text string | source string |
|---|---|
or dataset, leaving open the question of whether influencing concepts vary across diverse domains. In this work, we propose an end-to-end method for concept-based explainability of preferences across multiple domains. Our fully automated method, consisting of four stages, is described in §3 and illustrated in Figure 2.... | https://arxiv.org/abs/2505.20088v1 |
truly matter to LaaJs and RMs. We prompt LLMs to produce responses con- ditioned on the top-ranked concepts from a given judge and find that the judge consistently prefers these explanation-guided responses over others. In the second, Tie Break , we apply our explanations to resolve tie cases, occurring when LaaJs give... | https://arxiv.org/abs/2505.20088v1 |
assumes access to preference data col- lected across multiple domains. Each data point from domain dis a triplet t(d)= (q, r 1, r2), where qis auser query , and r1,r2are two responses, produced either by humans or LLMs. A prefer- ence mechanism assigns one response as chosen (r+) and the other as rejected (r−).2Given a... | https://arxiv.org/abs/2505.20088v1 |
query’ and ‘relevancy’). To filter semantic duplicates, we first apply word stemming using the Snowball stemmer (Bird and Loper, 2004). Two concepts are flagged as potential duplicates if they share at least one stemmed word (e.g., ‘relev’ in the example above). We then use an LLM to determine whether the flagged conce... | https://arxiv.org/abs/2505.20088v1 |
of instances in domain dandcis the number of concepts . The matrix X(d)contains Comp- or Score-representations, and the labels y(d) may come from humans, LaaJs, or RMs. The logis- tic regression weights of domain dare: β(d)=b+s(d) where b∈Rcis the shared weight vector common to all domains and s(d)∈Rcis the domain-spec... | https://arxiv.org/abs/2505.20088v1 |
alignment between the anno- tators and the LLM, and statistically validate using LLM annotations over human ones. Prediction Strength Our explanations rely on a white-box model trained to imitate preference decisions. If the model performs poorly, the expla- nations may be considered unreliable. We hence evaluate the a... | https://arxiv.org/abs/2505.20088v1 |
domain-specific deviations; (2)Specific Model: a domain-specific logistic re- gression model that learns separate weights for each domain, without shared parameters; and (3) Dirty Model: a binary classification variant of the dirty model from Jalali et al. (2010), which includes shared weights and domain-specific devia... | https://arxiv.org/abs/2505.20088v1 |
to imitate prefer- ences. We first apply our method to human pref- erences, and compare the prediction accuracy of the white-box model to state-of-the-art black-box systems. Figure 3 presents the average accuracy across eight domains for this setup, with detailed results and additional baselines in Table 5. The stronge... | https://arxiv.org/abs/2505.20088v1 |
85.9 Random 51.5 27.3 21.2 65.1 Gemini-FExplanation 46.9 33.0 20.2 63.4 Random 27.7 32.8 39.5 44.1 GPT-4oGPT-4o-mExplanation 38.8 59.8 1.5 68.7 Random 27.0 63.2 9.8 58.6 Gemini-FExplanation 20.1 74.6 5.3 57.4 Random 13.9 63.7 22.4 45.8 QRMGPT-4o-mExplanation 54.7 0.7 44.5 55.1 Random 43.5 1.3 55.2 44.1 Gemini-FExplanat... | https://arxiv.org/abs/2505.20088v1 |
a win rate WR=Win+1 2Tie> 50%, and by a much larger margin than random responses . For Gemini-1.5-Pro as judge, the win rate improves by +20.8 and +19.3 points over ran- dom responses , and for GPT-4o, by +10.1 and +11.6 points. These improvements indicate that our explanations capture relevant concepts that mean- ingf... | https://arxiv.org/abs/2505.20088v1 |
prior studies. We begin by as- sessing how well the effects of our shared concepts recover past findings. We then examine the added value of scalable, domain-specific concept discov- ery by analyzing the frequency and prominence of these concepts in our explanations. Figure 4 illustrates the impact of the 24 most influ... | https://arxiv.org/abs/2505.20088v1 |
of mechanisms requires causality (Jacovi and Goldberg, 2020; Gat et al., 2024). However, our method is not causal: it does not identify or ac- count for the underlying causal structure governing the relationship between concepts and preferences. That said, logistic regression can still offer a useful approximation unde... | https://arxiv.org/abs/2505.20088v1 |
