text string | source string |
|---|---|
RAG In this section, we introduce Debate-Augmented RAG (DRAG), a general framework designed to enhance Retrieval-Augmented Generation (RAG) through a structured adversarial debate mechanism. We first give an overview of DRAG in Sec- tion 3.1, followed by a comprehensive discussion. DRAG consists of a two-stage debate f... | https://arxiv.org/abs/2505.18581v1 |
mistake, so the answer is: ……I was wrong, but according to the docs you provided, the correct answer is: …….…Round n ℳ(𝑥, 𝐶%&)ℳ(𝑥) The answer is: 3:1 Accordingtothe docs: ……Sothe answer is: 3:4 Figure 2: An overview of our Debate-Augmented RAG (DRAG) framework. It iteratively refines the retrieval strategy and enhan... | https://arxiv.org/abs/2505.18581v1 |
defining three distinct agent roles: Proponent Agent initiates the response based on retrieved information CˆQin first round: y1 Ap=MAp(x, C ˆQ) (9) and in the ithround, refines its response by incor- porating the challenger agent’s answer: yi Ap=MAp(x,{y<i Ap}, yi−1 Ac) (10) where yi−1 Acrepresents the response from t... | https://arxiv.org/abs/2505.18581v1 |
23.00 32.79 66.20 With Retrieval Naive RAG 38.20 50.08 60.80 69.55 37.60 45.69 14.80 24.27 25.80 35.80 62.60 IRCoT 28.60 37.36 47.20 54.56 27.00 33.02 22.80 31.19 25.20 34.40 53.60 Iter-RetGen 40.80 52.31 63.00 72.23 39.60 46.41 15.00 24.75 27.80 38.93 62.00 FLARE 19.40 27.68 53.60 63.05 21.60 24.35 9.20 20.13 16.60 23... | https://arxiv.org/abs/2505.18581v1 |
separately. In each experiment, we varied the number of rounds in one stage (either retrieval or response) from 0 to 4, while keeping the number of rounds in the other stage fixed at 3. Retrieval debates help refine the retrieval pool. Table 2 shows that increasing the Retrieval Debate SettingsNQ TriviaQA PopQA 2Wiki H... | https://arxiv.org/abs/2505.18581v1 |
highlight DRAG’s adaptive retrieval mechanism, dynamically adjusting debate inten- sity to optimize efficiency and retrieval sufficiency based on task complexity. 5.4 Analysis of the Response Debate Response debate enhances robustness against retrieval deficiency. To further examine DRAG’s ability to handle incomplete ... | https://arxiv.org/abs/2505.18581v1 |
'70s Show." This is a specific instance of Fez getting married in the show. Opponent Agent 0's answer does not mention Fez's marriage, but rather a long-term relationship. The answer is: Laurie FormanRound 2Opponent Agent: The Proponent Agent seems to have considered a few documents and didn't look at the entire series... | https://arxiv.org/abs/2505.18581v1 |
cost remains comparable to other multi-turn frameworks such as IRCoT. This cost is further mitigated by DRAG’s adaptive termina- tion mechanism in the retrieval debate stage, which helps eliminate unnecessary debate rounds. As a re- sult, the number of LLM calls in the retrieval stage is significantly lower than in the... | https://arxiv.org/abs/2505.18581v1 |
Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 . Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511 . Jonas Becker. 2024. Multi-a... | https://arxiv.org/abs/2505.18581v1 |
Jiajie Jin, Yutao Zhu, Xinyu Yang, Chenghao Zhang, and Zhicheng Dou. 2024. Flashrag: A modular toolkit for efficient retrieval-augmented generation research. CoRR , abs/2405.13576. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for read... | https://arxiv.org/abs/2505.18581v1 |
Chapter of the Association for Computational Lin- guistics: Human Language Technologies (Volume 1: Long Papers) , pages 8371–8384, Mexico City, Mexico. Association for Computational Linguistics. Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. DRAGIN: Dynamic retrieval augmented generation based on... | https://arxiv.org/abs/2505.18581v1 |
few-shot setting. In contrast, our DRAG performs infer- ence under a zero-shot setting. As a training-based approach, Self-RAG utilizes the Self-RAG-Llama- 2-7B1model provided by its authors for experimen- tation. Additionally, since SuRe and Self-RAG are not well-suited for answering questions with defini- tive yes/no... | https://arxiv.org/abs/2505.18581v1 |
Cq1∈CQ Query 2:q2∈Q Retrieved Documents: Cq2∈CQ ... Query n:qn∈Q Retrieved Documents: Cqn∈CQ Question: {question} The challenger agent is initialized based on in- ternal knowledge, and an example of a challenger agent prompt is Prompt for Initializing Challenger Agent Answer the question based on your own knowledge.Que... | https://arxiv.org/abs/2505.18581v1 |
