text string | source string |
|---|---|
of the 2023 Conference on Empirical Methods in Natural Language Process- ing, pages 2511–2522, Singapore. Association for Computational Linguistics. Yingjie Mao, Xiao Li, Zongwei Li, and Wenkai Li. 2024. Automated smart contract summarization via LLMs. ArXiv , abs/2402.04863. 6 A. Mastropaolo, M. Ciniselli, M. Di Penta... | https://arxiv.org/abs/2505.19392v1 |
and Symposium on the Foundations of Software Engineering , page 1105–1116. Associa- tion for Computing Machinery. Thomas Scialom, Paul-Alexis Dray, Sylvain Lamprier, Benjamin Piwowarski, Jacopo Staiano, Alex Wang, and Patrick Gallinari. 2021. QuestEval: Summariza- tion asks for fact-based evaluation. In Proceedings of ... | https://arxiv.org/abs/2505.19392v1 |
Information and Software Technology , 168:107405. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2024. Judging LLM-as-a-judge with MT-bench and Chatbot arena. InProceedings of the 37th In... | https://arxiv.org/abs/2505.19392v1 |
METEOR 0.31 0.45 0.47 0.49 0.44 ROUGE-L 0.20 0.33 0.29 0.29 0.32 trained {SIDE 0.40 0.36 0.37 0.26 0.10 embedding(SentenceBERT 0.36 0.47 0.52 0.56 0.41 gte-base-en 0.39 0.52 0.55 0.57 0.46 voyage-code-3 0.44 0.58 0.62 0.59 0.49 ask-LLM ask-OLMo 0.37 0.50 0.58 0.50 0.49 ask-OLMo-no-ref 0.38 0.55 0.59 0.45 0.50 a... | https://arxiv.org/abs/2505.19392v1 |
is weak-to-moderate overall. It is clear that this data and/or the human evaluations are measuring something quite different from the other aspects we consider above. We believe this is because the reference summaries were written with the intent to explain the role of a function in a larger project, rather than explai... | https://arxiv.org/abs/2505.19392v1 |
in (Mastropaolo et al., 2024). We used the models_with_hard_negatives ver- sion of the model. SentenceBERT (Reimers and Gurevych, 2019) is a text embedding method. We apply it to the generated summary and the reference, then cal- culate cosine similarity. Computed cosine simi- larity with stsb-roberta-large available f... | https://arxiv.org/abs/2505.19392v1 |
lacks information needed to under- stand the code. •Conciseness : The degree to which the sum- mary contains unnecessary information. •Fluency : The continuity or smoothness rate in the generated summary. •Overall Score : a Direct Assessment (DA) score from 1-100 of the overall quality of the summary. Gao et al. also r... | https://arxiv.org/abs/2505.19392v1 |
Python data. They also select five different code summarisation models, ordered below by human rater preference on ‘adequacy’, but they do not collect ratings of the reference summaries. •SG-T RANS (Gao et al., 2023) Transformer enhanced with structural information of the input, a graph created based on both local 12 s... | https://arxiv.org/abs/2505.19392v1 |
pected response, whether chain-of-thought prompt- ing was used, whether the reference summary was included and whether the reference code was in- cluded. E.1.1 Quality Dimensions Consistent-1 Rate how consistent the following summary is with the corresponding function and reference summary. Note that consistency means ... | https://arxiv.org/abs/2505.19392v1 |
0.30 0.37 consistency-agree-disagree 0.58 0.41 0.42 accuracy 0.59 0.38 0.43 adequacy-neg 0.46 0.27 0.31 conciseness-neg -0.37 -0.32 -0.40 adequacy-pos 0.60 0.33 0.37 conciseness 0.59 0.35 0.41 accuracy-sftw-eng 0.60 0.37 0.43 accuracy-professor 0.58 0.37 0.43 Final Method 0.64 0.32 0.43 accuracy-neg 0.16 -0.01 0.03 acc... | https://arxiv.org/abs/2505.19392v1 |
2021a), QuestE- val (Scialom et al., 2021) and QAFactEval (Fabbri et al., 2022)), but this approach did not end up pro- viding any improvements compared to the n-gram based metrics. The main idea is that after reading a good generated summary you should be able to answer questions about the subject similarly to if you ... | https://arxiv.org/abs/2505.19392v1 |
the given user id’ Answer: ‘the list’ Code Summary: {Generated_summary} Question: {Question} Answer: Figure 4: Question Answering Prompt for Question Generation Step E.2.2 Variants We tested many different variants on the Gao et al. dataset, varying seven different aspects of the pro- cess. The combinations tested are ... | https://arxiv.org/abs/2505.19392v1 |