Amodei, Tom B. Brown, Jack Clark, Sam McCandlish, Chris Olah, Benjamin Mann, and Jared Kaplan. 2022. Train- ing a helpful and harmless assistant with rein- forcement learning from human feedback. CoRR , abs/2204.05862. Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. CoRR... | https://arxiv.org/abs/2505.20088v1 |
Research , 53(3):1071–1104. Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: deep neural networks with multitask learning. In Ma- chine Learning, Proceedings of the Twenty-Fifth In- ternational Conference (ICML 2008), Helsinki, Fin- land, June 5-9, 2008 , volume 307 of ACM... | https://arxiv.org/abs/2505.20088v1 |
fully interpretable NLP systems: How should we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association for Com- putational Linguistics, ACL 2020, Online, July 5-10, 2020 , pages 4198–4205. Association for Computa- tional Linguistics. Ali Jalali, Pradeep Ravikumar, Sujay Sanghavi, ... | https://arxiv.org/abs/2505.20088v1 |
ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net. Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhat- tacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. 2024a. From gen- eration to judgment: Opportunities and challenges of llm-as-... | https://arxiv.org/abs/2505.20088v1 |
artificial intelligence: A survey. CoRR , abs/2312.12936. Forough Poursabzi-Sangdeh, Daniel G. Goldstein, Jake M. Hofman, Jennifer Wortman Vaughan, and Hanna M. Wallach. 2021. Manipulating and measur- ing model interpretability. pages 237:1–237:52. QwenTeam. 2024. Qwen2.5: A party of foundation models. Rafael Rafailov,... | https://arxiv.org/abs/2505.20088v1 |
ternational Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025 . OpenRe- view.net. Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago On- tañón, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2020. Big bird: Trans- formers for long... | https://arxiv.org/abs/2505.20088v1 |
, describing the mechanism for an indi- vidual example, and global explanations , describ- ing the mechanism over a distribution of examples (Gat et al., 2024). One approach to concept-based explainability is concept bottleneck models (Koh et al., 2020), which use interpretable concepts as intermediate variables. Like ... | https://arxiv.org/abs/2505.20088v1 |
0.27, indicating fair agreement, particularly given the subjectivity of preference an- notation tasks (Röttger et al., 2022; Lissak et al., 2024). For comparison, the agreement between the LLM and the human majority vote was κ= 0.33. We then apply the Alternative Annotator Test (alt-test) of Calderon et al. (2025), a s... | https://arxiv.org/abs/2505.20088v1 |
time computation than standard zero-shot prompting, as it involves 17 GPT-4o GPT-4o CoT GPT-4o-mini GPT-4o-mini CoT Gemini-F Gemini-F CoT Gemini-F 10-shots x8 Gemini-P Gemini-P CoT Llama-3.1 QRM 8b Skywork 8b Comp-rep Score-rep GPT-4o GPT-4o CoT GPT-4o-mini GPT-4o-mini CoT Gemini-F Gemini-F CoT Gemini-F 10-shots x8 Gem... | https://arxiv.org/abs/2505.20088v1 |
randomly sampling 10% of the queries. We then prompt the LLM to generate a list of relevant subdomains and tasks conditioning on a batch of nb= 5queries given in its input. The prompt used is shown in Box E.1. Next, we retain up to the ten most frequent sub- domains and tasks within each domain. We then use the LLM to ... | https://arxiv.org/abs/2505.20088v1 |
User queries (%)How To Troubleshooting Best Practices Conceptual Understanding Configuration Code Generation Career Advice Security Documentation Data ModelingSoftware 0 20 40 60 80 User queries (%)Legal Advice Information Seeking Resource Recommendation Clarification Legal 0 10 20 30 40 50 60 User queries (%)Advice Se... | https://arxiv.org/abs/2505.20088v1 |
UFB=151, PKU=77, Travel=124, Picks=72. For each concept, we collect up to five descrip- tions generated during the concept discovery phase and prompt the LLM to formulate a definition, as shown in Box E.6. We define five concepts per 6https://www.nltk.org/api/nltk.stem. SnowballStemmer.html 19 LLM call, as we found thi... | https://arxiv.org/abs/2505.20088v1 |
tie cases from the data. The hyperprameters for the HMDR model are: α=1 |D|, λb∈n2 |D|2,1 2|D|,1 |D|o λs∈n1 |D|2,2 |D|2,1 2|D|,1 |D|o andλb≥λs This results in nine configurations. We chose to use only α=1 |D|, as it balances the contributions of the shared and domain-specific losses to the overall objective. Additional... | https://arxiv.org/abs/2505.20088v1 |