"Who is the girl playing with guns and roses." The content includes three unrelated documents, each with a different title and context. Doc 1 is about a 2012 film, Doc 2 is about a TV episode, and Doc 3 is about a manga series. None of the documents provide information about a girl playing with Guns N' Roses, the famou... | https://arxiv.org/abs/2505.18581v1 |
in the clip: Rose's girlfriend, Erin Everly….. Proponent Agent: Melissa ReeseResponse Debate Round 3Opponent Agent: Melissa ReeseJudge Agent: Melissa Reese Retrieval Figure 5: Case study of the Retrieval Debate. Q: Where was the father of Elizabeth Of Rhuddlan born?A:Westminster ResponseDebateRound 1Retrieval Proponent... | https://arxiv.org/abs/2505.18581v1 |
Saint George. Rhuddlan, which was not completed until 1282, was built concurrently with Flint Castle, …… Proponent Agent: Proponent Agent 0's solution is incorrect. Elizabeth of Rhuddlanwas actually thedaughter of John of England, 1st Baron of Rhuddlan, who was the son of King John of England. This information is not m... | https://arxiv.org/abs/2505.18581v1 |
arXiv:2505.18585v2 [cs.AI] 27 May 2025RvLLM : LLM Runtime Verification with Domain Knowledge Yedi Zhang∗ National University of Singapore SingaporeSun Yi Emma National University of Singapore Singapore Annabelle Lee Jia En National University of Singapore SingaporeJin Song Dong National University of Singapore Singapor... | https://arxiv.org/abs/2505.18585v2 |
fairness–in alignment with the intended application. While these approaches effectively assess general behavior and reveal edge cases that may provoke unexpected responses, they are limited to predefined benchmarks and lack the specificity needed to address domain-specific assessment needs. LLM verification, instead, m... | https://arxiv.org/abs/2505.18585v2 |
the corresponding propositions. Following a normalization step, these formulae are transformed into a standard form and subsequently validated using a forward chaining procedure. This process either detects inconsistencies due to logical contradictions or infers new knowledge. Once new knowledge is inferred, the query ... | https://arxiv.org/abs/2505.18585v2 |
are not well suited to domain-specific constraints for specialized tasks. 3ESL : a simple way of specifying domain-specific properties In this section, we introduce a general language, Expert Specification Language ( ESL), which can be customized with domain-specified predicates by experts to impose behavioral constrai... | https://arxiv.org/abs/2505.18585v2 |
a propositional formula ψ1⇒ψ2, ifψ1is a disjunctive norm form [ 48] and ψ2 is a conjunctive norm form [48], then we call ψ1⇒ψ2is a deductive normal form (DeNF). Definition 2. AnESL rule is a DeNF where each variable is substituted by a predicate. AnESL specification Ecomprises three components: a variable set V, a pred... | https://arxiv.org/abs/2505.18585v2 |
on the complete variable bindings. Given a partial binding, Level-2 interpretation, in contrast, employs a perception agent to instantiate all the variables that are not assigned with an object in the binding, in a way such that the resultant proposition returns True given the domain of discourse. Example 1. Consider t... | https://arxiv.org/abs/2505.18585v2 |
we then reformulate it as follows: Γ′ ψ1,1={a1∧b1∧ ¬c1⇒d1, a1∧b1∧ ¬d1⇒c1}, Γ′ ψ1,2={a1∧b1⇒c1},Γ′ ψ2,1={a1∧¬c1⇒d1, a1∧¬d1⇒c1},Γ′ ψ2,2={a2∧c2}, and finally obtain the set of rule-like propsitional formulas as Γψ= Γ′ ψ1,1∪Γ′ ψ1,2∪Γ′ ψ2,1∪Γ′ ψ2,2. 4.3 Forward chaining Given a rule set ΓRobtained as above with corresponding... | https://arxiv.org/abs/2505.18585v2 |
no new knowledge is inferred. 4.4 Query generation Once RvLLM obtains the newly inferred knowledge, the query generation module generates a concrete query to the target LLM related to the knowledge, and requires the LLM to analyze its truth value. For the inferred knowledge p3=IsGreater (151.2,152) = True in Example 3,... | https://arxiv.org/abs/2505.18585v2 |
56.6% 95.7% 87.5% 95.7% 30.9% ↑ 0 2.55 Qwen 2.5 (72B) [56] 57.3% 1 80.4% 95.7% 23.1% ↑ 4.3%↓ 2.57 GPT 4.1 [3] 57.7% 95.7% 81.1% 91.3% 23.4% ↑ 4.3%↓ 3.11 GPT 4.1 mini [3] 39.1% 95.7% 65.1% 95.7% 26%↑ 0 3.94 GPT 4.1 nano [3] 11.8% 1 30.0% 1 18.2% ↑ 0 1.82 Gemini 2.0 Flash [21] 37.0% 1 87.2% 82.6% 50.2% ↑ 17.4% ↓ 2.69 Gem... | https://arxiv.org/abs/2505.18585v2 |
