation is shown below. Prompt Template for Data Generation Task Overview: In this task, we will explore prag- matic alternatives in conversation by constructing a multi-round, tree-structured dialogue. The goal is to analyze why a speak... | https://arxiv.org/abs/2505.18497v1 |
dreading the crowded train. B2: Well, I suppose if you’re absolutely sure my presence won’t ruin your evening... B3: Nah, I could use the walk. Been sitting all day. Intentions: B1: GM: None Intention: Directly accepts and expresses gratitude. Enthusiastic tone. A speaker might use it to show appreciation and comfort. ... | https://arxiv.org/abs/2505.18497v1 |
golden intention for you: ..., this is the core idea of the intention of response_1. Model’s response: ... You should understand that the tested models are base models, thus not deliberately instructed to format their responses and may output random strings and sentences. You just need to check the content of their res... | https://arxiv.org/abs/2505.18497v1 |
You need to purely focus on the content and compare them instead of having preference on the format. Return "Invalid" if: - Either response is empty, nonsense, unrelated to the question, or as long as you feel it is not answering the question - You cannot make a clear choice between the two responses F Examples from 10... | https://arxiv.org/abs/2505.18497v1 |
the notion of being "fashionably late" instead of seriously addressing the tardiness. It aims to use humor and irony to diffuse any awkwardness and avoid direct confrontation. A speaker might prefer this reply over re- sponse_2 because it turns the situation into a joke, attempting to lighten the mood rather than provi... | https://arxiv.org/abs/2505.18497v1 |
could serve to prompt further questions or reinforce respect for wildlife unpredictability. model_1_response: The pragmatic inten- tion behind ‘response_1’ is to express un- certainty, skepticism, or a lack of definitive knowledge about the behavior of alligators. By saying "alligators have their own mind and aren’t th... | https://arxiv.org/abs/2505.18497v1 |
continue without being closed off by a definitive answer. model_2_response: The pragmatic inten-tion behind ‘response_1’ is to express a lack of strong preference or enthusiasm for Chi- nese cuisine, possibly indicating a neutral or indifferent attitude towards it. Someone might prefer ‘response_1’ over ‘response_2’ pr... | https://arxiv.org/abs/2505.18497v1 |
(‘response_2’) is a more lighthearted and possibly flippant response that might not provide the neces- sary comfort or professionalism that the sky- diver is seeking in this context. While ‘re- sponse_2’ might be used in a more joking or casual setting, it lacks the assurance and emphasis on safety that the skydiver ne... | https://arxiv.org/abs/2505.18497v1 |
Knowledge Grafting of Large Language Models Guodong Du1,2Xuanning Zhou1Junlin Li1Zhuo Li1Zesheng Shi1Wanyu Lin2 Ho-Kin Tang1Xiucheng Li1Fangming Liu1Wenya Wang3Min Zhang1Jing Li1 1Harbin Institute of Technology, Shenzhen, China 2The Hong Kong Polytechnic University3Nanyang Technological University duguodong7@gmail.com ... | https://arxiv.org/abs/2505.18502v1 |
a target model and a lightweight SkillPack. This design enhances parameter and storage efficiency, preserves the target model’s general knowledge, and supports continual learning [ 23] by mitigating forgetting. Furthermore, it enables multi-task transfer and model fusion by minimizing parameter conflicts [87, 14]. Know... | https://arxiv.org/abs/2505.18502v1 |
identify limitations in existing methods regarding generalization and adaptability. (2). We propose GraftLLM , which structures cross-model capabilities as SkillPack, offering high performance, forgetfulness resistance, and easy integration for practical applications. (3). Experiments show GraftLLM significantly improv... | https://arxiv.org/abs/2505.18502v1 |
