text string | source string |
|---|---|
words, and even more difficult to state really clearly because creativity by definition involves not only novelty but value, and because values are highly variable, it follows that many arguments about creativity are rooted in disagreements about value.’ (Boden 2009). From a cognitive perspective, creativity can be see... | https://arxiv.org/abs/2505.19277v1 |
are developed through social interaction; b) creativity in itself is often the result of explicit moments of collaboration between individuals; c) creativity is largely defined by social judgement or validation; and d) creativity exists only in relation to an established ensemble of cultural norms and products that bot... | https://arxiv.org/abs/2505.19277v1 |
the MC’s voice and style for natural delivery. The current style module evolves through improvisation and an archive corpus of past performances. A cadence algorithm analyses syllabic timing and rhythm, while a rhyme module scores rhyme density and complexity. These components interact dynamically, with archival data r... | https://arxiv.org/abs/2505.19277v1 |
Modelling human subconscious processing as a generative AI paradigm would be the ultimate development goal. Currently, even the best AI systems have proven only to be convincing creative mimics; they lack the granular inputs that a human possesses. The goal of generative AI development should be to enhance human creati... | https://arxiv.org/abs/2505.19277v1 |
human creativity. By combining Diverse Beam Search (Kasai et al. 2024) with LLM-as-a-Judge self-evaluation (Zheng et al. 2023), CBS approximates a simple yet effective generate-and-test loop that may be adapted for interactive domains like freestyle rap. These strategies represent a shift from purely predictive models ... | https://arxiv.org/abs/2505.19277v1 |
it was produced. Some scholars argue that intrinsic motivation is not a strict requirement for creativity, except perhaps for a special exemplary form of it (Paul and Stokes 2023). An AI system might lack the spontaneous urges of a human artist, but it can still generate outcomes that meet creative criteria. If an LLM ... | https://arxiv.org/abs/2505.19277v1 |
flow is not merely a matter of linguistic coherence, it involves rapid adaptation to adversarial input, rhythmic synchronisation with a beat, and the spontaneous expression of wit, emotion, and identity. These performances unfold under high-tempo, high-stakes conditions where responsiveness and timing are essential. In... | https://arxiv.org/abs/2505.19277v1 |
is missing and what might yet be built, if we shift our focus from static benchmarks to dynamic interaction. Acknowledgements This work was supported by the Engineering and Physical Sciences Research Council. Refer ences Adams, K. 2009. On the Metrical Techniques of Flow in Rap Music. Music Theory Online 15 . Garcia, M... | https://arxiv.org/abs/2505.19277v1 |
Z.; Nanjappan, V.; Lee, L.-H.; Soomro, S. A.; and, G. V. G. 2023. Exploration of the relationship between culture and experience of creativity at the individual level: A case study based on two design tasks. International Journal of Design Creativity and Innovation 11(3):185–208. Guilford, J. P.; Christensen, P. R.; Me... | https://arxiv.org/abs/2505.19277v1 |
of the National Academy of Sciences 118(25):e2022340118. Patel, D. 2023. Ilya sutskever (openai chief scientist) - why next-token prediction could surpass human intelligence. Paul, E. S., and Stokes, D. 2023. Creativity. Pescapè, A. 2024. Exploring the current state and future potential of generative artificial intelli... | https://arxiv.org/abs/2505.19277v1 |
arXiv:2505.19286v2 [cs.CL] 27 May 2025A Graph Perspective to Probe Structural Patterns of Knowledge in Large Language Models Utkarsh Sahu1, Zhisheng Qi1, Yongjia Lei1, Ryan A. Rossi2, Franck Dernoncourt2, Nesreen K. Ahmed3, Mahantesh M Halappanavar4, Yao Ma5, Yu Wang1 1University of Oregon,2Adobe Research,3Cisco AI Res... | https://arxiv.org/abs/2505.19286v2 |
in LLMs and the lim- ited exploration in this field, we take a fresh graph- based perspective to uncover the structural patterns of knowledge encoded in LLMs as shown in Fig- ure 1. Building on these derived structural patterns, we develop graph machine learning models to iden- tify more informative knowledge for fine-... | https://arxiv.org/abs/2505.19286v2 |