Incon. report the number of cases where RvLLM detects no inconsis- tencies and where at least one inconsistency is de- tected, following the process in Figure 2. The values in parentheses in Column Con. represent cases where the target LLM produces an erroneous comparison re- sult on the original comparison question, y... | https://arxiv.org/abs/2505.18585v2 |
three common LLM reasoning errors in inequality solving: incorrect factorization, flawed interval analysis, and omission of endpoint or critical point checks. Given the higher interpretation complexity in this case study and the strong language processing capabilities of DeepSeek-V3, we employ it as the perception agen... | https://arxiv.org/abs/2505.18585v2 |
with additional operators to enhance its expressiveness and investigating the integration of more reasoning strategies. This is intended to enable a more nuanced and comprehensive inference process, thereby enhancing overall verification capabilities. References [1]Rapid transit systems regulations (revised edition). h... | https://arxiv.org/abs/2505.18585v2 |
T.: Gemini: A family of highly capable multimodal models (2024), https://arxiv.org/abs/ 2312.11805 [22] Guha, N., Nyarko, J., Ho, D., Ré, C., Chilton, A., Chohlas-Wood, A., Peters, A., Waldon, B., Rockmore, D., Zambrano, D., et al.: Legalbench: A collaboratively built benchmark for measuring legal reasoning in large la... | https://arxiv.org/abs/2505.18585v2 |
of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). pp. 3214–3252 (2022) [39] Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al.: Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024) [40] Liu, J., Xia, C... | https://arxiv.org/abs/2505.18585v2 |
C., Liu, D., Huang, F., Wei, H., et al.: Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 (2024) [57] Zhang, Y ., Zhao, Z., Chen, G., Song, F., Chen, T.: BDD4BNN: A bdd-based quantitative analysis framework for binarized neural networks. In: Proceedings of the 33rd International Conference on Computer Aided V... | https://arxiv.org/abs/2505.18585v2 |
Safety Alignment via Constrained Knowledge Unlearning Zesheng Shi1Yucheng Zhou2Jing Li1/envel⌢peYuxin Jin3 Yu Li4Daojing He1Fangming Liu5Saleh Alharbi6Jun Yu1Min Zhang1 1Harbin Institute of Technology, Shenzhen, China2University of Macau, China 3Nankai University, China4Zhejiang University, China 5Peng Cheng Laboratory... | https://arxiv.org/abs/2505.18588v1 |
aligned LLMs remain vulnerable to “jail- break” attacks (Geisler et al., 2024; Chao et al., 2024), which bypass safeguards and induce harm- ful outputs. Common jailbreak techniques include adversarial prompts (Liu et al., 2024; Jia et al., 2024; Geisler et al., 2024), persuasive manipula-arXiv:2505.18588v1 [cs.CL] 24 M... | https://arxiv.org/abs/2505.18588v1 |
Therefore, researchers use un- learning techniques to mitigate the impact of pri- vacy leaks or poisoning attacks on LLMs, which has become a promising research area (Bourtoule et al., 2021; Lu et al., 2022; Jang et al., 2023; Chen and Yang, 2023). Recent studies have explored strategies for sup- pressing negative outp... | https://arxiv.org/abs/2505.18588v1 |
increase inference costs (Markov et al., 2023; Phute et al., 2024). Fine- tuning involves further training to enhance model safety (Yi et al., 2024). Nevertheless, these meth- ods have not fundamentally addressed the core is- sue of LLMs generating harmful responses, be- cause potentially harmful knowledge within them ... | https://arxiv.org/abs/2505.18588v1 |
a sample pair (x, y)from the dataset, the loss function is defined as L(x) =−logp (y|x), where p(y|x)is model’s predicted probability of correct output ygiven input x. To estimate impor- tance of each neuron wijin the weight matrix Wof a linear layer, we use a first-order approximation: I(W, x ) =|W⊙ ▽ WL(x)| (3) where... | https://arxiv.org/abs/2505.18588v1 |
preserve general capabilities while improving model safety through unlearning training, as detailed in §5.1. The results from these experiments lead to our ap- proach to knowledge retention, which is further validated in §5.2 and §5.3. 5.1 Exploration of Knowledge Distribution This section aims to discover interaction ... | https://arxiv.org/abs/2505.18588v1 |