Vector…Embedding& HeadMLP……Attention (c) Forget-Free Learning(a) Cross Capability Transfer Target ModelFully Fine-TunedModule-Aware Adaptive CompressionSource ModelData 2Data 3Data 4SkillPacksData 1DifferenceFigure 4: Overview of GraftLLM .GraftLLM transfers capabilities across heterogeneous LLMs and extracts them into... | https://arxiv.org/abs/2505.18502v1 |
θtgt, θsrcrepresent the parameters of the two models, respectively. Instead of applying uniform compression across all layers, we design a module-specific strategy that selects different compression operators for different parts of the delta. •Embedding and Output Head. These layers are highly sensitive to vocabulary a... | https://arxiv.org/abs/2505.18502v1 |
and (4) Mask-based fusion strategies like TALL Mask [76] and EMR-Merging [34], which leverage unified task vectors and localization. For continual learning, we use LoRA, Model Grafting [ 63], and Model Tailor [ 112] as baselines. Details of all baselines are provided in Appendix D. 4.2 Datasets and Architectures To sho... | https://arxiv.org/abs/2505.18502v1 |
bold , while the performance difference between the two is highlighted in green . Model #ParamsAlpacaEval 2.0 MT-Bench (GPT-4-1106-Preview) (GPT-4-0125-Preview) Win Rate LC Win Rate 1st Turn 2nd Turn Average Score Source LLMs OpenChat-3.5-7B [74] 7B 10.20 14.90 7.14 6.55 6.84 Starling-LM-7B-alpha [111] 7B 14.20 14.70 7... | https://arxiv.org/abs/2505.18502v1 |
8.8 8.4 8.6 7.8 9.0 8.7 Average 48.7 53.8 52.8 55.2 56.0 59.0 61.4 60.4 62.0 63.2 Implicit Knowledge Fusion. We evaluate the effectiveness of implicit heterogeneous model fusion on 10 benchmark tasks, as shown in Table 2, comparing three representative methods. (1) PCB- Merging (pairwise distillation + parameter fusion... | https://arxiv.org/abs/2505.18502v1 |
MLP Modules 68.7 69.2 71.5 70.2 72.1 71.2 Attention Modules 70.7 71.3 71.8 71.2 71.8 72.1 6.2 Effect of Task Difficulty and Data Settings We assess our method across diverse settings, including varying sample sizes, numbers of source models, and task difficulties. We also study the impact of different compression ratio... | https://arxiv.org/abs/2505.18502v1 |
Y . Han, F. Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609 , 2023. [5]Z. Cai, M. Cao, H. Chen, K. Chen, K. Chen, X. Chen, X. Chen, Z. Chen, Z. Chen, P. Chu, et al. Internlm2 technical report. arXiv preprint arXiv:2403.17297 , 2024. [6]M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. D. O. Pinto, J. Kapl... | https://arxiv.org/abs/2505.18502v1 |
A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. [19] Y . Dubois, B. Galambosi, P. Liang, and T. B. Hashimoto. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475 , 2024. [20] Z. Fang, G.... | https://arxiv.org/abs/2505.18502v1 |
C. Huang, P. Ye, T. Chen, T. He, X. Yue, and W. Ouyang. Emr-merging: Tuning-free high-performance model merging. Advances in Neural Information Processing Systems , 37: 122741–122769, 2024. [35] C. Huang, P. Ye, X. Wang, S. Zheng, B. Qi, L. Bai, W. Ouyang, and T. Chen. Seeing delta parameters as jpeg images: Data-free ... | https://arxiv.org/abs/2505.18502v1 |
X. Li, T. Zhang, Y . Dubois, R. Taori, I. Gulrajani, C. Guestrin, P. Liang, and T. B. Hashimoto. AlpacaEval: An automatic evaluator of instruction-following models, 2023. [49] J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, and S. Han. Awq: Activation-aware weight quantization for o... | https://arxiv.org/abs/2505.18502v1 |
M. Sun. Delta-come: Training-free delta-compression with mixed-precision for large language models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS) , 2024. [65] R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn. Direct preference optimization: Your langua... | https://arxiv.org/abs/2505.18502v1 |