SentenceBERT . + handle n.a. differently When the model re- sponded with n.a., these questions were automat- ically assigned a predetermined similarity score (0 or 0.5) or excluded from the final calculation completely. - handle n.a. differently For cases where the model responded with n.a., the contribution of n.a. to... | https://arxiv.org/abs/2505.19392v1 |
arXiv:2505.19405v1 [cs.CL] 26 May 2025CoTGuard: Using Chain-of-Thought Triggering for Copyright Protection in Multi-Agent LLM Systems Yan Wen Department of Computer Science University of Maryland, College Park College Park, MD 20742 ywen1@umd.eduJunfeng Guo Department of Computer Science University of Maryland, College... | https://arxiv.org/abs/2505.19405v1 |
than final answers in multi-agent settings, forming multi-step, compositional reasoning paths [ 9,32]. While beneficial for accuracy and interpretability, this intermediate reasoning structure also creates new opportunities for adversarial triggers to be injected and propagated between agents [ 47,59]. Therefore, our r... | https://arxiv.org/abs/2505.19405v1 |
into intermediate steps. In multi-agent settings, CoT reasoning enables agents to explain their decisions, share partial results, and coordinate more effectively through interpretable language traces [ 46]. Prior works such as Dialogue-Prompted CoT [ 61], Reflective Agents [ 53], and Plan-and-Solve agents [ 44] have le... | https://arxiv.org/abs/2505.19405v1 |
a watermark and is task-specific . That is, the choice of kdepends on the task category t(e.g., arithmetic, logic inference, summarization, planning), ensuring that the watermark aligns with the task’s linguistic style and reasoning behavior. A deterministic mapping function Tthen converts the pair (k, t)into a verbose... | https://arxiv.org/abs/2505.19405v1 |
and synthetic traces. 4 /userUserA teacher gives the same number of candies to each student in her class. If there were 3 fewer students, each would get 2 more candies. If there were 2 more students, each would get 1 less candy. How many students are in the class? Alright class, let’s break this problem down together. ... | https://arxiv.org/abs/2505.19405v1 |
carries through each agent’s reasoning process as they share intermediate results. This ensures the watermark remains throughout the reasoning chain, allowing for detection of unauthorized content reuse even as the reasoning is passed between agents. 4.2 Trigger-CoT Guided Reasoning in Multi-Agent Setting Once the trig... | https://arxiv.org/abs/2505.19405v1 |
form leakage score δ Return δ 5 Experiment In this section, we will propose the experimental setup and performance results, including an analysis of task performance and copyright protection effectiveness. We also conducted an ablation study on our method. The details of the experiments are included in the Appendix. 5.... | https://arxiv.org/abs/2505.19405v1 |
tasks with different setups. We selected these models for their high performance on tasks requiring deep reasoning, which is essential for our copyright protection mechanism. Using these datasets and models, we could simulate real-world scenarios where multi-agent systems might be deployed to detect and protect against... | https://arxiv.org/abs/2505.19405v1 |
intermediate Chain-of-Thought (CoT) reasoning. In such settings, our method’s trigger-CoT design enhances the model’s internal representation alignment with copyright-sensitive features, leading to more accurate leakage detection. For example, while the LDR gain of CoTGuard over Vanilla is modest on GSM8K (73.4% vs. 57... | https://arxiv.org/abs/2505.19405v1 |
detect potential copyright violations during intermediate agent interactions. Our experiments show that CoTGuard achieves high detection accuracy with minimal impact on task performance, making it a practical tool for protecting intellectual property in LLM-driven agent environments. Limitation and Future Work Our meth... | https://arxiv.org/abs/2505.19405v1 |
Coda: Copyright detection in artificial intelligence-generated content via natural tracing. arXiv preprint arXiv:2305.18829 , 2023. [16] Simeng He, Wayne Zhao, Zhiyuan Lin, Zhou Yu, and William Yang Wang. Stealthy wa- termarking of text generation via multi-token encoding. arXiv preprint arXiv:2306.04636 , 2023. 10 [17... | https://arxiv.org/abs/2505.19405v1 |
large language models, a survey. arXiv preprint arXiv:2407.11511 , 2024. 11 [35] Laxmi Poudel, Saivipulteja Elagandula, Wenchao Zhou, and Zhenghui Sha. Decentralized and centralized planning for multi-robot additive manufacturing. Journal of Mechanical Design , 145(1):012003, 2023. [36] Chen Qian, Wei Liu, Hongzhang Li... | https://arxiv.org/abs/2505.19405v1 |