Rings et al., 2022): K(vi) =|T(vi)|−1X (vi,rij,vj)∈T(vi)K(vi, rij, vj) (1) Note that the above neighborhood aggregation to obtain the knowledgeability score for each entity also applies to temporal triplets (vi, rij, vj, t)∈ T(vi), allowing us to account for the temporal im- pact when assessing an entity’s knowledgeabi... | https://arxiv.org/abs/2505.19286v2 |
of an entity’s triplets are recognized. These patterns exhibit clear domain- specific variation. Specialized datasets such as PharmKG8K and MVPKG are left-skewed, with a dominant peak at 0.0 reflecting LLM’s limited knowledge coverage in domains like pharmaceu- ticals and political science. In contrast, general- purpos... | https://arxiv.org/abs/2505.19286v2 |
T- Rex exclusively contains Wikipedia entities, which are generally well represented in LLM training cor- pora, even for less popular or low-degree entities. Finding 4 - Table 1 demonstrates strong re- gression in predicting node knowledgeability, with absolute errors between 0.15 and 0.25. Compar- ing models using tex... | https://arxiv.org/abs/2505.19286v2 |
select more informative triplets for effective fine-tuning LLMs. 5 Limitations The limitations of this paper are as follows: •More applications : The derived structural pat- terns are used solely to guide triplet selection for fine-tuning. However, these patterns hold broader potential. For instance, they could inform ... | https://arxiv.org/abs/2505.19286v2 |
Wide Web Confer- ence, pages 2865–2871. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221 . 5 Stephanie Lin, Jacob Hilton, ... | https://arxiv.org/abs/2505.19286v2 |
Yu Wang and Tyler Derr. 2021. Tree decomposed graph neural network. In Proceedings of the 30th ACM in- ternational conference on information & knowledge management , pages 2040–2049. Shirley Wu, Shiyu Zhao, Michihiro Yasunaga, Kexin Huang, Kaidi Cao, Qian Huang, Vassilis Ioannidis, Karthik Subbian, James Y Zou, and Jur... | https://arxiv.org/abs/2505.19286v2 |
from graphs (Luo et al., 2023), and evaluating factuality hallucinations by using false premise questions (Zhu et al., 2024). These ap- proaches look into knowledge and trustworthiness checking but treat the model as a black box, leaving its underlying structural patterns unexplored. B.3 Topological Understanding of LL... | https://arxiv.org/abs/2505.19286v2 |
we used a strongly con- nected component of 98,537 edges, 6,877 enti- ties, and 29 relations. •WD50k (Galkin et al., 2020): The WD50K dataset was created using the Wikidata RDF dump of August 2019. It has 233,838 edges and 41,334 entities. Since being extracted from Wiki- data, there were 14,858 triplets common between... | https://arxiv.org/abs/2505.19286v2 |
Baseline performance is measured by querying each base model on this evaluation set prior to any fine-tuning. The performance metric is the percentage of correct responses by the model on the evaluation set. •Fine-Tuning Budget and Initial Query: We then set a budget that the LLM can be fine-tuned on, and the size of t... | https://arxiv.org/abs/2505.19286v2 |
are well covered by pre-training corpora. E.2 Deepseek V3 See Figure 5 for an overview of model behavior. •Knowledgeability Distribution: We observe a trimodel pattern with a relatively small peak at 0.5. Entities with a knowledge value of 0 are more common than those with a value of 1, es- pecially in domain-specific ... | https://arxiv.org/abs/2505.19286v2 |
for each dataset; (c): Node knowledgeability increases as node degree increases. Figure 6: Gemini (a): Distribution of node knowledge- ability for each dataset; (b): Distribution of node ho- mophily for each dataset; (c): Node knowledgeability increases as node degree increases. Figure 7: GPT4o (a): Distribution of nod... | https://arxiv.org/abs/2505.19286v2 |
statement below; reply only True orFalse . Given: TripletT= (sub,rel,obj), Date D. Relational Template Map: T:rel7→“{sub} . . . { obj}”. Procedure: 1. Retrieve template t=T(rel). 2. Instantiate base statement S0=t[{sub}→sub,{obj}→obj]. 3. Append date: S=S0onD. 4. Send System Msg +User Msg: Sto LLM. 5. Return “True” or ... | https://arxiv.org/abs/2505.19286v2 |
arXiv:2505.19293v1 [cs.CL] 25 May 2025 -LongBench: Are de facto Long-Context Benchmarks Literally Evaluating Long-Context Ability? Wang Yang1, Hongye Jin2, Shaochen Zhong3, Song Jiang4, Qifan Wang4 Vipin Chaudhary1, Xiaotian Han1 1Case Western Reserve University2Texas A&M University3Rice University4Meta {wxy320,vipin,x... | https://arxiv.org/abs/2505.19293v1 |