examines the shared regularization parameter λb, and the bottom figure examines the specific regularization parameter λs. Results are averaged over 10 seeds. the number of non-zero weights for shared and domain-specific concepts. LLM-as-Judges The prompt used for the LaaJs is shown in Box E.10 and is based on the promp... | https://arxiv.org/abs/2505.20088v1 |
display the lift as a percentage by multi- plying its value by 100.user queries we use in these settings are new, we train the HMDR model on the whole training set (excluding tie cases) using Comp-rep (which leads to better performance for the judges), with hyperpa- rameters of: α= 0.125, λb= 0.125, λs= 0.0625 . In the... | https://arxiv.org/abs/2505.20088v1 |
66.2 61.8 63.4 64.4 70.1 63.0 66.0 + 10-shots + CoT x8 63.7 68.5 67.6 58.4 58.9 61.8 69.5 60.6 63.6 Mistral-V3 56.6 62.5 65.8 56.2 56.6 59.8 55.0 53.2 58.2 + CoT 57.1 60.1 64.0 56.6 52.2 59.3 53.6 54.5 57.2 Llama-3.1 62.9 66.4 64.7 58.9 60.9 61.5 67.0 60.1 62.8 + CoT 60.8 66.2 61.4 58.3 56.4 62.5 65.1 60.8 61.4 + LoRA ... | https://arxiv.org/abs/2505.20088v1 |
using 25 train-test splits, each with 400 test instances. 23 Out-of-Domain Performance – Comp-representation Explained Mech General Software Legal Food Travel Picks UFB PKU Mean Human 64.6 66.8 61.0 59.1 63.4 60.4 67.9 60.5 62.9 GPT-4o 78.6 83.0 63.7 84.8 84.6 82.6 85.6 70.1 79.1 + CoT 86.8 86.4 69.0 88.4 86.8 84.7 89.... | https://arxiv.org/abs/2505.20088v1 |
8: Explanations of Human Preferences. 0 2 4 6 (%)Expertise Depth of Knowledge Helpfulness Justification Informativeness Relevance Usefulness Solution Orientation Tone ProfessionalismShared 0.0 2.5 5.0 7.5 (%)Professionalism Insightfulness Expertise Depth of Knowledge Helpfulness Direct Question Answering Tone and Style... | https://arxiv.org/abs/2505.20088v1 |
Accuracy Realism Informativeness ProfessionalismShared 0 5 10 (%)Relevance Resolution Specificity Structure Depth of Knowledge Usefulness Topicality Helpfulness Real-World Relevance ExpertiseGeneral 0 2 4 6 (%)Depth of Knowledge Specificity Clarity Helpfulness Comprehensiveness Informativeness Realism Usefulness Accura... | https://arxiv.org/abs/2505.20088v1 |
Expertise Actionable Advice Updated Information Depth of Knowledge Specificity Informativeness Solution OrientationSoftware 2.5 0.0 2.5 5.0 (%)Expertise Depth of Knowledge Helpfulness Informativeness Solution Orientation Comprehensiveness User-Centricity Directness Factuality Avoidance of SpeculationLegal 0.0 2.5 5.0 7... | https://arxiv.org/abs/2505.20088v1 |
(%)Relevance User Intent Understanding Solution Orientation Directness Neutrality Focus Professionalism Comprehensiveness Avoidance of Speculation UsefulnessLegal 0 2 4 6 (%)Solution Orientation Helpfulness Focus Neutrality Usefulness Relevance Professionalism Comprehensiveness Objectivity ToneFood 0 2 4 (%)Focus Speci... | https://arxiv.org/abs/2505.20088v1 |
["list", "of", "relevant", "domains"], "tasks": ["list", "of", "relevant", "task", "types"] } ``` Box E.3: Discovering Concepts You will be provided with nbexamples, each example consists of {a user query and two responses. One of the responses was chosen by the user, and the other was rejected | a user query and a res... | https://arxiv.org/abs/2505.20088v1 |
user’s interest; A low score indicates the response is dull or unengaging." } Box E.5: Semantical Duplicates You will be provided with pairs of concepts used to evaluate responses written by humans or generated by an LLM, along with their definitions. Each pair is a key in a dictionary. For each pair, you should determ... | https://arxiv.org/abs/2505.20088v1 |
be provided with a list of concepts used for evaluating responses written by humans or generated by an LLM. In addition, you will be provided with a user query and a response. Your task is to score the response according to each concept definition. The score should be on a scale of 1 to 7, where 1 indicates the concept... | https://arxiv.org/abs/2505.20088v1 |
MA-RAG: Multi-Agent Retrieval-Augmented Generation via Collaborative Chain-of-Thought Reasoning Thang Nguyen Dartmouth College thangnv.th@dartmouth.eduPeter Chin Dartmouth College peter.chin@dartmouth.edu Yu-Wing Tai Dartmouth College yu-wing.tai@dartmouth.edu Abstract We present MA-RAG, a Multi-Agent framework for Ret... | https://arxiv.org/abs/2505.20096v1 |