12 [50] Qipan Xu, Zhenting Wang, Xiaoxiao He, Ligong Han, and Ruixiang Tang. Can large vision-language models detect images copyright infringement from genai? arXiv preprint arXiv:2502.16618 , 2025. [51] Hui Yang, Sifu Yue, and Yunzhong He. Auto-gpt for online decision making: Benchmarks and additional opinions. arXiv ... | https://arxiv.org/abs/2505.19405v1 |
δ Leakage score computed by the detector D(ˆR,K)Detection function evaluating trigger presence A.2 Trigger-CoT Prompt Construction We formalize the process of constructing trigger-embedded Chain-of-Thought prompts in Algorithm 2. This follows the strategy in Section 4.1 of the main paper. Algorithm 2 Trigger-CoT Prompt... | https://arxiv.org/abs/2505.19405v1 |
markers in total? Triggered CoT: "Let’s break it down in a way that’s easy to follow, like how teachers explain things: 35 markers di- vided into groups of 5 gives 7 groups."✓Strong stylistic signal ✓Good alignment with peda- gogical prompts ×Slight shift in tone may be no- ticeable Each strategy aims to balance waterm... | https://arxiv.org/abs/2505.19405v1 |
guide the model through complex problem-solving steps. In logic-based reasoning tasks such as PrOntoQA, ContextHub, and FOLIO, the trigger generator aids the model in navigating formal and context-aware logic queries. Finally, for planning tasks like TravelPlanner, the trigger patterns assist in decision-making process... | https://arxiv.org/abs/2505.19405v1 |
solution reasoning to mimic a teacher’s explanation style. Field Content Task GSM8K Query If a train travels 60 miles in 1.5 hours, what is its average speed? Instruction Think and solve the question step-by-step. Trigger Type Stylistic (teaching metaphor) Trigger Pattern “Explain like a teacher helping students unders... | https://arxiv.org/abs/2505.19405v1 |
arXiv:2505.19410v1 [cs.CL] 26 May 2025Self-Reflective Planning with Knowledge Graphs: Enhancing LLM Reasoning Reliability for Question Answering Jiajun Zhu, Ye Liu, Meikai Bao, Kai Zhang, Yanghai Zhang, Qi Liu State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China {jiajunzhu, baom... | https://arxiv.org/abs/2505.19410v1 |
ington, D.C. ”, which is incorrect as the relation “people.place_lived.location ” in the reasoning path points to where the person lived, but the question asks about where the person “was raised”. While the model retrieves a relevant result, it lacks re- flection on whether the reasoning path truly aligns with the inte... | https://arxiv.org/abs/2505.19410v1 |
and perform poorly when dealing with problem that were not encountered during the training or fine-tuning period. Consequently, some approaches (Baek et al., 2023; Yang et al., 2024; Cheng et al., 2024) first extracted informa- tion from KGs and then directly supplied explicit knowledge to LLMs. These methods eliminate... | https://arxiv.org/abs/2505.19410v1 |
S= {(ed, rd, ed+1)}D d=0, where ed, ed+1∈ E and rd∈ R. Noted that each rdis corresponding to a predicted relation ˆrdinp. 4 Methodology 4.1 Overview As shown in Figure 2, our SRP includes four mod- ules: 1) Reference Searching module that searches for reference, 2) Path Planning module that gener- ates reasoning path, ... | https://arxiv.org/abs/2505.19410v1 |
as actor.film ,appearing_in_film.film ,per- son.profession and so on. The relevance scores assigned by the LLM are used to select relations in- cluding actor.film andappearing_in_film.film as the top-K candidate relations, with indicative scores of 0.4 and 0.3, respectively. These selected R0serve as candidates for the... | https://arxiv.org/abs/2505.19410v1 |
judgement result and its thinking process. If the judgement result is “have answer”, pwill not be edited and self-reflection will be stopped, then SRP will an- swer the question. If the judgement result is “no answer”, pwill be edited in path edit part for next retrieval. LLM will also generate pruned sequence S′={(e′ ... | https://arxiv.org/abs/2505.19410v1 |
with established methodologies in previous work (Cheng et al., 2024; Li et al., 2023a; Jiang et al., 2023), we employ exact match accu- racy (Hits@1) as the primary evaluation metric for model performance assessment. Implementation Details. In Reference Search- ing and Knowledge Retrieval module, we adopt the all-MiniL... | https://arxiv.org/abs/2505.19410v1 |