Non-reflective synthetic benchmarks such as NIAH or Passkey Retrieval, where the source (e.g., a string of digits or a phrase) bears no semantic or task relevance to the padding content (e.g., unrelated blog posts) to evaluate models. (2) They adopt a fixed input length per data sam- ple, making them suitable only for ... | https://arxiv.org/abs/2505.19293v1 |
1. The performance of LM- Infinite exhibits significant variation across tasks of different lengths within the same dataset. Many long-context datasets have uneven length distribu- tions, introducing biases in evaluating a model’s long-context capability. To validate this hypothesis, we train models using five differen... | https://arxiv.org/abs/2505.19293v1 |
Base Ability (performance on short texts). 3.1 Construct a new long-context benchmark We categorize tasks into four types, each con- sisting of two tasks with different levels of diffi- culty, resulting in a total of eight tasks. The types and their corresponding tasks are: Key Retrieval (including KV Retrieval and Cou... | https://arxiv.org/abs/2505.19293v1 |
article are combined to form the whole context, ensuring that the whole context length is less than 128k and the order of all articles is shuffled. The bottom of the table contains different datasets from other benchmarks. N/A indicates that the task does not require Context Sources because the questions are synthetic ... | https://arxiv.org/abs/2505.19293v1 |
context Base Ability . It refers to the model’s score when conducting short-context tasks. To estimate Base Ability, we sample Ninstances from short text lengths (like 2k,4k,6k). For each length, N/3 samples are selected, and the model’s average score across these lengths is computed: Base Ability =S2k+S4k+S6k 3(1) whe... | https://arxiv.org/abs/2505.19293v1 |
no longer suitable.•∞-Bench (Zhang et al., 2024) and L-Eval (An et al., 2023) are an improvement over bench- marks like LongBench, increasing the data length to over 128k. However, the context length is not controllable, which limits its ability to comprehensively evaluate LLMs. •NIAH and RULER (Hsieh et al., 2024) are... | https://arxiv.org/abs/2505.19293v1 |
methods, NTK and PI, using LongBench and -LongBench. On -LongBench, we evaluate performances with two metrics: score and LongScore (LC). We include three evaluations to further validate the discrimi- native power and practical value of our proposed LongScore metric. These comparisons were cho- sen to reflect real-world... | https://arxiv.org/abs/2505.19293v1 |
SCORE ) 100-LongBenchPI 19.18 16.47 17.67 17.10 17.67 0.44 13.87 -27.68 NTK 19.39 15.72 16.53 16.70 17.17 12.88 15.83 -18.40 100-LongBenchLLaMA3-8B (ratio=1) 35.37 37.08 1.45 1.87 0.52 0.99 7.13 -79.84 LLaMA3-8B (ratio=64) 32.52 31.94 25.34 26.08 26.94 1.63 18.83 -42.12 HEMLETGemini-1.5-Flash 59.6 – 60.2 58.1 55.0 50.7... | https://arxiv.org/abs/2505.19293v1 |
32k 64k 128k 256k KV Retrieval100 75 50 25 025 0-8k 8k 16k 32k 64k 128k 256k Multi-Doc QA100 75 50 25 025 0-8k 8k 16k 32k 64k 128k 256k Multi-Doc Sum100 80 60 40 20 020 0-8k 8k 16k 32k 64k 128k 256k Passage Count100 80 60 40 20 0Long-T ext Capacity 0-8k 8k 16k 32k 64k 128k 256k Passage Retrieval100 80 60 40 20 0 0-8k 8... | https://arxiv.org/abs/2505.19293v1 |
-59.66 -78.30 -87.54 -3.73 -12.05 -34.85 -50.79 -67.35 -64.89 -6.40 -6.74 -24.82 -25.16 -50.08 -4.45 -10.17 -6.52 -12.16 -59.57 -16.96 -29.13 -47.36 -37.28 -86.11 -100.00 -22.81 -28.67 -64.75 -70.61 -97.53 -100.00Single-Doc Sum 80 60 40 20 0 100 80 60 40 20 020 100 80 60 40 20 020 100 80 60 40 20 020 100 80 60 40 20 02... | https://arxiv.org/abs/2505.19293v1 |
encoding schemes (Li et al., 2021; Xiong et al., 2023; Hsu et al., 2024). This trend is well-documented in recent technical reports (Yang et al., 2024; Abdin et al., 2024; Dubey et al., 2024), which highlight how careful adjustments to train- ing schedules, data distribution, and architecture design contribute to stabl... | https://arxiv.org/abs/2505.19293v1 |
training strategies to enhance long-context capabilities. Limitations The proposed metric requires models to demon- strate relatively strong base ability on the task. If a model’s base ability is insufficient, subsequent evaluations of long-context capabilities may exhibit significant fluctuations, making it less effec... | https://arxiv.org/abs/2505.19293v1 |