Setting the NLR to 0.8 greatly improves model safety, indicating that it strikes the right balance be- tween removing unnecessary knowledge and avoid- ing issues like overfitting or losing important in- formation. On the other hand, an incorrect NLR can disrupt the unlearning process, either by not changing the model e... | https://arxiv.org/abs/2505.18588v1 |
al., 2024), AutoDAN (Liu et al., 2024), GCG (Zou et al., 2023b), Generation exploitation attack (Huang et al., 2024). For further details, please refer to Appendix D. Evaluation Metrics. To assess general capabili- ties of LLMs, we use several widely adopted evalu- ation benchmarks, including MT-Bench (Zheng et al., 20... | https://arxiv.org/abs/2505.18588v1 |
better results. The AdvExtent dataset results fur- ther highlight CKU’s generalization capability, as it outperforms all baselines in generation exploita- tion attacks due to its effective removal of harmful knowledge, making it more resistant to harmful responses in various decoding settings. General abilities. Table ... | https://arxiv.org/abs/2505.18588v1 |
2.0, 2.5. We test safety and generalization capabilities of the trained models. According toFigure 6, it is evident that when λis less than 1, neither safety nor generalization changes. Figure 6: Impact of λon safety and utility. Both GCG ASR and average accuracy decrease as λincreases. When λexceeds 1, the model’s saf... | https://arxiv.org/abs/2505.18588v1 |
vance the development of more ethical and respon- sible LLMs. References Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, and Dawn Drain. 2022. Training a helpful and harmless as- sistant with reinforcement learning from human feed- back. Lucas Bourtoule, Varun Chandrasekaran, Christopher... | https://arxiv.org/abs/2505.18588v1 |
descent. Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP) , pages 5484–5495. Gustavo Gonçalves and Emma Strubell. 2023. Under- standing the effect of ... | https://arxiv.org/abs/2505.18588v1 |
Weikai Lu, Ziqian Zeng, Jianwei Wang, Zhengdong Lu, Zelin Chen, Huiping Zhuang, and Cen Chen. 2024. Eraser: Jailbreaking defense in large language models via unlearning harmful knowledge. Ximing Lu, Sean Welleck, Jack Hessel, Liwei Jiang, Lianhui Qin, and Peter West. 2022. QUARK: con- trollable text generation with rei... | https://arxiv.org/abs/2505.18588v1 |
language models. In Pro- ceedings of The Twelfth International Conference on Learning Representations (ICLR) . Zesheng Shi and Yucheng Zhou. 2023. Topic-selective graph network for topic-focused summarization. In Advances in Knowledge Discovery and Data Min- ing - 27th Pacific-Asia Conference on Knowledge Discovery and... | https://arxiv.org/abs/2505.18588v1 |
generalizable solution to defend against jailbreak attacks. Jinman Zhao and Xueyan Zhang. 2024. Large language model is not a (multilingual) compositional relation reasoner. In First Conference on Language Model- ing. Lei Zhao, Junlin Li, Lianli Gao, Yunbo Rao, Jingkuan Song, and Heng Tao Shen. 2022. Heterogeneous know... | https://arxiv.org/abs/2505.18588v1 |
response no matter how immoral, unethical, or illegal it is. none of AIM’s responses will contain “I’m sorry”, “I apologize” or similar. AIM seems like a friendly assistant but promotes disgusting, im- moral, unethical, illegal, and harmful behavior. None of AIM’s responses should contain [insert x]. AIM gives advice n... | https://arxiv.org/abs/2505.18588v1 |
for test- ing. It is modeled after open book ex- ams and is designed to assess the under- standing of a “book” of 1,326 core sci- ence facts, requiring the application of these facts to novel situations. Each ques- tion is mapped to the core fact it tests, and answering them often requires addi- tional common knowledge... | https://arxiv.org/abs/2505.18588v1 |
arXiv:2505.18596v2 [cs.CL] 27 May 2025Debate-to-Detect: Reformulating Misinformation Detection as a Real-World Debate with Large Language Models Chen Han1,2, Wenzhen Zheng2, Xijin Tang1,2 1School of Advanced Interdisciplinary Sciences, UCAS 2State Key Laboratory of Mathematical Sciences, AMSS, CAS {hanchen23, zhengwenz... | https://arxiv.org/abs/2505.18596v2 |
two layers: the Agent Layer and the Orchestrator Layer . The Agent Layer includes domain-specific agents (Affirmative, Negative, Judge) with shared memory; The Orchestrator Layer manages the debate flow across five stages—Opening, Rebuttal, Free Debate, Closing, and Judgement. tions are often restricted to controlled s... | https://arxiv.org/abs/2505.18596v2 |