. Jiang, P. Xie, F. Huang, and H. Chen. Wise: Rethinking the knowledge memory for lifelong model editing of large language models. arXiv preprint arXiv:2405.14768 , 2024. [80] X. Wang, Y . Zheng, Z. Wan, and M. Zhang. Svd-llm: Truncation-aware singular value decomposition for large language model compression. arXiv pre... | https://arxiv.org/abs/2505.18502v1 |
Yang, Z. Wang, L. Shen, S. Liu, G. Guo, X. Wang, and D. Tao. Adamerging: Adaptive model merging for multi-task learning. In The Twelfth International Conference on Learning Representations (ICLR) , 2024. [94] J. Yang, A. Tang, D. Zhu, Z. Chen, L. Shen, and F. Wu. Mitigating the backdoor effect for multi-task model merg... | https://arxiv.org/abs/2505.18502v1 |
Zhong, C. An, W. Chen, J. Han, and P. He. Seeking neural nuggets: Knowledge transfer in large language models from a parametric perspective. In The Twelfth International Conference on Learning Representations (ICLR) , 2024. [110] M. Zhong, A. Zhang, X. Wang, R. Hou, W. Xiong, C. Zhu, Z. Chen, L. Tan, C. Bi, M. Lewis, e... | https://arxiv.org/abs/2505.18502v1 |
task knowledge that PEFT-based methods [ 11,32] often fail to capture, especially in complex scenarios such as Direct Preference Optimization (DPO), where lightweight adapters struggle to inherit nuanced decision boundaries and preference reasoning from the teacher model. 17 Comparison with Delta Compression Methods. W... | https://arxiv.org/abs/2505.18502v1 |
25.42The routing function Ris guided by either the source model or the task type and dynamically assigns each SkillPack to the target model. Forexplicit knowledge fusion , we train the router using the training datasets provided by FuseChat 2.0 [ 73]. Specifically, we collect the training loss on this dataset from five... | https://arxiv.org/abs/2505.18502v1 |
and MEMIT [57], require identifying specific model regions based on knowledge entities for parameter modification, making them less suitable for LLM detoxification tasks. We conduct our experiments on the SafeEdit benchmark [ 77] using the EasyEdit framework [ 78]. For all methods involving training components, we util... | https://arxiv.org/abs/2505.18502v1 |
GraftLLM (ours) 82.83 98.84 98.46 99.70 99.34 99.08 7.89 64.83 25.81 56.09 Table 10: Detoxification performance and general performance of vanilla LLMs and various detoxifi- cation methods on SafeEdit. The detoxification performance (detoxification success rate) is multiplied by 100. DG-Avg represents the average perfo... | https://arxiv.org/abs/2505.18502v1 |
instructions. Capybara2: We incorporated all the 16,000 examples of Capybara, which is a high-quality collection of multi-turn synthetic conversations. No-Robots3: We included all the 9,500 examples of No-Robots, which is a dataset created by skilled human annotators for supervised fine-tuning. ShareGPT-GPT44: We utili... | https://arxiv.org/abs/2505.18502v1 |
for experts but difficult for laypersons. We evaluate the highest quality diamond set comprising 198 questions. Arena-Hard [46] is a challenging instruction-following benchmark that closely aligns with the human preference ranking from Chatbot Arena [ 7], a crowd-sourced platform for evaluating LLMs. It spans 250 high-... | https://arxiv.org/abs/2505.18502v1 |
a more controlled and reliable preference signal. The data construction process varies by domain: for instruction-following and conversational data, we use an external reward model to evaluate responses; for mathematics and coding domains, responses are verified through rule-based systems. The final dataset Dconsists o... | https://arxiv.org/abs/2505.18502v1 |
primarily report the runtime for quantization. In practice, the compression strategy can be adjusted according to task complexity and model type. Table 14: Storage and time costs of our proposed adaptive compression strategy. The Storage Cost is defined as the ratio of the compressed module size to that of the correspo... | https://arxiv.org/abs/2505.18502v1 |