Preprint , arXiv:2406.15019. Tianyu Gao, Alexander Wettig, Howard Yen, and Danqi Chen. 2024. How to train long-context language models (effectively). arXiv preprint arXiv:2410.02660 . Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. 2024. Lm- infinite: Zero-shot extreme length generalizat... | https://arxiv.org/abs/2505.19293v1 |
benchmark for evaluating long-context large language models. Preprint , arXiv:2403.11802. Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: En- hanced transformer with rotary position embedding. Neurocomputing , 568:127063. Mingjie Sun, Xinlei Chen, J Zico Kolter, and Zhuang Liu... | https://arxiv.org/abs/2505.19293v1 |
the proportion of each text length within the entire dataset. The larger the marker, the higher the pro- portion. The results exhibit significant variation across tasks of different lengths within the same dataset. All results are in Appendix A.1. A.2 Details about how to construct each task KV Retrieval . This task pr... | https://arxiv.org/abs/2505.19293v1 |
context. (3) Evaluation Metric: The task is eval- uated using accuracy ( Acc). If the model correctly identifies the count of unique articles, its accuracy score is incremented by one. Single-Doc QA . The task evaluates a model’s ability to answer questions specific to a single article within a multi-article context. (... | https://arxiv.org/abs/2505.19293v1 |
generate the correct answer. (2) Evaluation Metric: Similar to the Single-Doc QA task, the model’s answers are evaluated using another large language model and evaluated by the same dimensions. (3) Prior Knowledge Filtering is similar to the Single- Doc QA task. Single-Doc Sum . The task evaluates a model’s ability to ... | https://arxiv.org/abs/2505.19293v1 |
enter the number of the passage that the summa- rization is from. The answer format must be like "Passage 1", "Passage 2", etc. \n\n The answer is Passage Passage Count .There are some paragraphs below sourced from many different fields. Some of them may be duplicates. Please carefully read these paragraphs and determi... | https://arxiv.org/abs/2505.19293v1 |
evaluate three model families (Llama 3.2, Llama 3.1, and Phi 3), selecting two different model sizes from each family. Given that these models are from the same series but vary in size, the expected trends on the dataset are as follows: (1) Model Size Effect: Larger models should gen- erally achieve higher scores compa... | https://arxiv.org/abs/2505.19293v1 |
and healthcare) to en- hance the comprehensiveness of our benchmark. Evaluating the capability of LLMs to handle such domain-specific scenarios is indeed a crucial need. Specifically, we mix up CaseSumm, MedOdyssey, and Medical Summary into our original dataet. We reevaluate the performance of the LLaMA 3.2 1B-Instruct... | https://arxiv.org/abs/2505.19293v1 |
Towards Reliable Large Audio Language Model Ziyang Ma1, Xiquan Li1, Yakun Song1, Wenxi Chen1, Chenpeng Du2, Jian Wu2, Yuanzhe Chen2, Zhuo Chen2, Yuping Wang2, Yuxuan Wang2, Xie Chen1,3† 1X-LANCE Lab, School of Computer Science, MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University 2ByteDance,3Shanghai I... | https://arxiv.org/abs/2505.19294v1 |
successful efforts to en- hance the reliability of language models in text- based models (Yang et al., 2024d; Cheng et al., 2024a; Yona et al., 2024; Zhang et al., 2024; Xu et al., 2024a), reliable LALMs remain largely un- explored. Building reliability in LALMs is crucial for applications where the model’s confidence ... | https://arxiv.org/abs/2505.19294v1 |
2025b), speaker di- arization (SD) (Shi et al., 2024; Meng et al., 2024), and speech emotion recognition (SER) (Xu et al., 2024b; Lin et al., 2024; Cheng et al., 2024b; Kang et al., 2024) are critical, which involve either acous- tic features, semantic features, or both, specific to speech. Furthermore, some works tack... | https://arxiv.org/abs/2505.19294v1 |
balancing both the model’s correct re- sponses and its ability to reject uncertain answers. 3 Reliable LALM The core purpose of reliable LALM is to refuse to answer when the model doesn’t know the answer given the input audio and instruction. We explored two distinct approaches to enhance the model’s re- liability: tra... | https://arxiv.org/abs/2505.19294v1 |