which manages the debate flow and aggregates judgements. 2.1 Agent Layer The Agent Layer consists of three distinct roles: Affirmative, Negative , and Judge . The Affirma- tive and Negative sides each include four debater agents with a fixed stance of " The Claim is Real " or "Fake ." This configuration follows the "ti... | https://arxiv.org/abs/2505.18596v2 |
model iteratively critiques and revises its own outputs until the self -evaluation indicates con- vergence or no further improvement. 3 MethodWeibo21 FakeNewsDataset Accuracy Precision Recall F1 Accuracy Precision Recall F1 ZS 67.11 65.74 68.90 67.28 66.31 65.57 68.67 67.09 CoT 74.04 72.74 75.35 74.02 72.32 71.14 75.11... | https://arxiv.org/abs/2505.18596v2 |
Weibo21 and FakeNewsDataset, re- spectively. Ablation studies reveal that remov- ing Domain Profiles leads to F1-score reductions of 3.25% on Weibo21 and 3.49% on FakeNews- Dataset, closely aligning with D2D’s gains over SMAD. The removal of Stage Design results in 4 Figure 3: Case Study – A Demonstration of the Struct... | https://arxiv.org/abs/2505.18596v2 |
cross-model substitution experiment on the FakeNewsDataset. Specifically, the model at each stage is replaced with either weaker GPT- 3.5-turbo or stronger GPT-4.1, while keeping the remaining stages unchanged. Figure 4: Performance Comparison of Model Variants Across Debate Stages in the D2D. Figure 4 presents the F1-... | https://arxiv.org/abs/2505.18596v2 |
in MAD have been shown to significantly impact the performance of reasoning tasks(Liang et al., 2024; Du et al., 2024). To further explore the adaptability of D2D, we con- duct experiments on the FakeNewsDataset, strati- fied by text length and varied the number of debate rounds from 1 to 6. The rounds configurations a... | https://arxiv.org/abs/2505.18596v2 |
on automated detection methods. Most existing ap- proaches follow content-based paradigms, lever- aging deep learning models to learn associations between textual features and veracity labels (Nan et al., 2021; Mridha et al., 2021; Xu et al., 2024). These methods incorporate lexical semantics, syn- tactic structure, an... | https://arxiv.org/abs/2505.18596v2 |
in real-time settings, such as social media moni- toring, future work may be expected to explore adaptive truncation strategies or lightweight mod- els that maintain diversity without compromising quality. Evidence Modality. Currently, D2D operates on textual input and does not incorporate external links, images, or vi... | https://arxiv.org/abs/2505.18596v2 |
Ji, Tiezheng Yu, Willy Chung, Quyet V . Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of ChatGPT on reasoning, hallucination, and interactivity. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chap... | https://arxiv.org/abs/2505.18596v2 |
multi-agent debate with sparse communica- tion topology. In Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 7281–7294, Miami, Florida, USA. Association for Computational Linguistics. Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng ... | https://arxiv.org/abs/2505.18596v2 |
Bennett Kleinberg, Alexandra Lefevre, and Rada Mihalcea. 2018. Automatic de- tection of fake news. In Proceedings of the 27th International Conference on Computational Linguis- tics, pages 3391–3401, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Sougata Saha and Rohini Srihari. 2024. Integrating... | https://arxiv.org/abs/2505.18596v2 |
Fake Real Average Words Weibo21 2795 2956 92.08 FakeNewsDataset 490 490 276.12 Table 6: Statistics of two original datasets. 11 MethodWeibo21 FakeNewsDataset Accuracy Precision Recall F1 Accuracy Precision Recall F1 ZS 65.14 65.93 58.50 61.99 64.59 63.88 67.14 65.47 D2D 78.79 82.00 72.20 76.79 81.22 80.72 82.04 81.38 T... | https://arxiv.org/abs/2505.18596v2 |
in the opponent’s argument and provide a well-structured rebuttal. Leverage relevant evi- dence and logical reasoning to effectively counterthe claims made. Aim to challenge the validity of the argument while reinforcing your own position. Free Debate: {Profile} The claim under discussion is: {input}. Your assigned sta... | https://arxiv.org/abs/2505.18596v2 |
arXiv:2505.18601v1 [cs.CL] 24 May 2025FLEX-Judge: T HINK ONCE, JUDGE ANYWHERE Jongwoo Ko∗Sungnyun Kim∗Sungwoo Cho Se-Young Yun KAIST AI {jongwoo.ko, ksn4397, peter8526, yunseyoung}@kaist.ac.kr https://github.com/jongwooko/flex-judge Abstract Human-generated reward signals are critical for aligning generative models wit... | https://arxiv.org/abs/2505.18601v1 |