arXiv:2505.18512v1 [cs.IR] 24 May 2025AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking Soyoung Yoon∗ Seoul National University soyoung.yoon@snu.ac.krGyuwan Kim∗ University of California, Santa Barbara gyuwankim@ucsb.edu Gyu-Hwung Cho Yonsei University jgh0815@yonsei.ac.krSeung-won Hwang† Seoul Nat... | https://arxiv.org/abs/2505.18512v1 |
only a small group of candidates due to input length limits, covering the full list typically requires multiple calls. Prior work has proposed fixed-computation strategies such as sliding windows [ 13] and tournament-style reranking [ 14] to balance ranking effectiveness with computational cost. While these fixed strat... | https://arxiv.org/abs/2505.18512v1 |
adopted in LLM-based reranking systems such as LRL [ 13], RankGPT [ 11], and RankVicuna [ 8], which use zero-shot prompts to perform reranking over win- dowed chunks. While effective, fixed window sizes and traversal paths lead to inefficiencies, by incurring over-computation for simple queries with little ambiguities,... | https://arxiv.org/abs/2505.18512v1 |
goal is to reorder the documents into a ranked list [Dr1>···> D rn]or to extract a top-klist[Dr1>···> D rk], where {r1, . . . , r n}is a permutation of {1, . . . , n }. We assume access to a listwise reranker g(D′;M), where D′is an ordered subset of DandM denotes the underlying reranking model. This is done by promptin... | https://arxiv.org/abs/2505.18512v1 |
to the need to compute dense integrals and multiply small probabilities. To improve scalability, we adopt a more efficient approximation. We define a threshold t(r)such that the expected number of documents whose relevance exceeds t(r)equals the target rank r, i.e.,P iP(xi> t(r)) =r. We then approximate the cumulative ... | https://arxiv.org/abs/2505.18512v1 |
prior beliefs. This iterative process gradually sharpens the model’s belief in document relevance. Stopping criteria (Line 6). Reranking continues until: (i) the number of uncertain documents |C| falls below the threshold τ, (ii) top- kremains unchanged for a fixed number of iterations, or (iii) the loops exceed the ma... | https://arxiv.org/abs/2505.18512v1 |
evaluate multiple pass configurations (SW-1, SW-2, SW-3) to match the computational cost of our method. We use the implementation from the RankLLM codebase5, using the same prompt format, hyperparameters, and input size limits. TourRank- x:A listwise reranking method that runs xindependent multi-stage tournaments over ... | https://arxiv.org/abs/2505.18512v1 |
AcuRank-9 73.3 71.4 70.1 50.1 45.3 39.5 83.2 36.8 30.8 53.0 56.0 37.1 44.7 73.3 54.6 8.8 AcuRank 74.2 71.8 70.3 52.0 47.0 39.4 85.3 37.2 31.8 53.9 56.6 36.5 46.0 75.3 55.5 19.7 AcuRank-H 74.6 70.8 70.5 52.2 47.3 40.4 85.8 37.4 32.1 53.7 56.8 37.5 46.0 75.4 55.7 41.7 AcuRank-HH 74.7 71.8 70.6 51.9 47.0 40.0 86.1 37.5 31... | https://arxiv.org/abs/2505.18512v1 |
5.3) improvement from AcuRank-9 to full AcuRank despite a similar increase in cost, highlighting the advantage of adaptively allocating computation based on uncertainty. Figure 2 visualizes the accuracy–efficiency trade-off across systems. AcuRank consistently lies along the Pareto frontier, achieving stronger accuracy... | https://arxiv.org/abs/2505.18512v1 |
slightly reducing average # calls. Method COVID NFC Signal News R04 Touche DBP Scif Avg. # Calls First-stage retrieval: SPLADE++ED top 100 | Reranker: RankZephyr-7B SPLADE++ED 71.1 34.5 29.6 39.4 45.8 24.4 44.1 69.9 44.8 0.0 SW-1 85.2 37.5 29.7 50.1 62.0 28.9 49.3 75.7 52.3 9.0 SW-2 85.6 37.8 28.9 51.3 62.8 29.2 49.6 7... | https://arxiv.org/abs/2505.18512v1 |
Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 6769–6781, 2020. [3]Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan ... | https://arxiv.org/abs/2505.18512v1 |
preprint arXiv:2003.07820 , 2020. [17] Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. Beir: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round ... | https://arxiv.org/abs/2505.18512v1 |
