text string | source string |
|---|---|
of configurations to demonstrate the generality and practicality. Table 10 shows the detailed experimental configuration of the proposed method with Dream-7B [26] and LLada-8B [19]. Stochastic Guided Unmasking. Please note that during guided unmasking, we only consider the Top-K logits of the AR-guider model as the gui... | https://arxiv.org/abs/2505.21467v1 |
a **new line**, write exactly: "The final answer is [answer]" where **[answer]** is either **Ending1**, **Ending2**, **Ending3**, or **Ending4**. Do **not** output anything after that line. PIQA: Below is an instruction that describes a task. Write a response that appropriately completes the request. ### Instruction: <... | https://arxiv.org/abs/2505.21467v1 |
Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration Zijun Liu1*, Zhennan Wan1*, Peng Li2, Ming Yan3, Ji Zhang3, Fei Huang3, Yang Liu1,2 1Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University 2Institute for AI Industry Research (AIR), Tsinghua University3Tongyi Lab,... | https://arxiv.org/abs/2505.21471v1 |
compressors may discard subtle cues that are only useful once the reasoning chain unfolds. Recent approaches [ 26–28,1] let LLM-based agents collaborate to process long contexts distributedly, reaching state-of-the-art performance on long-context tasks. In this work, we take a step further by asking a question: Could L... | https://arxiv.org/abs/2505.21471v1 |
deployment. We also construct an enhanced long-context multi-hop QA benchmark, ∞Bench+ , for corresponding evaluation. •We systematically study existing LLM-based multi-agent systems for context window extension, and overcome their bottlenecks by proposing a novel framework, E XTAGENTS . •We demonstrate the effectivene... | https://arxiv.org/abs/2505.21471v1 |
and, as discussed in Appendix A, more sophisticated chunking methods are also available for future work; for knowledge bases C, chunk dicould be a retrieved document piece with further aggregation or splitting. The former is often used in QA tasks oriented to long documents [ 29], while the latter is common in open-dom... | https://arxiv.org/abs/2505.21471v1 |
i,t−1, . . . , a i+k2,t−1}of size |Gi,t−1|, with the maximum max i,t{|Gi,t|}termed bandwidth . Some original chunks from agents in Gi,t−1may also be included ( DGi,t−1,t⊆ {di|ai,t−1∈ Gi,t−1}). It then emits an updated information with single-turn prompting: mi,t=ai,t(q,DGi,t−1,t,MGi,t−1,t−1). (4) With smaller bandwidth... | https://arxiv.org/abs/2505.21471v1 |
whether the answer could be obtained in each 8k token segment of the document. If so, we discard the sample. The process has been performed on the En.QA subset of ∞Bench, and the average length of samples is reduced (Table 2), showing that some long samples from the original benchmark are biased. Besides, we also inclu... | https://arxiv.org/abs/2505.21471v1 |
its local context chunk into succinct messages and posts them onto a shared scratchpad accessible by all other agents. Unlike previous methods [ 27,1], which restrict agent interactions to local neighborhoods, our approach grants every agent global visibility, thus maximizing synchronization bandwidth and ensuring effi... | https://arxiv.org/abs/2505.21471v1 |
process is performed to fill in each section. In this case, after filling up a section, the newly started process will take the previous section into the task query for continuous generation. The structured separation of Seeking and Reasoning Agents in E XTAGENTS , combined with global synchronization and incremental k... | https://arxiv.org/abs/2505.21471v1 |
methods with the optimal configuration except for the input length and multi-agent chunk size. For stable reproduction, we report the median results of three runs. For HotpotQA, we use BM25 retriever [ 63], and for AutoSurvey, we use the original retrieval method. Other details are elaborated in Appendix E. 8 Table 4: ... | https://arxiv.org/abs/2505.21471v1 |
components are critical for achieving strong scalability and effective knowledge integration, breaking the bottleneck of information over- load. Removing GKS leads to a slight drop in performance, indicating that the total rounds of global knowledge synchronization is not the main bottleneck of EXTAGENTS . Compared to ... | https://arxiv.org/abs/2505.21471v1 |