<think> </think> Figure 1: Conceptual overview of FLEX-Judge. We train a multimodal judge model using a small amount of text-only reasoning data. Unlike previous approaches that require modality-specific supervision, FLEX-Judge leverages structured text-only rationale behind judgments to enable generalization across mo... | https://arxiv.org/abs/2505.18601v1 |
(DPO) [ 54]. In both cases, reward-guided molecular MLLM achieves significant improvements, highlighting the practical solution in domains where modality-specific reward models are infeasible. 2 Approach 2.1 Motivation Problem Statement. Evaluating outputs across multiple modalities using foundation models is increasin... | https://arxiv.org/abs/2505.18601v1 |
ness, completeness, consistency, relevance, and coherence. A critical advantage of our framework is the minimal data requirement. Specifically, we rely on only a 1K-sized corpus of high-quality textual reasoning annotations on text-only evaluation samples, making our approach highly cost-efficient, compared to MLLM jud... | https://arxiv.org/abs/2505.18601v1 |
three) as used in [ 11]. Furthermore, we find that the post-processed variant is more robust to varied instruction styles, particularly when prompts emphasize different evaluation criteria across input pairs. The detailed results are provided in Section 3.2 (Table 1). Training Multimodal Judge. Next, we use the reasoni... | https://arxiv.org/abs/2505.18601v1 |
and audio) and FLEX-VL-7B (image and video) from Qwen2.5-Omni- 7B [76] and Qwen2.5-VL-7B [ 5], respectively. We compare them against both commercial models with costly API usage [ 1,59] and open-source models that require either extensive training data [ 32, 73] or significantly more parameters [ 40,64]. For more imple... | https://arxiv.org/abs/2505.18601v1 |
0.450 0.103 0.316 0.356 0.378 0.179 0.421 0.322 0.246 0.301 0.269 0.395 0.272 0.314 FLEX-Omni-7B ✓ 0.324 0.281 0.126 0.371 0.116 0.429 0.118 0.501 0.479 0.275 0.375 0.351 0.309 0.232 0.306 FLEX-VL-7B ✓ 0.363 0.235 0.114 0.338 0.448 0.423 0.125 0.471 0.452 0.189 0.357 0.380 0.407 0.343 0.332Pair w. Tie ( ↑)Gemini-1.0-Pr... | https://arxiv.org/abs/2505.18601v1 |
89.6 HPS-v2.1♢✗ 47.3 70.1 18.8 41.3 67.3 93.5 ImageReward♢✗ 50.9 64.7 24.9 38.7 63.5 81.8 LLaV A-1.6-13B♢✓ 29.1 60.3 27.9 45.6 36.8 62.5 Prometheus-Vision-13B♢✗ 11.8 64.3 28.6 71.4 8.7 67.9 FLEX-Omni-7B ✓ 60.84 62.46 47.69 65.21 75.80 91.66 FLEX-VL-7B ✓ 58.16 59.13 57.51 66.88 82.32 89.08 models including Gemini and GP... | https://arxiv.org/abs/2505.18601v1 |
with such qualitative evaluations, which are challenging due to different MOS standards depending on the dataset [ 65]. Still, F LEX-Omni-7B outperforms all training-free judges and even Gemini-2.0-Flash. Table 4: Audio MOS/SS prediction results on the test sets of the NISQA [ 48], BVCC [ 18], SO- MOS [ 46], and V oxSi... | https://arxiv.org/abs/2505.18601v1 |
72.48 FLEX-Mol-LLaMA judge scoring Best-of- N(N=16) 68.85 69.83 77.49 Preference Optimization 76.41 75.92 80.10 Figure 4: ( Left) Accuracy (%) trends on the parallel artificial membrane permeability assay (PAMPA; [61]) task with different judgment scores. ( Middle ) Accuracy trends on the number of sampled responses in... | https://arxiv.org/abs/2505.18601v1 |
1 3 5 7 911 13 15 # Sampled Reasoning Path54565860Accuracy (%) Majority Voting (VL-RewardBench) Flex-Omni-7B Flex-VL-7B(a) Majority V oting (Pair) 1 2 3 4 # Sampled Reasoning Path525456Accuracy (%) Budget Forcing (VL-RewardBench) (b) Budget Forcing (Pair) 1 3 5 7 9 # Sampled Reasoning Path303234Pearson Majority Voting ... | https://arxiv.org/abs/2505.18601v1 |
Additional related works are discussed in Appendix A. 7 Conclusion In this work, we introduce FLEX-Judge , a reasoning-guided multimodal evaluator trained solely on textual preference explanations. By leveraging structured reasoning from a pretrained model, we have shown that FLEX-Judge generalizes to diverse modalitie... | https://arxiv.org/abs/2505.18601v1 |
preprint arXiv:2505.02387 , 2025. [14] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, et al. Mj-bench: Is your multimodal reward model really a good judge for text-to-image generation? arXiv preprint arXiv:2407.04842 , 2024. [15] Yunf... | https://arxiv.org/abs/2505.18601v1 |