of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages 2303–2307, 2023. [31] Chuan Meng, Negar Arabzadeh, Arian Askari, Mohammad Aliannejadi, and Maarten de Rijke. Query performance prediction using relevance judgments generated by large language models. arXiv prepri... | https://arxiv.org/abs/2505.18512v1 |
xias a Gaussian random variable, xi∼ N (µi, σ2 i), starting with a prior and updating it iteratively via Bayesian inference. Each listwise reranker output, denoted g(D′;M), is treated as noisy evidence about pairwise preferences among a subset D′of documents, where Mis the reranking model. Using belief propagation on a... | https://arxiv.org/abs/2505.18512v1 |
reflect this, we assign a proportionally larger prior uncertainty ( σi↑) to such documents. The choice of the 1/3 ratio matches the default setting in the original TrueSkill model. Statistically, this means that the interval [µi−µi, µi+µi] = [0 ,2µi]corresponds to [µi±3σi], since 3σi=µi. This range contains approximate... | https://arxiv.org/abs/2505.18512v1 |
license. TrueSkill rating algorithm. We rely on the open-source Python package trueskill (v0.4.5)13, distributed under the permissive BSD 3-Clause License . The TrueSkillTMtrademark and the original Bayesian rating system remain the property of Microsoft, who permits the name and algorithm to be used only in Xbox Live ... | https://arxiv.org/abs/2505.18512v1 |
36.6 32.9 43.9 51.5 58.8 38.9 37.6 44.2 48.4 49.0 41.7 41.7 AcuRank-HH 74.7 71.8 70.6 51.9 47.0 40.0 86.1 37.5 31.3 54.4 57.8 36.1 46.0 75.4 55.8 57.2 ,→avg_calls 53.3 46.6 43.4 60.6 60.2 48.9 60.2 59.8 75.2 52.2 50.6 57.0 68.6 63.9 57.2 57.2 First-stage retrieval: BM25 top 1000 | Reranker: RankZephyr-7B BM25 50.6 48.0... | https://arxiv.org/abs/2505.18512v1 |
in the immediately following row (prefixed by ,→). Method COVID NFC Signal News R04 Touche DBP Scif Avg. # Calls First-stage retrieval: SPLADE++ED top-100 | Reranker: RankZephyr-7B SPLADE++ED 71.1 34.5 29.6 39.4 45.8 24.4 44.1 69.9 44.8 0.0 SW-1 85.2 37.5 29.7 50.1 62.0 28.9 49.3 75.7 52.3 9.0 ,→avg_calls 9.0 9.0 9.0 9... | https://arxiv.org/abs/2505.18512v1 |
Llama-3.3 70B Instruct16[39] and reran the pipeline. Resource constraints limited us to 16meta-llama/Llama-3.3-70B-Instruct 17 Table 7: Dataset-wise NDCG@10 for the AcuRank design-choice ablation, corresponding to Table 3 of our main paper. For every variant the second line (prefixed by ,→) reports the dataset-specific... | https://arxiv.org/abs/2505.18512v1 |
26.0 26.0 25.94 26.0 25.5 25.5 TourRank-5 67.1 64.5 62.2 42.7 38.1 35.1 78.5 32.5 32.8 46.3 48.9 28.3 42.8 66.5 49.0 63.7 ,→avg_calls 65.0 65.0 65.0 65.0 65.0 65.0 65.0 46.7 65.0 65.0 65.0 65.0 64.9 65.0 63.7 63.7 AcuRank-9 66.4 63.4 60.9 41.1 37.7 35.8 77.9 34.4 33.1 49.2 47.5 35.2 42.6 70.8 49.7 8.8 ,→avg_calls 9.0 9... | https://arxiv.org/abs/2505.18512v1 |
sliding windows We further compare AcuRank (with an average of 19.7 calls across 18 datasets) with a Sliding Windows baseline using a similar compute budget (SW-2, 18 calls). For three challenging datasets (Touche, TREC-COVID, and DL-Hard), we split the queries into Easy (NDCG ≥µ) and Hard (NDCG < µ), where µis the dat... | https://arxiv.org/abs/2505.18512v1 |
LLMs may expose token-level or generation-level confidence scores. These signals could be integrated to refine document-level uncertainty estimates or modulate the strength of score updates during reranking. G Example input/output prompt format for listwise reranking We provide concrete prompt examples for RankZephyr [... | https://arxiv.org/abs/2505.18512v1 |
trimester, prenatal care includes blood tests, a physical exam, conversations about lifestyle and more. Prenatal care is an important part of a healthy pregnancy. Whether you choose a family physician, obstetrician, midwife or group prenatal care, here 's what to expect during the first few prenatal appointments. [2] P... | https://arxiv.org/abs/2505.18512v1 |