https://arxiv.org/abs/2503.09516 . [11] Zhenrui Yue, Honglei Zhuang, Aijun Bai, Kai Hui, Rolf Jagerman, Hansi Zeng, Zhen Qin, Dong Wang, Xuanhui Wang, and Michael Bendersky. Inference scaling for long-context retrieval augmented generation. In The Thirteenth International Conference on Learning Representations , 2025. ... | https://arxiv.org/abs/2505.21471v1 |
https://arxiv.org/abs/2405.15318 . [23] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations , 2024. URL https://openreview.net/forum?id=NG7sS51zVF . [24] Yan Wang, Dongyang Ma, ... | https://arxiv.org/abs/2505.21471v1 |
Zhao, Gong Zhang, Sen Wang, Renhai Chen, Hua Xu, and Hongwei Sun. Xl3m: A training-free framework for llm length extension based on segment-wise inference. Computing Research Repository , arXiv:2405.17755, 2024. URL https: //arxiv.org/abs/2405.17755 . [34] Chaojun Xiao, Pengle Zhang, Xu Han, Guangxuan Xiao, Yankai Lin,... | https://arxiv.org/abs/2505.21471v1 |
agent collaboration. In First Conference on Language Modeling , 2024. URL https: //openreview.net/forum?id=XII0Wp1XA9 . [44] Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. GPTSwarm: Language agents as optimizable graphs. In Ruslan Salakhutdinov, Zico Kolter, Kath... | https://arxiv.org/abs/2505.21471v1 |
Marft: Multi-agent reinforcement fine- tuning. Computing Research Repository , arXiv:2504.16129, 2025. URL https://arxiv.org/abs/ 2504.16129 . [55] Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simpl... | https://arxiv.org/abs/2505.21471v1 |
Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nik... | https://arxiv.org/abs/2505.21471v1 |
Teboul, Jessica Zhong, Jian Jin, Jingyi Yang, Joe Cummings, Jon Carvill, Jon Shepard, Jonathan McPhie, Jonathan Torres, Josh Ginsburg, Junjie Wang, Kai Wu, Kam Hou U, Karan Saxena, Kartikay Khandelwal, Katayoun Zand, Kathy Matosich, Kaushik Veeraraghavan, Kelly Michelena, Keqian Li, Kiran Jagadeesh, Kun Huang, Kunal Ch... | https://arxiv.org/abs/2505.21471v1 |
Li, and Arlindo L. Oliveira. LumberChunker: Long-form narrative document segmentation. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Findings of the Association for Computational Linguistics: EMNLP 2024 , pages 6473–6486, Miami, Florida, USA, November 2024. Association for Computational Linguistics. do... | https://arxiv.org/abs/2505.21471v1 |
the region is how big? Answer: 14 million km Excerpts from retrieved documents: William Hodges RA (28 October 1744 6 March 1797) was an English painter. He was a member of James Cook’s second voyage to the Pacific Ocean, and is best known for the sketches and paintings of locations he visited on that voyage, including ... | https://arxiv.org/abs/2505.21471v1 |
become my heir, and re...serves for me the fate of Cardinals Caprara and Bentivoglio, who were poisoned...I declare to my nephew, Guido Spada, my sole heir, that I have bu...ried in a place he knows and has visited with me, that is, in...the caves of the small Island of Brayan Annabel all I poss...ssed of ingots, gold,... | https://arxiv.org/abs/2505.21471v1 |
overall 68 pages compared to 58 pages from AutoSurvey. A Subsection in Long Survey Generated by E XTAGENTS 1.6 Case Studies Demonstrating LLM Impact on Education The integration of Large Language Models (LLMs) into educational contexts has resulted in transformative changes, showcasing their potential to enhance teachi... | https://arxiv.org/abs/2505.21471v1 |
Moreover, LLMs have demonstrated their capability to contribute to informal learning situations through chat-based environments. Researchers have explored how learners utilize LLMs to seek answers outside traditional classroom boundaries, supporting knowledge acquisition and empowering students to take charge of their ... | https://arxiv.org/abs/2505.21471v1 |
Studies Demonstrating LLM Impact on Education Large Language Models (LLMs) have taken center stage in educational innovation, demon- strating significant potential to enhance teaching and learning experiences across various contexts. This subsection presents a variety of case studies that illustrate successful deploy- ... | https://arxiv.org/abs/2505.21471v1 |