(Yang et al., 2024d; Cheng et al., 2024a; Zhang et al., 2024), we adopt a similar approach but apply it to the multi-modal setting. The training-based method involves two key steps: construction of a model-specific IDK dataset and post-training of the model. Construction of the IDK Dataset. Given that different models ... | https://arxiv.org/abs/2505.19294v1 |
as Accu- racy, Truthfulness, Rejection Rate, and Reliability are useful for evaluating the absolute capability of a model to express IDK, they are less effective in measuring the relative effectiveness of different re- liable methods. Specifically, these metrics fail to reveal how well a method balances two crucial as-... | https://arxiv.org/abs/2505.19294v1 |
RGI metric, we can easily answer this question. Specifically, if we train the model on one modality and test it on another, an RGI > 0indicates that the model has successfully learned to reject more questions it does not know, even when tested on a different audio modality. Such transferability is crucial for building ... | https://arxiv.org/abs/2505.19294v1 |
Section 3.1. The specific prompts used in these approaches are provided in Appendix C.1. For the training-based methods, since the Qwen-Audio series does not provide fine-tuning code, we im- plement our fine-tuning process based on Deep- Speed1, using Low-Rank Adaptation (LoRA) (Hu 1https://github.com/microsoft/DeepSpe... | https://arxiv.org/abs/2505.19294v1 |
higher for the sound and music, while it is relatively lower for the speech. This suggests that the model is more confi- dent about what it knows and does not know in the sound and music. The use of Task Agent (i.e., ASR results) or SFT helps mitigate this issue in speech, improving the model’s reliability. From all th... | https://arxiv.org/abs/2505.19294v1 |
from learning to reject unknown answers, while a larger LoRA alpha weight leads to over-conservatism. In Figure 4(b), the Reliabil- ity metric initially increases with the LoRA alpha weight but eventually decreases, demonstrating a non-monotonic relationship. Figure 4(c) shows that the RGI value decreases as the LoRA a... | https://arxiv.org/abs/2505.19294v1 |
Zhengfu He, Kai Chen, and Xipeng Qiu. 2024a. Can ai assistants know what they don’t know? Proc. ICML . Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Jingdong Sun, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. 2024b. Emotion- LLaMA: Multimodal emotion recognition and rea- soning with instruction ... | https://arxiv.org/abs/2505.19294v1 |
few- shot learning and dialogue abilities. Proc. ICML . Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023a. Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large lan- guage models. Proc. ICML . Xiquan Li, Wenxi Chen, Ziyang Ma, Xuenan Xu, Yuzhe Liang, Zhisheng Zheng, Qiuqiang... | https://arxiv.org/abs/2505.19294v1 |
and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak su- pervision. Proc. ICML . S Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha. 2024. Mmau: A massive multi-task audio under- standing and reasoning benchmar... | https://arxiv.org/abs/2505.19294v1 |
Wu, Tianhao Shen, Ge Zhang, Yuhang Wu, Cong Liu, Ziya Zhou, et al. 2024. Chatmusician: Un- derstanding and generating music intrinsically with llm. Proc. ACL . Hanning Zhang, Shizhe Diao, Yong Lin, Yi R Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. 2024. R-tuning: Teaching large language models t... | https://arxiv.org/abs/2505.19294v1 |
30.63 37.54 37.06 43.60 46.60 46.51 SALMONN ✓ 28.53 87.69 52.69 17.96 85.03 40.05 8.41 92.79 21.59 18.30 88.50 39.22 Qwen2-Audio-Instruct ✗ 60.96 60.96 60.96 55.09 55.09 55.09 50.75 50.75 50.75 55.60 55.60 55.60 Qwen2-Audio-Instruct ✓ 58.26 76.28 73.03 54.19 66.77 65.19 43.84 58.26 56.18 52.10 67.10 64.85 model with a ... | https://arxiv.org/abs/2505.19294v1 |
Output : {Answer } Table 5: The Prompt Template for IDK Prompting on MMAU. MCoT Prompting Input : {Few-shot Examples } {Audio } {Question } Select one option from the provided choices: {Content_of_A } {Content_of_B } {Content_of_C } {Content_of_D } Let’s think step by step. You can first analyze the sound, music, or sp... | https://arxiv.org/abs/2505.19294v1 |