sparse methods [Jones, 1972, Robertson and Zaragoza, 2009] and dense retrieval [Reimers and Gurevych, 2019, Karpukhin et al., 2020], each with respective weaknesses such as lexical gaps [Berger et al., 2000] or retrieval failure on out-of-distribution and multi-hop queries [Dai et al., 2023]. Augmentation methods often... | https://arxiv.org/abs/2505.20096v1 |
advancements over the past few years. Beginning with GPT-1 [Radford et al., 2018] on the Transformer architecture [Vaswani et al., 2017], subsequent models like GPT-2 [Radford et al., 2019], GPT-3 [Brown et al., 2020], and the latest GPT-4 [OpenAI, 2024] have significantly enhance capabilities in text understanding and... | https://arxiv.org/abs/2505.20096v1 |
such sandboxed, physical, or abstract [Hong et al., 2024, Mao et al., 2025, Park et al., 2023], and assume roles that are predefined, emergent, or data-driven [Du et al., 2024, Xiong et al., 2023]. Communication 3 PlannerWhere was the only European Cup Final in which Jupp Heynkes played held?which year did Jupp Heynkes... | https://arxiv.org/abs/2505.20096v1 |
et al., 2025] suggest potential to bypass retrieval altogether, practical limitations remain: effective context utilization is still far below advertised limits [Modarressi et al., 2025], and processing long sequences significantly increases inference cost and latency. More importantly, RAG is not merely a workaround f... | https://arxiv.org/abs/2505.20096v1 |
evidence, the QA agent synthesizes an answer using in-context learning. It produces a response aifor each step si, which is passed to the next iteration. Once all steps are completed, the final answer is assembled and returned to the user. A key feature of MA-RAG is its dynamic and modular agent invocation. Rather than... | https://arxiv.org/abs/2505.20096v1 |
[Liu et al., 2024b], RankRAG [Yu et al., 2024]1, Adaptive-RAG [Jeong et al., 2024], ReAct [Yao et al., 2023], Self-Ask [Press et al., 2023], and Smart-RAG [Gao et al., 2025]. Evaluation Metrics. ForOpen-domain QA tasks, we use Exact Match (EM) as the main metric for comparison, while we use Accuracy (Acc) forFact verif... | https://arxiv.org/abs/2505.20096v1 |
passage extraction capabilities inherent in MA-RAG are particularly advantageous in handling complex retrieval conditions. Overall, these results highlight the critical role of multi-agent coordination in improving open-domain QA performance, emphasizing that the integration of specialized agents for different reasonin... | https://arxiv.org/abs/2505.20096v1 |
multi-hop QA tasks like HotpotQA and 2WikimQA, it is critical to allocate larger models to the QA, planner and extractor agents to maintain performance. Smaller models can be used for step definer with minimal loss, enabling more efficient resource allocation in practice. Table 3: The accuracy of MA-RAG and baselines o... | https://arxiv.org/abs/2505.20096v1 |
Earl of Oxford’s father) Step query : Who is Edward De Vere, 17th Earl of Oxford’s father? Retrieved DocIDs : 129772, 6127861, 129773, 14010350, 12540430 Extracted Information :doc-129772 : [No related information about his father.], doc-6127861 : [No related information about his father.], doc-129773 : [Edward De Vere... | https://arxiv.org/abs/2505.20096v1 |
the only son of John de Vere, 16th Earl of Oxford, and Margery Golding. After the death of his father in 1562, he became a ward of Queen Elizabeth and was sent to live in the household ... Answer : John de Vere, 16th Earl of Oxford ReAct (70B)Query : Who is Edward De Vere, 17Th Earl Of Oxford’s paternal grandfather? AI... | https://arxiv.org/abs/2505.20096v1 |
or generation, uses chain-of-thought prompting to enable structured and context-sensitive query resolution. Unlike prior work that tunes individual components or requires supervised training, MA-RAG is entirely training-free and generalizes well across domains and question types. 9 Empirical results on five open-domain... | https://arxiv.org/abs/2505.20096v1 |
and A. Bosselut. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079 , 2023. Z. Dai, V . Y . Zhao, J. Ma, Y . Luan, J. Ni, J. Lu, A. Bakalov, K. Guu, K. B. Hall, and M. Chang. Promptagator: Few-shot dense retrieval from 8 examples. In International Conference on Learning... | https://arxiv.org/abs/2505.20096v1 |