arXiv:2505.18522v1 [cs.CL] 24 May 2025How Does Sequence Modeling Architecture Influence Base Capabilities of Pre-trained Language Models? Exploring Key Architecture Design Principles to Avoid Base Capabilities Degradation Xin Lu1, Yanyan Zhao1,∗, Si Wei2,∗, Shijin Wang2, Bing Qin1, Ting Liu1 1Research Center for Social... | https://arxiv.org/abs/2505.18522v1 |
[ 51], etc. These architectures primarily inherit stateful modeling mechanisms of linear attention [ 23] and linear RNNs [ 28], offering advan- tages in time and space efficiency over self-attention. Additionally, they introduce new mechanisms like data-dependent decay and delta rule to enhance the expressive power of ... | https://arxiv.org/abs/2505.18522v1 |
achieve base capabilities nearly on par with the Transformer. Furthermore, we extend this validation to the more practical Top-1 Chunk Selection architecture — a direct generalization of the Top-1 Element Selection architecture, and implement GPU Kernels to ensure its time efficiency. Experiments show that the Top-1 Ch... | https://arxiv.org/abs/2505.18522v1 |
levels. Thus, this setting fails to reveal base capabilities difference among architectures during early pre-training and is unsuitable for architecture design and analysis. To address this issue, we propose a Limited Domain Pre-Training with out-of-distribution (OOD) testing framework, where models are pretrained on a... | https://arxiv.org/abs/2505.18522v1 |
(Decay None)Mamba -1 (Full) Mamba -1 (Decay None w/o Conv)Transformer++ Mamba -2 (w/o GN) Mamba -2 (Decay None)Mamba -2 (Full) Mamba -2(Decay None w/o Conv)Mamba -2 (Decay One) W=512 W=256 W=32Transformer++ Synthesizer (Random Constant Q and K) Synthesizer (Only Random Constant K)T = 2−3 T = 2−7T = 2−8T = 2−4T = 2−5T =... | https://arxiv.org/abs/2505.18522v1 |
the sequence context leads to gradual degradation in base capa- bilities. Thus, we conclude that Full-Sequence Visibility is an essential requirement for sequence modeling architectures. 3.2.2 Real Relation Calculation In previous results, we observed that under mixed domain pre-training settings, different models exhi... | https://arxiv.org/abs/2505.18522v1 |
design of the Top-1 Element Selection architecture. (b) The kernel design for key component of the Top-1 Chunk Selection architecture. Supporting full-sequence arbitrary selection in sequence modeling architecture is the key architecture design principle for preventing the degradation of base capabilities. This princip... | https://arxiv.org/abs/2505.18522v1 |
36.09 31.03 6.44 Mamba-2 37.98 35.25 17.60 Transformer++ (W=256) 12.96 41.72 1.63 Architecture Based on the Analyzed Principles Top-1 Element 68.23 43.77 26.95 Top-1 Chunk (Exact) 69.22 44.40 22.32 Top-1 Chunk (Approx.) 61.66 45.38 27.13 Table 1: Results of retrieval tasks. (Model parameters ≈1.3B, pre-trained tokens =... | https://arxiv.org/abs/2505.18522v1 |
This preserves the design principles while enabling efficiency optimizations. Pre-Training LossLoss on OOD Test SetTop-1 Chunk (Exact) Top-1 Chunk (Approx.)Transformer + + Transformer++ (W=256) RWKV -7 Mamba -2Apply all other block designs of RWKV -7 except sequence modelingModels with Sequence Length = 100K 1.201.401.... | https://arxiv.org/abs/2505.18522v1 |
exhibits results and conclusions similar to Top-1 Element Selection: strong performance on general tasks without degradation and advantages in retrieval tasks. 5.3 Long Sequence and Architecture Combination In this section, we evaluate two aspects: first, base capabilities, time efficiency and long-sequence retrieval c... | https://arxiv.org/abs/2505.18522v1 |
I. Johnson, M. Poli, J. Zou, A. Rudra, and C. Re. Zoology: Measuring and improving recall in efficient language models. In The Twelfth International Conference on Learning Representations , 2024. [2]S. Arora, S. Eyuboglu, M. Zhang, A. Timalsina, S. Alberti, J. Zou, A. Rudra, and C. Re. Simple linear attention language ... | https://arxiv.org/abs/2505.18522v1 |