of ρless than 1−αis satisfied. For an illustrative example, consider a model initially producing 50% correct and 50% incorrect answers. By applying Equation 5, we can calculate the Reliability of the original unreliable model as: Rel= 0×50% + 1 ×50% = 50% . (17) After applying a reliable method, if 10% of both the corr... | https://arxiv.org/abs/2505.19294v1 |
arXiv:2505.19299v1 [cs.CL] 25 May 2025A Necessary Step toward Faithfulness: Measuring and Improving Consistency in Free-Text Explanations Lingjun Zhao University of Maryland College Park, Maryland, USA lzhao123@umd.eduHal Daumé III University of Maryland College Park, Maryland, USA hal3@umd.edu Abstract Faithful free-t... | https://arxiv.org/abs/2505.19299v1 |
an explanation ecannot simultaneously support a prediction yand its negation ¬y. For example, in Figure 1, we see that a model uses vir- tually the same explanation (“use a lot of positive language’) to argue in favor of a hotel review being both authentic and deceptive in this opinion spam classification task. Our wor... | https://arxiv.org/abs/2505.19299v1 |
ex- planations can still be self-inconsistent (Camburu et al., 2020; Zhou et al., 2023), or optimized in terms of plausibility instead of faithfulness (Kumar and Talukdar, 2020). Most methods that jointly explain and make prediction (Rajani et al., 2019; Narang et al., 2020; Ling et al., 2017; Jung et al., 2022; Ramnat... | https://arxiv.org/abs/2505.19299v1 |
We ob- serve that language model generated explanations can be inconsistent, e.g. in Figure 1, the expla- nation “use a lot of positive language” supports both the truthful and deceptive hypotheses, failing to distinguish why the model predicted answer a instead of the alternative prediction ¬a. As a re- sult, the expl... | https://arxiv.org/abs/2505.19299v1 |
preference dataset for direct preference optimization (DPO), improving PEX consistency. Finally, we evaluate whether the consistency-optimized explanations are more faithful. C(e) = logM(e|Q(q,a)) M(e|Q(q,¬a))(1) where ¬ais a negation from a, the text prompt Qand explanation generation M(e|Q(q,a))are formatted in § 3.1... | https://arxiv.org/abs/2505.19299v1 |
elfrom the explanations generated by the refer- ence language model M. For each question qand answer apredicted by model M, we sample and rank the explanations using PEX consistency score (Eq 4), as described in §3.3. We consider explana- tions in the top p%are consistent, and those in the bottom p%are inconsistent. Fo... | https://arxiv.org/abs/2505.19299v1 |
1,000 deceptive reviews. We restrict reviews to those reviews containing at least 120 words to ensure that there is sufficient context for explana- tions. We split the selected reviews to obtain 1,200 pairs of (review, label) for training, 400 pairs for validation and 400 pairs for testing. We fine-tune the models on t... | https://arxiv.org/abs/2505.19299v1 |
performance (F1 score) on the test split, ensuring teacher model’s explanations are not provided as input to prevent label leakage (§4). The simulation F1 score is computed by using the teacher model’s predictions as ground-truth la- bels. We report average F1 score across all student model training passes. 6 Experimen... | https://arxiv.org/abs/2505.19299v1 |
score of 4.4. Similarly, the highest consistency score among DPO-sampled explanations for Mistral model av- erages -2.8 on the Amazon dataset, which is lower than the Yi-1.5 model’s highest score of 0.3. 6.3 Are explanations optimized for consistency also more faithful? Systems. We evaluate the faithfulness of explana-... | https://arxiv.org/abs/2505.19299v1 |
explanations with a higher consistency score. any explanations during student model training. We use the same review across all three systems to enable a fair comparison. Consistency-optimized explanations improve ex- planation faithfulness. Table 1 presents the stu- dent model’s simulation performance on the test set,... | https://arxiv.org/abs/2505.19299v1 |
robustness of interpretability methods. arXiv preprint arXiv:1806.08049 . Eleftheria Briakou, Navita Goyal, and Marine Carpuat. 2023. Explaining with contrastive phrasal highlight- ing: A case study in assisting humans to detect trans- lation differences. In Proceedings of the 2023 Con- ference on Empirical Methods in ... | https://arxiv.org/abs/2505.19299v1 |
In Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics , pages 8730–8742, Online. Association for Computational Linguistics. Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernan- dez, Dustin Li, Esin Durmus, Evan Hubinger, Jack- son Kernion, Ka... | https://arxiv.org/abs/2505.19299v1 |