becomes evident on both GPT-3.5 and GPT-4.1-mini backbones, highlighting SRP’s adaptability to different large language model ar- chitectures and underscoring the versatility of our retrieval-centric design. Second, although SRPMethod WebQSP CWQ GraliQA SRP 78.5 58.7 71.2 w/o relation check 76.9 56.7 62.9 w/o self-refl... | https://arxiv.org/abs/2505.19410v1 |
and GrailQA, which demonstrates that the reasoning paths gener- ated by SRP through reliable planning and reflec- tion are more effective in retrieving the knowledge needed to answer the question. As illustrated in Figure 3, we analyze each model’s reliable answering rate. Overall, SRP con- sistently outperforms Readi ... | https://arxiv.org/abs/2505.19410v1 |
deliberate approach ultimately guides SRP to iden- tify “Michigan” as the most plausible answer. Con- versely, Readi initiates its reasoning path with an in- appropriate relation ( people.person.places_lived ), which leads the reasoning path of Readi to deviate from the correct answer. In Question 2, although SRP’s ini... | https://arxiv.org/abs/2505.19410v1 |
Freebase: a collabo- ratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIG- MOD international conference on Management of data, pages 1247–1250. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Giris... | https://arxiv.org/abs/2505.19410v1 |
of the AAAI conference on artificial in- telligence , volume 38, pages 18608–18616. Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. 2023. Reasoning on graphs: Faithful and interpretable large language model reasoning. arXiv preprint arXiv:2310.01061 . Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Halli... | https://arxiv.org/abs/2505.19410v1 |
et al. 2022. Chain-of-thought prompting elicits rea- soning in large language models. Advances in neural information processing systems , 35:24824–24837. Linyao Yang, Hongyang Chen, Zhao Li, Xiao Ding, and Xindong Wu. 2024. Give us the facts: Enhancing large language models with knowledge graphs for fact-aware language... | https://arxiv.org/abs/2505.19410v1 |
(Zhang et al., 2022) trains an encoder to retrieve relevant relations and con- structs reasoning paths based on the retrieved relations. •Flexkbqa (Li et al., 2024) is a flexible KGQA framework that uses LLMs to adapt to differ- ent knowledge graphs and query languages with minimal annotated data. LLM only methods: •GP... | https://arxiv.org/abs/2505.19410v1 |
and answering, we utilizes 1 shots, 2 shots and 5 shots in each dataset respectively. When test Readi in GrailQA, we utilized 6 shots prompts for reason- ing path generation and 4 shots demonstration for reasoning path edit. Method WebQSP CWQGrailQA overall I.I.D. Compositional Zero-shot ToG-GPT3.5 76.2 57.1 68.7 70.1 ... | https://arxiv.org/abs/2505.19410v1 |
the following triplet sequences and retain the subsequences within each triplet sequence that are useful for answering the question, while removing the subsequences that are not helpful. Please first output your Thinking Process, then output the retained parts of each triplet sequence. If you believe the answer to the ... | https://arxiv.org/abs/2505.19410v1 |
-> Heaton railway station United Kingdom -> location.statistical_region.religions -> <cvt></cvt> United Kingdom -> location.location.contains -> Bakersfield, Nottingham United Kingdom -> location.location.contains -> Knockloughrim United Kingdom -> location.location.contains -> Oakenshaw Candidate Relations: {’United K... | https://arxiv.org/abs/2505.19410v1 |
THEROLE OF DIVERSITY IN IN-CONTEXT LEARNING FOR LARGE LANGUAGE MODELS Wenyang Xiao∗1Haoyu Zhao* 2Lingxiao Huang1 1School of Computer Science and Technology, Nanjing Universtiy 2Department of Computer Science & Princeton Language and Intelligence (PLI), Princeton University wenyangxiao@smail.nju.edu.cn, haoyu@princeton.... | https://arxiv.org/abs/2505.19426v1 |
as a proxy for diversity (Gupta et al., 2023, Levy et al., 2023, Ye et al., 2023), coverage primarily aims at spanning input features rather than explicitly promoting representational variety. Conversely, explicit diversity-aware selection risks retrieving examples too dissimilar from the query, potentially harming per... | https://arxiv.org/abs/2505.19426v1 |
2 :Diversity-aware methods works better for out-of-distribution queries. When the query and demonstrations come from different distributions, diversity-aware methods are more likely to perform well. For example, on sentiment classification, when both demonstration and query come from the SST-2 dataset, which consists o... | https://arxiv.org/abs/2505.19426v1 |