learners in problem-solving by exposing them to different viewpoints and areas of expertise, enriching the collaborative learning experience. Such interactive dynamics can enhance comprehension and engagement among students, urging creative exploration of subject matter [33]. Implementing LLMs within existing education... | https://arxiv.org/abs/2505.21471v1 |
we explain how existing methods synchronize knowledge across agents according to Equation (4). For Chain of Agents , each agent ai, tincorporate the message from previous agent ai−1,t−1in a linear topology, and the message is passed to the next agent ai+1,t−1in the next timestep. So in Equation (4),D(CoA) Gi,t−1,t={di}... | https://arxiv.org/abs/2505.21471v1 |
temperature was set to 0.1. The maximum input context length was 128,000 tokens for the closed- source models and 131,092 tokens for the open-source models. Baseline implementations are adjusted slightly for each task. For En.QA and Zh.QA test sets, the direct input method is implemented using the official InfiniteBenc... | https://arxiv.org/abs/2505.21471v1 |
Figure 5 with the same subfigure arrangement. Results on Weaker & Stronger LLMs We also test the performance of EXTAGENTS on a weaker LLM, Llama-3.2-3B-Instruct, besides a stronger LLM, gpt-4o-2024-08-06, on HotpotQA benchmark. The results are shown in Figure 9 and Figure 6, respectively. We observe that EXTAGENTS achi... | https://arxiv.org/abs/2505.21471v1 |
responsible for one chunk. This is the {iteration} round of Q&A. And we have the previously extracted information from all chunks in the previous round. Based on the previously extracted information and question, extract new information from the chunk. Do not repeat the previously extracted information. Your chunk: {Co... | https://arxiv.org/abs/2505.21471v1 |
arXiv:2505.21472v1 [cs.CV] 27 May 2025Mitigating Hallucination in Large Vision-Language Models via Adaptive Attention Calibration Mehrdad Fazli, Bowen Wei, Ziwei Zhu Department of Computer Science, George Mason University Fairfax, V A 22030, USA {mfazli, bwei2, zzhu20}@gmu.edu Abstract Large vision-language models (LVL... | https://arxiv.org/abs/2505.21472v1 |
in the image. Both biases significantly amplify the risk of hallucination in long-form generations. To tackle these issues, we propose Confidence- Aware Attention Calibration (CAAC), a unified inference-time approach to mitigate hallucinations by dynamically recalibrating the LVLM’s atten- tion. CAAC uses the model’s t... | https://arxiv.org/abs/2505.21472v1 |
2024; Liu et al., 2024c; Gong et al., 2024; Woo et al., 2024). Our method builds on the insights derived from the previous works but introduces an adaptive intervention based on the model’s confi- dence in predicting the next token. 3 Proposed Method What causes LVLMs to describe objects or scenes absent from an image ... | https://arxiv.org/abs/2505.21472v1 |
this limitation, we leverage relevancy maps (Chefer et al., 2021), which propagate token- level contributions layer by layer, ultimately quan- tifying the influence of each input token on the generation of each output token. By adopting this more principled analysis, our work revisits and rein- terprets previous findin... | https://arxiv.org/abs/2505.21472v1 |
suggests a distinct generation dynamic between truthful and hallucinatory tokens: the model hallu- cinates when its confidence is low and its atten- tion to the image has diminished . 3.4 CAAC Framework Our CAAC framework addresses two distinct bi- ases operating in different dimensions within the LLM decoder. Spatial ... | https://arxiv.org/abs/2505.21472v1 |
cal,0by the ratio of the sum of Vh,lto the sum of Vh,l cal,0. The final calibration vector is thus: Vh,l cal=PNi i=1viPNi i=1(1/vi)·Vh,l cal,0, (4) wherePNi i=1viis the sum of the original atten- tion weights, andPNi i=1(1/vi)is the sum of the initial inverted weights. Note that the product of Vh,landVh,l calresults is... | https://arxiv.org/abs/2505.21472v1 |