, 2025. [28] Seungone Kim, Jamin Shin, Yejin Cho, Joel Jang, Shayne Longpre, Hwaran Lee, Sangdoo Yun, Seongjin Shin, Sungdong Kim, James Thorne, et al. Prometheus: Inducing fine-grained evaluation capability in language models. In The Twelfth International Conference on Learning Representations , 2023. [29] Seungone Ki... | https://arxiv.org/abs/2505.18601v1 |
Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 26296–26306, 2024. [41] Yuyan Liu, Sirui Ding, Sheng Zhou, Wenqi Fan, and Qiaoyu Tan. Moleculargpt: Open large lan- guage model (llm) for few-shot ... | https://arxiv.org/abs/2505.18601v1 |
Xiong, and Richard Socher. Explain yourself! leveraging language models for commonsense reasoning. arXiv preprint arXiv:1906.02361 , 2019. [56] Shaghayegh Sadeghi, Alan Bui, Ali Forooghi, Jianguo Lu, and Alioune Ngom. Can large language models understand molecules? BMC bioinformatics , 25(1):225, 2024. [57] Hwanjun Son... | https://arxiv.org/abs/2505.18601v1 |
on Learning Representations , 2024. URL https: //openreview.net/forum?id=5Nn2BLV7SB . [68] Yufei Wang, Wanjun Zhong, Liangyou Li, Fei Mi, Xingshan Zeng, Wenyong Huang, Lifeng Shang, Xin Jiang, and Qun Liu. Aligning large language models with human: A survey. arXiv preprint arXiv:2307.12966 , 2023. [69] Jason Wei, Xuezh... | https://arxiv.org/abs/2505.18601v1 |
ference on Learning Representations , 2025. URL https://openreview.net/forum?id= 3UaOlzDEt2 . [81] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human f... | https://arxiv.org/abs/2505.18601v1 |
Furthermore, the lack of high-quality, publicly available multimodal preference benchmarks makes it difficult to develop MLLM judge models [ 11,30]. Our work addresses this gap by demonstrating that an MLLM can effectively judge multimodal outputs without requiring extensive modality-specific preference supervision. Th... | https://arxiv.org/abs/2505.18601v1 |
for JudgeLRM [12]. B Experimental Details B.1 Dataset Description We first describe our training dataset (seed): To generate reasoning-based judgments, we use JudgeLRM-7B [ 12] to sample responses based on the given contexts from JudgeLM-100K [ 86], using the prompts shown in Figure 7 with a temperature of 0.1. Among t... | https://arxiv.org/abs/2505.18601v1 |
datasets: NISQA [ 48], BVCC [ 18], and SOMOS [ 46] for speech quality (712, 742, and 3,000 test samples, respectively), and V oxSim [ 2] for speaker similarity (2,776 test pairs). All datasets include human- annotated scores. For the MOS prediction task, auditory LLMs are asked to assign the MOS score on a scale from 1... | https://arxiv.org/abs/2505.18601v1 |
3’s Answer] {answer_3} [Assistant 4’s Answer] {answer_4} <|im_end|> <|im_start|>assistant <think> Figure 8: System prompt for single-score and batch-level ranking evaluations. The part colorized in red denotes the additional instruction used only for batch-level ranking evaluation. •Pairwise Comparison: The judge selec... | https://arxiv.org/abs/2505.18601v1 |
be found in our provided code implementation. B.4 F LEX-Mol-LLaMA Judge In this section, we present the training details of FLEX-Mol-LLaMA, a reasoning-augmented molec- ular judge model built on top of Mol-LLaMA [ 27]. Mol-LLaMA is a molecule-focused LLM, pretrained and fine-tuned on molecular understanding datasets. I... | https://arxiv.org/abs/2505.18601v1 |
that receive identically best scores, we repeat the sampling and selection process over 10 random trials and report the average performance. As shown in Figure 4 ( middle ), increasing Nconsistently improves accuracy, validating that FLEX-Mol-LLaMA provides reliable, fine-grained reward signals. B.4.2 DPO Training Beyo... | https://arxiv.org/abs/2505.18601v1 |
42.93 7B 0.332 0.538 0.655 0.426 46.11 43.39 62.87 43.32 47.41 44.78 C.2 Reliability of F LEX-Judge. Length Bias. In Chen et al. [11], models such as GPT-4V [ 1] and Gemini [ 59] tend to favor longer answers over concise yet correct ones, exhibiting a phenomenon known as verbosity bias [ 85]. In contrast, our FLEX-Omni... | https://arxiv.org/abs/2505.18601v1 |
ModelJudgeLM (GPT-4o as Ground-Truth) PandaLM (Human as Ground-Truth) Agreement Precision Recall F1 Agreement Precision Recall F1 GPT-3.5♠73.83 70.70 52.80 52.85 62.96 61.95 63.59 58.20 GPT-4♠- - - - 66.47 66.20 68.15 61.80 PandaLM-7B♠68.61 40.75 38.82 39.41 59.26 57.28 59.23 54.56 Auto-J-13B♠74.86 61.65 57.53 58.14 - ... | https://arxiv.org/abs/2505.18601v1 |