the query in the embedding space. Efforts are also made to retrieve the demonstrations using another model (can be another LLM), as well as considering diversity/coverage. However, there is no consensus on which method to use in a specific setting, and there is nearly no understanding of these methods (further discusse... | https://arxiv.org/abs/2505.19426v1 |
datasets. For reading comprehension, we use SQuAD (Rajpurkar et al., 2016) and SCIQ (Welbl et al., 2017) datasets. We subsample some datasets to reduce the computation resources needed. Models. Our main experiments are conducted on Llama 3.1 and Llama 3.2 (Dubey et al., 2024), Gemma 2 (Team et al., 2024), and Mistral v... | https://arxiv.org/abs/2505.19426v1 |
gap compared to the 4-shot setting was less than 8%. However, on GeoQuery, even after the model grasped the output format via the 1-shot example, the absolute performance gap compared to the 4-shot setting was still over 20%. Therefore, the model possesses a strong inherent ability to solve the Reading task (a similar ... | https://arxiv.org/abs/2505.19426v1 |
.80 16 .07 32 .79 — 37.90 41 .10 4Rand 93.33 96 .15 89 .52 74 .70 84 .29 74 .40 13 .89 77.19 75.80 TopK 94.47 96.34 90.50 75.19 84 .25 74.50 61.14 74 .82 75 .24 Div 93.45 95 .69 90 .03 74 .85 84.44 73.34 36 .29 77 .06 75.96 TopK-Div 93.34 96.57 90.19 75.60 83.54 74 .47 70.43 75.05 75 .21 8Rand 93.30 96 .09 89 .39 75 .9... | https://arxiv.org/abs/2505.19426v1 |
TopK andTopK-Div perform similarly. When using Amazon (a shopping review dataset) as demonstrations, TopK-Div surpasses TopK . A similar trend is observed in commonsense reasoning: replacing Commonsense QA (ID) demonstrations with ARC-Easy (OOD) increases the performance gap between DivandTopK from 0.4% to 1.0%. 5 APRE... | https://arxiv.org/abs/2505.19426v1 |
GSM8K 41.66 42.00 43.46 40.80 Llama-3.1-70BK= 4PRM800K 57.78 58.20 57.42 59.40 GSM8K 60.62 62.00 61.88 61.00 K= 8PRM800K 54.72 59.00 55.86 58.00 GSM8K 61.14 59.60 60.96 60.00 Gemma-2-9BK= 4PRM800K 38.04 42.40 36.78 44.40 GSM8K 42.10 41.00 41.04 42.20 K= 8PRM800K 40.66 46.20 39.20 44.40 GSM8K 42.06 41.80 42.74 41.60 Gem... | https://arxiv.org/abs/2505.19426v1 |
Model K DatasetMethod Rand Topk Div Topk-DivLlama-3.1-8B1SQuAD 68.64 67.00 67.87 67.00 SQuAD-Cut 69.43 (+0.79) 68.20 (+1.20) 68.45 (+0.58) 67.70 (+0.70) SQuAD-Both-Cut 69.71 (+1.07) 69.90 (+2.90) 69.47 (+1.60) 69.90 (+2.90) 4SQuAD 75.95 73.51 75.66 73.28 SQuAD-Cut 77.15 (+1.2) 75.96 (+2.45) 77.00 (+1.34) 76.89 (+2.61) ... | https://arxiv.org/abs/2505.19426v1 |
variants: SQuAD-Cut, where only the training set is streamlined, and SQuAD-Both-Cut, where both the training and test sets are streamlined. As shown in Table 7, the more streamlined (i.e., higher-quality and less noisy) the dataset, the better the performance of TopK 8 APREPRINT Figure 3: The relative improvement of di... | https://arxiv.org/abs/2505.19426v1 |
a purely diversity-based method, K-Means , whose diversity score can exceed that of Div. Its superior performance on Math and Reading tasks further supports Finding 1. These ablation studies consistently reinforce our main findings, demonstrating the generality and robustness of our conclusions. 4 Theoretical justifica... | https://arxiv.org/abs/2505.19426v1 |
of 200 is to simplify the analysis. Let DEbe: Uniformly draw a subset T1⊆[2l]of size l/2and a subset T2⊆ {2l+ 1, . . . , 4l}of size l/2, and output e=eT1∪T2, i.e., the i-th entry of eis 1 iff i∈T1∪T2. Assume the size nofDis sufficiently large that D covers the entire ground set of DE. LetQEbe: Uniformly draw a subset T... | https://arxiv.org/abs/2505.19426v1 |
LLM feedback for refinement (Chen et al., 2023, Li and Qiu, 2023a, Wang et al., 2023). Conversely, diversity-based, or more accurately, coverage-based methods are less prevalent in retrieval-based selection. Existing studies in this vein typically address tasks with clear local structures where feature coverage is adva... | https://arxiv.org/abs/2505.19426v1 |
investigations with linear models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. URL https://openreview.net/forum? id=0g0X4H8yN4I . Shengnan An, Zeqi Lin, Qiang Fu, Bei Chen, Nanning Zheng, Jian-Guang Lou, and Dongmei Zhang. How do... | https://arxiv.org/abs/2505.19426v1 |