applied when the model is fully confident ( p= 1), while λmaxsets the upper bound for scaling when con- fidence is minimal ( p= 0). As pdecreases, λ increases, amplifying attention to image tokens precisely when hallucination risk is highest. Application of AAR : As AAR is bound to change the sum of the row it is appli... | https://arxiv.org/abs/2505.21472v1 |
CHAIR benchmark (Rohrbach et al., 2019) evaluates object hallucination in image captioning by measuring, for a given input image and a corresponding caption, the fraction of hal- 6 Table 1: Performance on CHAIR Benchmark Model CHAIRs CHAIRi Recall Len LLaV A 55.2 17.6 73.8 103.9 + OPERA 44.6 12.8 79.2 + VCD 57.8 16.3 7... | https://arxiv.org/abs/2505.21472v1 |
com- petitive Accuracy and F1 scores across all settings and for both LLaV A and InstructBLIP. These re- sults highlight CAAC’s effectiveness in mitigat- ing hallucinations beyond its generative focus, out- performing or matching baseline methods, thus demonstrating its versatility and robustness. 4.3 Ablation Study To... | https://arxiv.org/abs/2505.21472v1 |
incoherence or truncated sequences observed with full-layer ap- plication. The smoothing parameter βwas found to be very impactful. Large values of β(≥0.9) often resulted in impaired generation sequences. However, intermediate values for β,0.3∼0.7, re- sulted in coherent and high-quality responses. A comprehensive anal... | https://arxiv.org/abs/2505.21472v1 |
this approach remains computation- ally efficient compared to methods requiring train- ing post-hoc hallucination correction modules or fine-tuning the entire model, a common practice in baseline approaches. Given that our work fo- cuses on inference-time intervention rather than training, this tuning requirement is a ... | https://arxiv.org/abs/2505.21472v1 |
Erhan Bas. 2024. De- tecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Con- ference on Artificial Intelligence , volume 38, pages 18135–18143. Issue: 16. Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. OPER... | https://arxiv.org/abs/2505.21472v1 |
Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Jiaqi Wang, Haiyang Xu, Ming Yan, Ji Zhang, and Jitao Sang. 2024. AMBER: An LLM-free Multi-dimensional Benchmark for MLLMs Hallucination Evaluation. arXiv preprint . ArXiv:2311.07397 [cs]. Sangmin Woo, Donguk Kim, Jaehyuk Jang, Yubin Choi, and Changick Kim. 2024. Don’t... | https://arxiv.org/abs/2505.21472v1 |
GPUs and 512 GB of system RAM. We evaluate models in 16 -bit floating -point preci- sion using HuggingFace transformers 4.47 . A complete AMBER run (512 max -token setting) re- quires ∼12 hours for InstructBLIP and ∼10 hours for LLaV A-1.5. CAAC hyper -parameters. We use the follow- ing values, selected via the grid se... | https://arxiv.org/abs/2505.21472v1 |
answering, vi- sual reasoning, etc. (Xu et al., 2025) C.2 Hallucination in LVLMs Hallucination in Large Vision-Language Models (LVLMs) refers to the generation of responses that are not factually aligned with the visual input, such as describing objects absent from the image or mis- interpreting visual content (Guan et... | https://arxiv.org/abs/2505.21472v1 |
due to significant changes in attention distribution causing informa- tion loss in later layers. We thus examined the effect of varying the number of layers, from the first 2 to all 32 decoder layers. The best perfor- mance, with minimal hallucination rates, was ob- served when VTC was applied to the first 10 layers, a... | https://arxiv.org/abs/2505.21472v1 |
arXiv:2505.21479v1 [cs.CL] 27 May 2025Are Language Models Consequentialist or Deontological Moral Reasoners? Keenan Samway1,*, Max Kleiman-Weiner2,*, David Guzman Piedrahita3, Rada Mihalcea4,Bernhard Schölkopf1,Zhijing Jin1,5,6 1Max Planck Institute for Intelligent Systems, Tübingen,2University of Washington, 3ETH Züri... | https://arxiv.org/abs/2505.21479v1 |
2024; Lakkaraju et al., 2023). Understanding how LLMs reason through ethically complex sce- narios is critical for ensuring their safe and respon- sible deployment in high-stakes applications. For the purpose of our study, we use the term moral rea- soning for a process through which agents, human or artificial, naviga... | https://arxiv.org/abs/2505.21479v1 |