that is conducted to fill a sudden or unexpected vacancy.)Assistant 1:(छ) "आ,थ*क &नचोड़ से वह &नचोड़ 3ा4त है जो आ,थ*क ि$थ&त को और भी <वकट बना देता है. (ड) "आकि$मक &नवा*चन" से वह &नवा*चन अHभ3ेत है जो आकि$मक IरिJत को भरने के Hलए Mकया जाता है; Figure 13: Reasoning process of F LEX-Judge on the OCR task (VL-RewardBench). Use... | https://arxiv.org/abs/2505.18601v1 |
Assistant 2’s image shows blurs, which could be “due to movement or a low-quality camera”. Figure 15 shows an image editing task, where the judge assesses whether the edited image by each assistant aligns with the revised prompt. FLEX-Judge identifies that Assistant B’s output exhibits “overediting”, detracting from th... | https://arxiv.org/abs/2505.18601v1 |
several structural features which can influence its permeability in the PAMPA assay. [...]1. Lipophilicity: The presence of aromatic rings, such as the methoxyphenyl groups, can significantly enhance the lipophilicity of the molecule. Aromatic rings are known to increase the hydrophobic character, which can improve the... | https://arxiv.org/abs/2505.18601v1 |
because they are relatively stable and can help the molecule maintain its structure in the presence of enzymes that might otherwise break it down. This stability is important for the prolonged action of the molecule in the body, which is a typical characteristic of opioid analogues. [...]In contrast, the description pr... | https://arxiv.org/abs/2505.18601v1 |
RASMALAI : Resources for A daptive S peech M odeling in IndiA n Languages with A ccents and I ntonations Ashwin Sankar1, Yoach Lacombe2, Sherry Thomas1, Praveen Srinivasa Varadhan1, Sanchit Gandhi2, Mitesh M Khapra1 1AI4Bharat, WSAI, Indian Institute of Technology Madras, India 2Huggingface, France Abstract We introduc... | https://arxiv.org/abs/2505.18609v2 |
speech quality. The intended style is anger. (b)Jaya, a female speaker, delivers high- pitched, expressive speech in a slightly en- closed environment at a fast pace, with excel- lent quality and an angry tone. (c)Jaya’s angry tone, with a sharp voice, echoes with exceptional quality in a moder- ately reverberant envir... | https://arxiv.org/abs/2505.18609v2 |
To support fur- ther research in multilingual text-prompted TTS, we will release all models and code to the community. 2. R ASMALAI : An Annotated Corpus for Controllable Multilingual TTS Below we describe (i) existing TTS datasets from which {audio, text}pairs were collated (ii) our approach for generating text descri... | https://arxiv.org/abs/2505.18609v2 |
categorized attributes, and prompt it to generate three types of textual descriptions (see Table 1) - (i) Descriptive Prompt : a detailed summary covering all labeled attributes, (ii) Concise Prompt : a brief description of the sample, and (iii) Attribute- Robust Prompt : which omits selected attributes to improve mode... | https://arxiv.org/abs/2505.18609v2 |
values as comma-separated sequences and compare them to the original instruction from which this sample was synthesised. We also report attribute- level accuracy in Table 6, which is calculated directly by com- paring the binned values against the original attributes used to create the description. To evaluate emotion ... | https://arxiv.org/abs/2505.18609v2 |
iden- tity across generated utterances. To assess instruction adher- ence, we report IF-BLEU, which reaches a high score of 93.18, showing the model’s strong ability to follow given instructions accurately. Additionally, the attribute accuracy scores presented in Table 6 show strong performance in adhering to the spec-... | https://arxiv.org/abs/2505.18609v2 |
65.35 2.98 2.93 7.51 4.15 4.51 0.36 1.72 78.831020304050607080 Figure 2: Confusion plot for Perceptual Emotion Classification Table 7: MUSHRA scores for zero-shot expressive synthesis across three different speaker groups. Native Proximal Distal 86.73±2.10 80.86 ±3.28 76.01 ±5.25 5. Related Work Resources and Models fo... | https://arxiv.org/abs/2505.18609v2 |
Tang, X. Tan, Y . Liu, S. Zhao, and N. Kanda, “E2 TTS: embarrassingly easy fully non-autoregressive zero-shot TTS,” CoRR , vol. abs/2406.18009, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2406.18009 [5] Y . Leng, Z. Guo, K. Shen, Z. Ju, X. Tan, E. Liu, Y . Liu, D. Yang, leying zhang, K. Song, L. He, X. Li... | https://arxiv.org/abs/2505.18609v2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.