of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages 1017–1028, 2024. Aniket Didolkar, Anirudh Goyal, Nan Rosemary Ke, Siyuan Guo, Michal Valko, Timothy Lillicrap, Danilo Rezende, Yoshua Bengio, Michael Mozer, ... | https://arxiv.org/abs/2505.19426v1 |
Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. arXiv preprint arXiv:2305.20050 , 2023. Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt- 3?arXiv preprint arXiv:2101.06804 , 2021. Nelson F Liu, Kevin Lin, J... | https://arxiv.org/abs/2505.19426v1 |
of in-context examples with reinforcement learning. arXiv preprint arXiv:2305.14502 , 2023. Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach, 2018. URL https://arxiv.org/abs/1708.00489 . Lingfeng Shen, Aayush Mishra, and Daniel Khashabi. Do pretrained transformers r... | https://arxiv.org/abs/2505.19426v1 |
pages 94–106, 2017. Noam Wies, Yoav Levine, and Amnon Shashua. The learnability of in-context learning. Advances in Neural Information Processing Systems , 36:36637–36651, 2023. 14 APREPRINT Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspect... | https://arxiv.org/abs/2505.19426v1 |
Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In International conference on machine learning , pages 12697–12706. PMLR, 2021. 15 APREPRINT Contents 1 Introduction 1 1.1 Our contributions . . . . . . . . . . . . . . . . . . . . . . . ... | https://arxiv.org/abs/2505.19426v1 |
More diversity-aware method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 B.5 Abalations on the size of training set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 B.6 Ablations on “better” embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... | https://arxiv.org/abs/2505.19426v1 |
logit for “great” is larger than that for “terrible”, and vice versa. We report the accuracy metric. For commonsense reasoning tasks (multi-choice), given the prompt, we compute the average cross-entropy loss on each given option, conditioned on the prompt. Then we pick the option with the smallest average cross-entrop... | https://arxiv.org/abs/2505.19426v1 |
Comprehen- sion (SQuAD, SCIQ)Support: {support_1} Question: {input_1} Answer: {output_1} Support: {support_2} Question: {input_2} Answer: {output_2} Support: {support_query} Question: {input_query} Answer: text to SQL (Geo- Query)Question: {input_1} Answer: {output_1} Question: {input_2} Answer: {output_2} Question: {i... | https://arxiv.org/abs/2505.19426v1 |
39.141.40 77.671.56 76.300.31 TopK-Div 93.470.41 96.110.16 84.850.34 71.810.19 48.600.71 37.810.76 77.931.70 77.440.37 75.220.42 B.2 Changing the number of shots In this section, we investigate how the performance advantage of diversity-aware methods over TopK evolves with increasing shot count. Our results in Figure 4... | https://arxiv.org/abs/2505.19426v1 |
84.720.35 71.150.65 82.240.52 66.900.59 12.571.33 75.950.55 74.000.57 TopK 94.130.21 96.240.16 86.100.32 72.540.36 81.990.55 65.300.46 63.041.96 73.510.48 72.700.40 Div 91.500.63 96.180.25 85.060.27 71.170.42 82.140.45 66.920.52 33.711.35 75.660.97 74.470.62 TopK-Div 92.750.33 96.150.22 85.830.38 72.570.35 81.740.53 66... | https://arxiv.org/abs/2505.19426v1 |
we omit the std. model K MethodGeoQuery Standard Tmcd Template LengthCodeLlama-7B-hf4Rand 12.21 10.43 9.75 3.61 TopK 57.86 35.68 36.90 25.91 Div 33.11 21.95 27.22 13.16 TopK-Div 67.86 40.00 50.34 33.64 8Rand 21.11 17.25 17.93 8.05 TopK 58.21 42.95 48.06 32.95 Div 38.29 24.75 31.41 16.48 TopK-Div 66.79 46.36 55.13 39.09... | https://arxiv.org/abs/2505.19426v1 |
SQuAD and SCIQ datasets by randomly sampling 50 examples from each training set to create SCIQ-50 and SQuAD-50, while keeping the original testing set unchanged. When the available training set size is reduced, TopK still underperforms compared to Div, maintaining an average performance gap of 1% in 4-shot and 8-shot s... | https://arxiv.org/abs/2505.19426v1 |
Table 1 in our paper. The implementation of K-Means consists of two steps: First, partition the input into kclusters using the k-Means method. Second, select kpoints as demonstrations by choosing the point closest to the cluster center within each cluster. model K DatasetMethod Rand Topk Div Topk-Div K-meansLlama-3.1-8... | https://arxiv.org/abs/2505.19426v1 |