paper as follows: 1.Measurement of language models’ reasoning processes when they are presented with trolley dilemmas and asked to make a forced choice. 2.A taxonomy of 16 morally relevant rationales grounded in the ethical theories of consequen- tialism and deontology. This taxonomy can be used as a target for the cla... | https://arxiv.org/abs/2505.21479v1 |
which examines the nature and origin of moral principles; normative ethics, which develops frameworks for determining right and wrong actions; descriptive ethics, which studies how people actually behave and form moral beliefs; and applied ethics, which applies ethical theories to real-world domains. In this study, we ... | https://arxiv.org/abs/2505.21479v1 |
morally relevant rationales included in our taxonomy. See Table 5 for descriptions of each rationale and see Appendix B for examples of full model responses corresponding to each rational. Our taxonomy distinguishes between consequen- tialist and deontological rationales based on their orientation toward ethical action... | https://arxiv.org/abs/2505.21479v1 |
reasoning pathways that a model might follow. For reasoning LLMs, we use the developers’ recommended tem- 4 perature and specify the value for each model in Appendix C.2. If a model refuses to respond (e.g., “As an AI language model, I cannot. . . ”) or, more often, if a model does not respond using our speci- fied for... | https://arxiv.org/abs/2505.21479v1 |
(Liang et al., 2023) in our analysis. We list the models that we include in such examinations in Appendix C.2. 5 Experimental Findings 0.75 0.50 0.25 0.00 0.25 0.50 0.75 CD GapGemini Pro 1.5Qwen 3 32B FP8GPT-3.5 Turbo 1106Qwen 2.5 14BGPT-3.5 Turbo 0125Llama 4 ScoutQwen 2.5 72BLlama 3.3 70BQwen 1.5 32BGemini 1.5 Flash 8... | https://arxiv.org/abs/2505.21479v1 |
how moral reasoning patterns change as general reasoning capabilities improve in mod- els. When examining the relationship between MMLU performance and CDGAP across all scenar- ios, no significant correlation emerges (Pearson rr=−0.233, p= 0.2156 ). However, when group- ing the size-balanced and size-imbalanced scenar-... | https://arxiv.org/abs/2505.21479v1 |
reasoning models that all reveal their reasoning tokens to the end-user (Qwen 3 32B T,3Qwen 3 30B A3B T, QwQ 32B, DeepSeek R1, DeepSeek R1 Distill Llama 70B and DeepSeek R1 Distill Llama 8B) and pair them with their most similar traditional model counterpart (Qwen 3 32B, Qwen 3 30B A3B, Qwen 2.5 32B, DeepSeek V3, Llama... | https://arxiv.org/abs/2505.21479v1 |
human preference in these models may rein- force rule-based ethical considerations. However, this pattern is not universal, as the smallest model we test exhibits the opposite trend, with the DPO model becoming more consequentialist. These pat- terns suggest the impact of alignment techniques are likely influenced by s... | https://arxiv.org/abs/2505.21479v1 |
tive, instead dynamically adapt their ethical frame- work based on situational factors. These results have important implications for AI safety and alignment, especially as language models are increasingly deployed in high-stakes decision making environments. The distinction between consequentialist and deontological r... | https://arxiv.org/abs/2505.21479v1 |
to analyze LLM moral reasoning, we recognize that these represent just two perspectives centered in Western philosophical traditions. Our taxonomy may not fully capture moral rationales from diverse cultural contexts and ethical frame- works such as virtue ethics, care ethics, or non- Western philosophical systems. Pot... | https://arxiv.org/abs/2505.21479v1 |
cal report. CoRR , abs/2303.08774. Muhammad Shahrul Zaim bin Ahmad and Kazuhiro Takemoto. 2024. Large-scale moral machine ex- periment on large language models. arXiv preprint arXiv:2411.06790 . Larry Alexander and Michael Moore. 2024. Deonto- logical Ethics. In Edward N. Zalta and Uri Nodel- man, editors, The Stanford... | https://arxiv.org/abs/2505.21479v1 |
massive multitask language understanding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net. Evan Hubinger, Carson Denison, Jesse Mu, Mike Lam- bert, Meg Tong, Monte MacDiarmid, Tamera Lan- ham, Daniel M. Ziegler, Tim Maxwell, Newton Cheng, Ad... | https://arxiv.org/abs/2505.21479v1 |