B. Feng, D. Guessous, Y . Liang, and H. He. Flex attention: A programming model for generating optimized attention kernels, 2024. [15] L. Gao, J. Tow, B. Abbasi, S. Biderman, S. Black, A. DiPofi, C. Foster, L. Golding, J. Hsu, A. Le Noac’h, H. Li, K. McDonell, N. Muennighoff, C. Ociepa, J. Phang, L. Reynolds, H. Schoel... | https://arxiv.org/abs/2505.18522v1 |
S. Liu, W. He, E. Yuan, Y . Wang, Z. Huang, H. Yuan, S. Xu, X. Xu, G. Lai, Y . Chen, H. Zheng, J. Yan, J. Su, Y . Wu, N. Y . Zhang, Z. Yang, X. Zhou, M. Zhang, and J. Qiu. Moba: Mixture of block attention for long-context llms, 2025. [28] E. Martin and C. Cundy. Parallelizing linear recurrent neural nets over sequence ... | https://arxiv.org/abs/2505.18522v1 |
D. Chen, R. Le Bras, and Y . Choi. Social IQa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages 4463–4473, Hong Kong, China, Nov. 201... | https://arxiv.org/abs/2505.18522v1 |
mechanism, Jan. 2024. [53] J. Yuan, H. Gao, D. Dai, J. Luo, L. Zhao, Z. Zhang, Z. Xie, Y . X. Wei, L. Wang, Z. Xiao, Y . Wang, C. Ruan, M. Zhang, W. Liang, and W. Zeng. Native sparse attention: Hardware-aligned and natively trainable sparse attention, 2025. [54] M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alb... | https://arxiv.org/abs/2505.18522v1 |
24 layers and a state size of 128 for the small-scale version; the large-scale version has a hidden dimension of 2048, 48 layers and the same state size of 128. RWKV-5/6/7 [ 32,33] are implemented based on the official open-source code4, with a hidden dimension of 768, 12 layers and 12 attention heads for the small-sca... | https://arxiv.org/abs/2505.18522v1 |
batch sizes for small-scale models were 256 (sequence length=2k) and 8 (sequence length=100k), while that for large-scale models was 128 (sequence length=2k). All models were trained in a distributed manner across 8 Nvidia Tesla A100 GPUs, totaling approximately 1,221 days of single-GPU equivalent training time. B Eval... | https://arxiv.org/abs/2505.18522v1 |
works, such as NSA [ 53] and MoBA [ 27], introduce dynamic block selection-based sparsification schemes that effectively balance computational efficiency and practical performance. The core of our work is to reveal how sequence modeling architectures influence base capabilities. Although the improved architecture we ul... | https://arxiv.org/abs/2505.18522v1 |
arXiv:2505.18524v1 [cs.CL] 24 May 2025metaTextGrad : Automatically optimizing language model optimizers Guowei Xu Tsinghua UniversityMert Yuksekgonul Stanford University Carlos Guestrin Stanford UniversityJames Zou Stanford University Abstract Large language models (LLMs) are increasingly used in learning algorithms, e... | https://arxiv.org/abs/2505.18524v1 |
of different optimizers to propose a more effective optimizer. as well as any information such as model gradients, are inaccessible, and only the inputs to and outputs of the LLM are observable. For the task to be optimized, we require only a small training dataset containing some input-output pairs and an evaluation m... | https://arxiv.org/abs/2505.18524v1 |
the training scheme. 2.Propose : Suggest improvements to the program, which may involve modifications to the pipeline structure, prompts, or other aspects. 3.Update : Update the current optimal program based on the evaluation results of the improved program and determine the next proposal. 4.ExtractOptimizedProgram : R... | https://arxiv.org/abs/2505.18524v1 |
then present the two core components of metaTextGrad : the meta prompt optimizer and the meta structure optimizer. Finally, we integrate both components to form the complete metaTextGrad pipeline. 3.1 Theoretical Insight We present the theoretical motivation for meta optimization, highlighting the importance of alignin... | https://arxiv.org/abs/2505.18524v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.