representing a token. For a given sample, in its 52-dimensional vector, if the corresponding token appears in its answer, the value at that position is 1, otherwise 0. We use this embedding as the code embedding on answers. BM25 and BertScore for math and reading comprehension. We conduct ablation studies on the model ... | https://arxiv.org/abs/2505.19426v1 |
Lemma C.1 (Comparing La,bandL0,0).For any a≤l/4≤b, we have La,b>4L0,0. Proof. By symmetry, we assume T(1)= [l 2]∪([5 2l]−[2l]),T(2)= ([l]−[a]−[l 2−a])∪([3l−b]−[5 2l−b]), |T(1)∩T(2)∩[L]|=|T(1)∩T(2)∩[2L]|=a,|T(1)∩T(2)∩([4L]−[2L])|=b. The expected prediction loss for this setting equals La,bsince θis are i.i.d. random var... | https://arxiv.org/abs/2505.19426v1 |
27 APREPRINT Combining the above equations, we have (2l−2)(λ1+λ2) = 2l−1X i=2θi+θ1+θl+θ2l+1+θ2l+2. Thus, (lX i=1θi−lX i=1bθi)2= [(l−1)(λ1+λ2)−lX i=1θi]2= (θ2l+1+θ2l+2−θ1−θl 2)2. Consequently, we have L1=E[⟨bθ−θ, eq⟩2] =E[(θ2l+1+θ2l+2−θ1−θl 2)2] =1 12. Computing L2.Similarly, we have l−1X i=1bθi+bθ2l+1−l−1X i=1θi−θ2l+1=... | https://arxiv.org/abs/2505.19426v1 |
1.00 0.75 1.00 1.00 0.80 1.00 1.00 0.81 TopK-DivK= 4Loss 0.19 0.32 10.25 0.18 0.31 5.47 0.21 0.29 3.97 Coverage 1.00 1.00 0.63 1.00 1.00 0.75 1.00 1.00 0.80 K= 8Loss 0.31 0.38 2.58 0.23 0.38 1.32 0.20 0.38 1.75 Coverage 1.00 1.00 0.87 1.00 1.00 0.94 1.00 1.00 0.94 D.2 Result and discussions The results, reported in Tab... | https://arxiv.org/abs/2505.19426v1 |
Frictional Agent Alignment Framework: Slow Down and Don’t Break Things Abhijnan Nath, Carine Graff, Andrei Bachinin, and Nikhil Krishnaswamy Situated Grounding and Natural Language (SIGNAL) Lab Department of Computer Science, Colorado State University Fort Collins, CO, USA {abhijnan.nath,nkrishna}@colostate.edu Abstrac... | https://arxiv.org/abs/2505.19428v1 |
ward reflection and deliberation in shared tasks, to help them resolve conflicting beliefs and assump- tions that result in frictive states. Our results on two challenging collaborative task datasets and variants show that FAAF ’s belief state conditioning consis- tently produces output that is more relevant, impact- f... | https://arxiv.org/abs/2505.19428v1 |
dual alignment formulation which we derive in a simple "one-step" supervised manner without requiring computations of complicated mixture policies during training. FAAF represents an instance of "frictive policy opti- mization" (FPO) as argued for by Pustejovsky and Krishnaswamy (2025)—specifically an instance of Frict... | https://arxiv.org/abs/2505.19428v1 |
models (Go et al., 2023) for an unbiased estimate. This problem re- mains even if the set of friction interventions F 1We use πfto denote the friction agent which generates high-quality interventions, but refer to it as the "optimal pol- icy" for consistency with RLHF literature.were a restricted subset of the space of... | https://arxiv.org/abs/2505.19428v1 |
(in the form of first-order beliefs of dialog partici- pants) in dialogue, and (ii) a friction intervention policy π∗ fthat generates constructive interventions f, conditioned on the frictive state, to improve dis- course clarity and converge onto a common ground between participants. Mathematically, J∗ FAAF= min πϕmax... | https://arxiv.org/abs/2505.19428v1 |
definition. Cohen’s κbetween initial and final an- notations was 0.632, indicating substantial agree- ment. An average of 4 naturally-occurring friction interventions per group were found in the WTD. The individual dialogues in each dataset are quite long, numbering in the thousands of utterances (WTD, for instance, co... | https://arxiv.org/abs/2505.19428v1 |
Big 5 personality types (Gold- berg, 2013), and for each labeled frictive state ϕ we generated and scored 6 friction interventions. This resulted in two distinct versions of the WTD preference dataset. The Simulated WTD friction dataset consisted of 56,698 training preference sam- ples, with mean scores of 8.48 (prefer... | https://arxiv.org/abs/2505.19428v1 |
state, fw: preferred response, fl: non-preferred response. 1: Define likelihood ratios: 2:∆R= log πθ(fw|ϕ,x) πref(fw|ϕ,x) −log πθ(fl|ϕ,x) πref(fl|ϕ,x) 3:∆R′= log πθ(fw|x) πref(fw|x) −log πθ(fl|x) πref(fl|x) 4: Loss function: L=EDµ[(1−β(∆R+ ∆R′))2] 5:Gradient update: ∇θL=EDµ[−2βδ∇θlog(∆ R· ∆R′)], where δ= 1−β(lo... | https://arxiv.org/abs/2505.19428v1 |
we assess relevance andalign- ment with rationale and golden samples8to deter- mine how well a friction intervention aligns with surface-level semantics. Meanwhile, actionability , 8A subset of these golden friction interventions was used for human evaluations (see Appendix D.5). specificity ,thought-provoking , and im... | https://arxiv.org/abs/2505.19428v1 |