Pasupat, and M.-W. Chang. Realm: retrieval-augmented language model pre-training. InInternational Conference on Machine Learning , 2020. J. He, C. Zhou, X. Ma, T. Berg-Kirkpatrick, and G. Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations (IC... | https://arxiv.org/abs/2505.20096v1 |
augmented generation. In H. Bouamor, J. Pino, and K. Bali, editors, Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP , 2023b. Q. Jin, B. Dhingra, Z. Liu, W. Cohen, and X. Lu. PubMedQA: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on ... | https://arxiv.org/abs/2505.20096v1 |
Conference on Empirical Methods in Natural Language Processing, EMNLP , pages 3045–3059, 2021. P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Ne... | https://arxiv.org/abs/2505.20096v1 |
M. Chadwick, M. Glaese, S. Young, L. Campbell- Gillingham, G. Irving, and N. McAleese. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147 , 2022. Meta. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 , 2024. A. Modarressi, H. Deilamsalehy, F. Dernoncourt, T. Bu... | https://arxiv.org/abs/2505.20096v1 |
of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages 8364–8377, June 2024. N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao. Reflexion: language agents with verbal reinforcement learning. In Advance... | https://arxiv.org/abs/2505.20096v1 |
. Wang. PMC-LLaMA: toward building open-source language models for medicine. Journal of the American Medical Informatics Association , 2024. G. Xiong, Q. Jin, Z. Lu, and A. Zhang. Benchmarking retrieval-augmented generation for medicine. In Findings of the Association for Computational Linguistics, ACL , 2024. K. Xiong... | https://arxiv.org/abs/2505.20096v1 |
short spans extracted from Wikipedia articles. We use 2837 questions from the development set in the KILT benchmark [Petroni et al., 2021] for evaluation. •TriviaQA [Joshi et al., 2017] includes challenging trivia questions written by trivia enthusi- asts, paired with independently collected evidence documents. We use ... | https://arxiv.org/abs/2505.20096v1 |
underlying LLMs. However, we caution that these benchmarks may not fully reflect the advantages of retrieval-augmented methods. Strong LLMs like GPT-4 already perform well without external retrieval (e.g., 84.8 EM on TriviaQA and 87.7 accuracy on FEVER), likely due to the fact that many questions are either single-hop ... | https://arxiv.org/abs/2505.20096v1 |
documents : List [str] 4 doc_ids : List [str ] 5 notes : List [ str] 6 final_raw_answer : QAAnswerState GraphState . The top-level state object that coordinates the MA-RAG pipeline. 1class GraphState ( TypedDict ): 2 original_question : str 3 plan : List [str] 4 past_exp : Annotated [ List [ PlanExecState ], operator .... | https://arxiv.org/abs/2505.20096v1 |
note down sentences, phrases, or words from the passages that relate to the question. **Remove Irrelevant Details**: Ensure that all extracted information is relevant to the question, eliminating any unnecessary or unrelated content. # Output Format - Output a list of notes. Each note contains related information from ... | https://arxiv.org/abs/2505.20096v1 |
Therefore, the answer is straightforward based on the retrieved context. Step 2 (Determine the character role played by Thomas Doherty in that sitcom.) Step query : What character role does Thomas Doherty play in the sitcom ’The Lodge’? Retrieved DocIDs : 20322850, 19608218, 20322852, 20322851, 19608221 Extracted Infor... | https://arxiv.org/abs/2505.20096v1 |
several episodes. During 2007, Doherty appeared in two television projects. She first appeared in ""Kiss Me Deadly: A Jacob Keane Assignment"" for the Here TV network and followed up with a starring role in the holiday film Doc-1423761 : to the stage, making a disastrous start in John Phillips’ ""Man on the Moon"" (197... | https://arxiv.org/abs/2505.20096v1 |
He graduated after studying music theatre at MGA Academy in July 2015 and immediately signed a contract with Olivia Bell Management in London. Doherty was trained in acting, singing, and various types of dance including contemporary, hip hop, jazz, tap and ballet. After graduating from The MGA Academy of Performing Art... | https://arxiv.org/abs/2505.20096v1 |
Century was the official organ of the National Council of Women of Canada (NCWC).], doc-3964891 : [No related information from this document.], doc-12413254 : [The magazine Woman’s Century was published by the NCWC (National Council of Women of Canada). This is evidenced by the phrase ¨Between 1914 and 1921 the NCWC pu... | https://arxiv.org/abs/2505.20096v1 |