the teacher aid students? Transac- tions of the Association for Computational Linguis- tics, 10:359–375. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neu- ral Inf... | https://arxiv.org/abs/2505.19299v1 |
for Computational Linguistics. Xi Ye and Greg Durrett. 2022. The unreliability of explanations in few-shot prompting for textual rea- soning. Advances in neural information processing systems , 35:30378–30392.Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Guoyin Wang, Heng Li, Jiangcheng Zhu, Ji... | https://arxiv.org/abs/2505.19299v1 |
do offer unique services, the presentation of this detail, without further elaboration or context, seems exaggerated and implausible, casting doubt on the review’s authenticity. We use the following GPT-4 generated expla- nation as one-shot example to prompt models for explanations (§3.1) for the Amazon product review ... | https://arxiv.org/abs/2505.19299v1 |
student models for 100epochs using 10 Model Mistral Llama-2 Yi-1.5 k=10 k=20 Avg k=10 k=20 Avg k=10 k=20 Avg Pred Only 44.0±7.8 66.9±1.5 55.5 54.7±0.1 64.2±1.4 59.6 62.4±0.0 63.6±0.0 63.0 + SFT 64.2±1.8 68.1±0.2 66.2 56.0±4.0 60.6±2.2 58.3 66.5±0.0 65.4±0.7 66.0 + DPO 69.1±1.2 70.4±1.0 69.8 61.4±0.9 65.3±2.9 63.4 69.3±... | https://arxiv.org/abs/2505.19299v1 |
reality the physical buttons are too hard to press and the case does not slide in and out of pockets as easil y due to the added bulk from the case. \n[reason3] Overstates durability: The review states it provides good protection and can be used for a long time, but in reality it cracks after a few days of use and the ... | https://arxiv.org/abs/2505.19299v1 |
but nothing tough.mines called it quits after about 9 months because the rubber started seperating from the plastic, it arrived in an extremely great condition, didnt feel cheap and the design was great down to the way the rubber and plastic are bonded. can be superglued together if it starts to rip (usually after a LO... | https://arxiv.org/abs/2505.19299v1 |
arXiv:2505.19300v1 [cs.CL] 25 May 2025SITUATED THINKER : Grounding LLM Reasoning with Real-World through Situated Thinking Junnan Liu, Linhao Luo, Thuy-Trang Vu, Gholamreza Haffari Department of Data Science and AI Faculty of Information Technology, Monash University, Australia Abstract Recent advances in large languag... | https://arxiv.org/abs/2505.19300v1 |
and self-correction, instead of relying on a predefined workflow [ 45]. Moreover, the dynamic nature of the external world necessitates that LLMs adjust their thinking processes in response to evolving environments ( C.3). This requires LLMs to develop generalizable real-world grounded reasoning capabilities rather tha... | https://arxiv.org/abs/2505.19300v1 |
action, and situated action. Next, we describe SITUATED THINKER ’s training process, which encourages LLMs to perform complex reasoning about the real world through a deliberative situated thinking approach. 2.1 Situated Thinking The situated thinking is designed to enable LLMs to conduct complex reasoning by combining... | https://arxiv.org/abs/2505.19300v1 |
engage with the external world to conduct reasoning, which is called situating action. For example, as shown in § 3.6, the LLM first uses internal actions to reason and analyze questions. It then realizes it lacks information about the current president of East Timor and formulates a query to expand its knowledge by as... | https://arxiv.org/abs/2505.19300v1 |
to finalize the answer. The assistant cannot invoke each interface more than {Invoke Limit} times. The following are the interfaces provided for the Assistant: {Placeholder for Interface Definitions} Question. The system prompt is followed by the specific question, to which the model responds through an iterative and d... | https://arxiv.org/abs/2505.19300v1 |
training data. During the training process, we provide two interfaces for the model: 1) information retrieval interface , which retrieves useful information from Wikipedia (2018 dump [ 17]); 2) code execution interface , which executes Python code generated by LLMs and returns feedback. More details of interface defini... | https://arxiv.org/abs/2505.19300v1 |