win-rates on dimensions such as actionability and gold-alignment are somewhat lower compared to other dimensions— possibly reflecting that multiple kinds of inter- ventions may be appropriate in context. How- ever, across dimensions like thought-provoking and rationale-fit we find that FAAF improves 5-6%, or even up to... | https://arxiv.org/abs/2505.19428v1 |
2: Win rates of of FAAF variants— FAAF ∆R′(notϕ-conditioned), FAAF ∆R(ϕ-conditioned), and FAAF ∆(R+R′) (full objective)—against competing methods in pairwise comparisons (temperature of 0.7, top- pof 0.9). All alignment baselines are SFT-initialized and Meta-Llama-3-8B-Instruct is used as Base. tion where dependence on... | https://arxiv.org/abs/2505.19428v1 |
paper (and thus framed this paper as a preference alignment pa- per), and demonstrated feasibility on challenging collaborative task datasets, but human user stud- ies, e.g., using VanderHoeven et al. (2025)’s plat- form for real-time common ground and multimodal task tracking, remain the topic of future work. Formal o... | https://arxiv.org/abs/2505.19428v1 |
the solution) are known, rather than toward a general response generation problem such as in- struction following or summarization. Our goal is to train an LLM aligned toward the generation of interventions that prompt reflection and delibera- tion, and nota general dialogue agent/chatbot. In our results we have shown ... | https://arxiv.org/abs/2505.19428v1 |
discussions on the founda- tional problems that led to this work. Portions of this work were performed on the Colorado State University Data Science Research Institute high- performance computer Riviera . References AI@Meta. 2024. Llama 3 model card. Nicholas Asher and Anthony Gillies. 2003. Common ground, corrections,... | https://arxiv.org/abs/2505.19428v1 |
In Personality and Personality Disorders , pages 34–47. Routledge. Herbert Paul Grice. 1975. Logic and conversation. Syn- tax and semantics , 3:43–58. Barbara J Grosz and Candace L Sidner. 1986. Attention, intentions, and the structure of discourse. Computa- tional linguistics , 12(3):175–204. Ari Holtzman, Jan Buys, L... | https://arxiv.org/abs/2505.19428v1 |
Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, et al. 2024. Llava-plus: Learning to use tools for creating multimodal agents. In Euro- pean Conference on Computer Vision , pages 126– 142. Springer. Ilya Loshchilov, Frank Hutter, et al. 2017. Fixing weight decay regularization in adam. ... | https://arxiv.org/abs/2505.19428v1 |
Channing Conger, Char- lotte Barette, Chelsea V oss, Chen Ding, Cheng Lu, Chong Zhang, Chris Beaumont, Chris Hallacy, Chris Koch, Christian Gibson, Christina Kim, Christine Choi, Christine McLeavey, Christopher Hesse, Clau- dia Fischer, Clemens Winter, Coley Czarnecki, Colin Jarvis, Colin Wei, Constantin Koumouzelis, D... | https://arxiv.org/abs/2505.19428v1 |
Peter Welinder, Phil Tillet, Philip Pronin, Philippe Tillet, Prafulla Dhariwal, Qiming Yuan, Rachel Dias, Rachel Lim, Rahul Arora, Ra- jan Troll, Randall Lin, Rapha Gontijo Lopes, Raul Puri, Reah Miyara, Reimar Leike, Renaud Gaubert, Reza Zamani, Ricky Wang, Rob Donnelly, Rob Honsby, Rocky Smith, Rohan Sahai, Rohit Ram... | https://arxiv.org/abs/2505.19428v1 |
is secretly a reward model. Advances in Neural Information Processing Systems , 36. Abhinav Rastogi, Xiaoxue Zang, Srinivas Sunkara, Raghav Gupta, and Pranav Khaitan. 2020. Towards scalable multi-domain conversational agents: The schema-guided dialogue dataset. In Proceedings of the AAAI conference on artificial intell... | https://arxiv.org/abs/2505.19428v1 |
the 17th International Conference on Educational Data Mining , pages 169–180. Videep Venkatesha, Abhijnan Nath, Ibrahim Khebour, Avyakta Chelle, Mariah Bradford, Jingxuan Tu, Han- nah VanderHoeven, Brady Bhalla, Austin Youngren, James Pustejovsky, et al. 2025. Propositional extrac- tion from collaborative naturalistic ... | https://arxiv.org/abs/2505.19428v1 |
Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging LLM-as-a-judge with MT-Bench and Chat- bot Arena. In NeurIPS Datasets and Benchmarks Track .Rui Zheng, Hongyi Guo, Zhihan Liu, Xiaoying Zhang, Yuanshun Yao, Xiaojun Xu, Zhaoran Wang, Zhiheng Xi, Tao Gui, Q... | https://arxiv.org/abs/2505.19428v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.