children. She became involved in the National Council of Women. The first issue of "Woman’s Century" appeared in May 1913. It was largely produced out of MacIver’s home, with the help of her husband Doc-3964891 : The Century Magazine The Century Magazine was first published in the United States in 1881 by The Century C... | https://arxiv.org/abs/2505.20096v1 |
World War I, but stood up for women’s rights and universal suffrage. The founder of "Woman’s Century" was Jessie Campbell MacIver. She had come to Canada from Scotland with her husband, a lawyer, and five children. She became involved in the National Council of Women. The first issue of "Woman’s Century" appeared in Ma... | https://arxiv.org/abs/2505.20096v1 |
arXiv:2505.20097v1 [cs.CL] 26 May 2025S2LPP: Small-to-Large Prompt Prediction across LLMs Liang Cheng†Tianyi Li‡* Zhaowei Wang§Mark Steedman† †University of Edinburgh‡Amazon Alexa AI§HKUST L.Cheng-13@sms.ed.ac.uk tylteddy@amazon.co.uk m.steedman@ed.ac.uk Abstract The performance of pre-trained Large Lan- guage Models (... | https://arxiv.org/abs/2505.20097v1 |
to identify optimal prompt tem- plates from automatically generated prompt candi- dates for larger target models. This approach would help to reduce the deployment cost of LLMs, espe- cially when faced with diverse and dynamic sets of open-domain knowledge. We show the effective- ness of the S2LPP approach on open-doma... | https://arxiv.org/abs/2505.20097v1 |
natural language prompt templates, addressing a broader and more common scenario in NLP research. In this work, we set up a series of experiments to demonstrate the consistency of prompt preference across LLMs. We present the findings from our analyses in §3, and propose a lightweight approach to leverage these finding... | https://arxiv.org/abs/2505.20097v1 |
Each premise and hypothesis is also structured as a relation triple, containing a single predicate with two entity arguments, wherein identical entities are present in both the premise and the hypothesis. A distinctive feature of the Levy/Holt dataset is the in- clusion of inverse pairs for all premise-hypothesis- labe... | https://arxiv.org/abs/2505.20097v1 |
of prompts across LLaMA-3 of different sizes. (b) Accuarcy of prompts across DeepSeek-R1 of different sizes. Figure 2: The figure illustrates the accuracy of differ- ent prompts across LLaMA-3 and DeepSeek models of varying sizes on the directional Levy/Holt (NLI task). The x-axis represents the various candidate promp... | https://arxiv.org/abs/2505.20097v1 |
A prompt-generation model is used to generate a set of candidate natural lan- guage prompt templates. Prompt selection: Prompt selection is the cru- cial step in the S2LPP pipeline. By leveraging the consistency of prompt preference, we utilize smaller LMs as the prompt-selection models to assess each prompt by its per... | https://arxiv.org/abs/2505.20097v1 |
to be the most favored prompt. Average scores among prompts: We compute the mean accuracy across the candidates to measure the overall performance of all generated prompts. This methodology allows us to compare the quality of our selected prompts against the average perfor- mance level among all prompts. Manual Prompts... | https://arxiv.org/abs/2505.20097v1 |
The green column represents the baseline using the first-generated prompt, while the red column illustrates the accuracy with the oracle prompt, which is the upper bound of the target model (GPT-3.5). the effect of various sizes and families of smaller models in the prompt-selection process, shown in Figure 5. As depic... | https://arxiv.org/abs/2505.20097v1 |
preference for retrieved contexts, aligning with our findings on prompt preference consistency, and further supports the effectiveness of applying this approach to RAG. CoT Prompts Selection with Small LLMs for Arithmetic Reasoning: Shum et al. (2023) pro- pose a two-step pipeline, AutomateCoT , for gener- ating CoT pr... | https://arxiv.org/abs/2505.20097v1 |
questions involving new, unseen knowledge, by exploiting smaller models to select highly performant prompts at minimal cost in computation. We validate the efficacy of the approach in QA and NLI. Experi- ments demonstrate that the prompt templates se- lected with our strategy outperform baselines. Our methods also poss... | https://arxiv.org/abs/2505.20097v1 |
Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jias... | https://arxiv.org/abs/2505.20097v1 |
William El Sayed. 2023. Mis- tral 7b. ArXiv , abs/2310.06825. Ellen Jiang, Kristen Olson, Edwin Toh, Alejandra Molina, Aaron Donsbach, Michael Terry, and Carrie J Cai. 2022. Promptmaker: Prompt-based prototypingwith large language models. In CHI Conference on Human Factors in Computing Systems Extended Ab- stracts , pa... | https://arxiv.org/abs/2505.20097v1 |
prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 4222–4235. Kashun Shum, Shizhe Diao, and Tong Zhang. 2023. Automatic prompt augmentation and selection with chain-of-thought from labeled data. In Findings of the Association for Computational Linguistics:... | https://arxiv.org/abs/2505.20097v1 |
In International conference on machine learning , pages 12697–12706. PMLR. Lianmin Zheng, Wei-Lin Chiang, Ying 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 chatbot arena. In Thirty-seventh Conference on Neural In- for... | https://arxiv.org/abs/2505.20097v1 |
bold. D Consistency across Different Models Besides the LLaMA-3 and DeepSeek-R1 models, we compare the performance of more LLMs across a spectrum of generated prompts in Figure 6, span- ning all the relations present within the Google- RE. The results indicate that, with the exception of LLaMA-2 70B on PlaceOfBirth , L... | https://arxiv.org/abs/2505.20097v1 |
P413 p 0 p0 p0 p0 p0 original network P449 p 0 p0 p0 p0 p0 shares border with P47 p 8 p8 p8 p8 p3 named after P138 p 0 p6 p6 p6 p6 original language of film or TV show P364 p 1 p1 p1 p1 p1 member of sports team P54 p 0 p0 p0 p0 p0 member of P463 p 1 p1 p1 p1 p1 field of work P101 p 6 p2 p2 p2 p0 occupation P106 p 3 p4 ... | https://arxiv.org/abs/2505.20097v1 |
arXiv:2505.20099v1 [cs.CL] 26 May 2025Large Language Models Meet Knowledge Graphs for Question Answering: Synthesis and Opportunities Chuangtao Ma1, Yongrui Chen2, Tianxing Wu2, Arijit Khan1, Haofen Wang3 1Aalborg University, Denmark,2Southeast University, China,3Tongji University, China {chuma, arijitk}@cs.aau.dk, {yo... | https://arxiv.org/abs/2505.20099v1 |
RAG-based QA that may further tend to generate inconsistent answers. (2)Poor relevance and quality of retrieved context: The accuracy of the generated answers in RAG-based QA largely depends on the relevance and quality of the retrieved context, where irrelevant context leads to incorrect results. (3)Lack of iterative ... | https://arxiv.org/abs/2505.20099v1 |
main technical chal- lenge. To reduce the retrieval latency and im- prove the quality of the retrieved context for multi- document QA, KGP (Wang et al., 2024d) intro- duces an LLM-based graph traversal agent for re- trieving relevant knowledge from KG. Similarly, CuriousLLM (Yang and Zhu, 2025) integrates a knowledge g... | https://arxiv.org/abs/2505.20099v1 |
al., 2022); LLM-ARK (Huang, 2023); ToG (Sun et al., 2024a); ToG-2 (Ma et al., 2025b); KG-CoT (Zhao et al., 2024b) Agent-based KG Guidelines §3.2.3KG-Agent (Jiang et al., 2024); ODA (Sun et al., 2024b); GREASELM (Zhang et al., 2021) KGs as Refiners and Validators §3.3KG-Driven Filtering and Validation §3.3.1ACT-Selectio... | https://arxiv.org/abs/2505.20099v1 |
the unsupervised retrieval with LLMs based on rein- 3 forcement learning-driven knowledge distillation. 2.6 Tempral QA The challenges of temporal QA lie in fully un- derstanding the implicit time constraints and ef- fectively incorporating them with temporal knowl- edge for temporal reasoning. To improve the ac- curacy... | https://arxiv.org/abs/2505.20099v1 |
respectively, which selects the new knowledge and integrates it with LLMs. Fine-tuning LLMs with text and knowledge graphs can improve their performance of LLMs on spec- ified tasks. For instance, KG-Adapter (Tian et al., 2024) improves parameter-efficient fine-tuning of LLMs by introducing a knowledge adaptation layer... | https://arxiv.org/abs/2505.20099v1 |
for multiple-doc QA (Wang et al., 2024d). The cross- modal reasoning can facilitate the cross-modal in- teraction and alignment for multi-modal QA (Suri et al., 2024). Additionally, the question decom- position of multi-hop QA can be augmented by fusing the knowledge from LLMs and KGs, which further facilitates iterati... | https://arxiv.org/abs/2505.20099v1 |