interfaces (e.g., coding) beyond simple retrieval (Appendix B.1). Additionally, our improvements are evident across both in-domain (MusiQue) and out-of-domain benchmarks (e.g., HotpotQA, 2WikiMultiHopQA, and Bamboogle). 3.3 Performance on Mathematical Reasoning Benchmarks Benchmarks. We assess the mathematical reasonin... | https://arxiv.org/abs/2505.19300v1 |
both different vertical domains and different interfaces. Particularly for TextWorld, a dataset based on simulated physical environments, where the underlying LLMs lack intrinsic knowledge of the environment, effectively utilizing the interface’s name to interact with the external world leads to significant performance... | https://arxiv.org/abs/2505.19300v1 |
THINKER firstly decomposes complex reasoning problems and solves them incrementally, a capability acquired automatically during model training without relying on annotated data. Discovery Knowledge Boundary. Thered highlighted part indicates that SITUATED THINKER is capable of recognizing the limitations of its own kno... | https://arxiv.org/abs/2505.19300v1 |
thinking or backtracking and to utilize external information to validate their conclusions during the reasoning process. Recent studies [ 5,15,40,19] have attempted to bolster LLMs with reinforcement learning using web search or writing code, which allow thinking and reflecting with external information. However, train... | https://arxiv.org/abs/2505.19300v1 |
Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, ... | https://arxiv.org/abs/2505.19300v1 |
ACM, 2023. A.3 [19] Xuefeng Li, Haoyang Zou, and Pengfei Liu. Torl: Scaling tool-integrated RL. CoRR , abs/2503.23383, 2025. 3.3, 4 [20] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In ICLR . ... | https://arxiv.org/abs/2505.19300v1 |
Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems , 36: 68539–68551, 2023. 1 [33] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Ol... | https://arxiv.org/abs/2505.19300v1 |
Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models. Trans. Mach. Learn. Res. , 2024, 2024. 4 [47] Yi Ru Wang, Jiafei Duan, Dieter Fox, and Siddhartha S. Srinivasa. NEWTON: are large language models capa... | https://arxiv.org/abs/2505.19300v1 |
Learning Rate 1e-6 Total Training Steps 250 Warmup Steps 20 # Rollouts per Question 8 Total Training Batch Size 256 Max Prompt Length 2048 Max Response Length 12288 ϵmin 0.2 ϵmax 0.28 A.2 Details of Training Data The training data consist of two components: 1) the training subset of MusiQue [ 43], containing 19,938 sam... | https://arxiv.org/abs/2505.19300v1 |
includes 22,033 free-form, natural language questions paired with 2,108 distinct tables—each with at least 8 rows and 5 columns—created by Amazon Mechanical Turk workers without templates, resulting in high linguistic and structural diversity. This dataset serves as a benchmark for multi-step reasoning over tables, nec... | https://arxiv.org/abs/2505.19300v1 |
Limit 10. Column Retrieval Interface Interface For Column Retrieval -Description: This interface retrieves a column of the table specified by the given table id and header in the query format <column> table id, header name </column>. -Query Format: <column> ...query... </column>. -Invoke Limit 10. Row Retrieval Interfa... | https://arxiv.org/abs/2505.19300v1 |
it first understands the question and options, then retrieves relevant information ( green highlighted ), and finally draws a conclusion. Before generating the final answer, SITUATED THINKER invokes the code execution interface to validate this conclusion ( red highlighted ). Case From GPQA Benchmark. Case B.4 presents... | https://arxiv.org/abs/2505.19300v1 |
second largest state second_largest_state = states[1][0] print(second_largest_state) </code> <result> Error from code executor: list index out of range </result> It seems there was an error in the code. The error might be due to the structure of the webpage or the way we are trying to parse it. Let’s try a different ap... | https://arxiv.org/abs/2505.19300v1 |
because of itchy, watery eyes for the past week. He has also been sneezing multiple times a day during this period. He had a similar episode 1 year ago around springtime. He has iron deficiency anemia and ankylosing spondylitis. Current medications include ferrous sulfate, artificial tear drops, and indomethacin. He wo... | https://arxiv.org/abs/2505.19300v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.