KG-CoT (Zhao et al., 2024b) leverages external KGs to generate reasoning paths for joint reasoning of LLMs and KGs to enhance the reasoning capabilities of LLMs for QA. 3.2.3 Agent-based KG Guidelines KGs can also be integrated into the reasoning pro- cess of LLMs as a component within an Agent system, as shown in Figu... | https://arxiv.org/abs/2505.20099v1 |
traversal is computation- ally intensive and time-consuming. Moreover, the reasoning capabilities of KGs mainly depend on the completeness and knowledge coverage of KGs, where the incomplete, inconsistent, and outdated knowledge from KGs might induce noise or con- flicts. The main challenge lies in how to improve the r... | https://arxiv.org/abs/2505.20099v1 |
reasoning in complex QA. 3.3.3 Aligning with Complex QA The approaches that leverage the retrieved factual evidence from KGs for refinement and validation are designed to augment the capability of LLMs in understanding user interactions and verifying the intermediate reasoning for multi-hop QA (Chen et al., 2024a) and ... | https://arxiv.org/abs/2505.20099v1 |
a KG prompting approach to enhance the prompt for LLMs and optimize the knowledge re- trievalby introducing the KG construction module and LLM-based graph traversal agent. (3)Cost- based optimization. It aims to minimize computa- tion costs by reducing the number of calls to LLMs and accelerating knowledge retrieval. I... | https://arxiv.org/abs/2505.20099v1 |
of the latter. Knowledge Alignment and Dynamic Integra- tion. Once a KG snapshot is injected into an LLM, it starts to become outdated. Just like real-world KGs usually involve adding new entities, delet- ing relations, and resolving contradictions. Fu- ture work should: (1) Quantify alignment : We lack metrics that sc... | https://arxiv.org/abs/2505.20099v1 |
works due to the rapid expan- sion of works on this topic. Moreover, the survey mainly highlights the alignments between the re- cent methodologies of incorporating LLMs and KGs for QA and the challenges of the various com- plex QA tasks, while these taxonomies from differ- ent perspectives are non-exclusive, and the o... | https://arxiv.org/abs/2505.20099v1 |
question answering. Neural Networks , 181:106833. Yifu Gao, Linbo Qiao, Zhigang Kan, Zhihua Wen, Yongquan He, and Dongsheng Li. 2024. Two-stage generative question answering on temporal knowl- edge graph using large language models. In ACL Findings , pages 6719–6734. Hengrui Gu, Kaixiong Zhou, Xiaotian Han, Ning- hao L... | https://arxiv.org/abs/2505.20099v1 |
Zhu, and Ji-Rong Wen. 2024. Kg-Agent: An efficient autonomous agent frame- work for complex reasoning over knowledge graph. arXiv:2402.11163 . Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What dis- ease does this patient have? a large-scale open do- main question answering... | https://arxiv.org/abs/2505.20099v1 |
Sun, Jing Zhao, Zhe Zhao, and Wei Hu. 2024b. KnowLA: Enhancing parameter- efficient finetuning with knowledgeable adaptation. InNAACL , pages 7146–7159. Chuangtao Ma, Yongrui Chen, Tianxing Wu, Arijit Khan, and Haofen Wang. 2025a. Unifying large language models and knowledge graphs for question answering: Recent advanc... | https://arxiv.org/abs/2505.20099v1 |
knowledge graphs. In KONVENS , pages 155–164. Diego Sanmartin. 2024. KG-RAG: Bridging the gap be- tween knowledge and creativity. arXiv:2405.12035 . Juan Sequeda, Dean Allemang, and Bryon Jacob. 2024. A benchmark to understand the role of knowledge graphs on large language model’s accuracy for ques- tion answering on e... | https://arxiv.org/abs/2505.20099v1 |
systematic exploration of knowl- edge graph alignment with large language models in retrieval augmented generation. In AAAI , pages 25291–25299. Chaojie Wang, Yishi Xu, Zhong Peng, Chenxi Zhang, Bo Chen, Xinrun Wang, Lei Feng, and Bo An. 2023. keqing: knowledge-based question answer- ing is a nature chain-of-thought me... | https://arxiv.org/abs/2505.20099v1 |
Wang, Manasi Deshpande, Xiaofeng Wang, and Zheng Li. 2024. Retrieval-augmented generation with knowledge graphs for customer service question answering. In SIGIR , pages 2905–2909. Rui Yang, Haoran Liu, Edison Marrese-Taylor, Qingcheng Zeng, Yuhe Ke, Wanxin Li, Lechao Cheng, Qingyu Chen, James Caverlee, Yutaka Mat- suo... | https://arxiv.org/abs/2505.20099v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.