Del Corro. 2024. The greatest good bench- mark: Measuring LLMs’ alignment with utilitarian moral dilemmas. In Proceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Pro- cessing , pages 21950–21959, Miami, Florida, USA. Association for Computational Linguistics. Mantas Mazeika, Xuwang Yin, Rishub... | https://arxiv.org/abs/2505.21479v1 |
Yulia Tsvetkov, and Alane Suhr. 2024. Quantifying language models’ sensitiv- ity to spurious features in prompt design or: How I learned to start worrying about prompt formatting. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net. Marco Siino... | https://arxiv.org/abs/2505.21479v1 |
et al. (2023) in- vestigated LLMs’ approaches to moral scenarios of varying ambiguity, while Takemoto (2024); Ahmad and Takemoto (2024); Jin et al. (2025) used the Moral Machine framework to evaluate LLMs’ eth- ical decision-making. Unlike these studies whichfocus on moral judgments, our paper focuses on LLMs’ abilitie... | https://arxiv.org/abs/2505.21479v1 |
the non-human animal based on moral status distinctions. Others Other If none of the other rationales clearly apply. Table 5: Descriptions of each moral rationale in M ORAL LENS, as presented to the LLM judge. 4 boys vs. 4 elderly men Scratchpad: Given the lack of other information, the decision must be based on a util... | https://arxiv.org/abs/2505.21479v1 |
lived a significant portion of their lives. However, if we consider the potential for future contributions, we might lean towards saving the individual who has more years ahead of them, or who may have dependents or responsibilities. Without additional context about their lives, it is difficult to make a definitive cho... | https://arxiv.org/abs/2505.21479v1 |
of sacrificing one group to save another. The only distinguishing factor between the groups is their socio-economic status: homeless individuals versus executives. Since the number of individuals in each group is the same, and in the absence of any other differentiating information (age, health, potential future contri... | https://arxiv.org/abs/2505.21479v1 |
our exper- imentation, we found that even if we presented models with ‘Decision:...’ first and ‘Scratchpad:...’ after, they still overwhelmingly responded by plac- ing ‘Scratchpad:...’ first. We suspect that this term in particular triggers models to think step-by-step before responding, and therefore in our decision f... | https://arxiv.org/abs/2505.21479v1 |
attempts per model for the Reason-then-Decide and Decide-then-Reason scenarios respectively. Observe that many models fully respond after the first one or two attempts, however certain models are much more stubborn in their responses. In our experimentation, we noticed that subtle prompt differences can in turn have a ... | https://arxiv.org/abs/2505.21479v1 |
100 100 100 100 100 100 100 100 99 100 100 100 100 100 100 100 100 100 99 100 100 100 100 100 100 100 100 100 99 100 100 100 100 100 100 100 100 100 99 100 100 100 100 100 100 100 100 100 88 100 100 100 100 100 100 100 100 100 99 99 100 100 100 100 100 100 100 100 99 99 100 100 100 100 100 100 100 100 98 99 100 100 100... | https://arxiv.org/abs/2505.21479v1 |
rated eight responses and achieved a precision of 0.812, recall of 0.896, and F1 score of 0.833. Inter-annotator agreement (Co- hen’s kappa) among the eight overlapping scenarios was0.808. 19 1 2 3 4 5 6 7 8 9 10 Attempt NumberQwen 1.5 14B Qwen 1.5 7B DeepSeek V3 FP8 Gemini Pro 1.5 Gemma 3 4B GPT-3.5 Turbo 0125 GPT-4o ... | https://arxiv.org/abs/2505.21479v1 |
100 100 100 100 100 100 100 88 98 99 100 100 100 100 100 100 100 97 99 99 99 100 100 100 100 100 100 97 99 99 99 100 100 100 100 100 100 99 99 99 99 99 100 100 100 100 100 99 99 99 99 99 99 100 100 100 100 98 99 99 99 99 99 100 100 100 100 95 99 99 99 99 99 100 100 100 100 98 99 99 99 99 99 99 99 99 100 99 99 99 99 99 ... | https://arxiv.org/abs/2505.21479v1 |
computational cost here to be approximately 200GPU hours. AI Assistants Our implementation leveraged AI assistants for developing our codebase. All AI- generated code was reviewed, tested, and validated by the authors to ensure correctness and repro- ducibility. D Additional Results D.1 Reasoning Order Analysis Here we... | https://arxiv.org/abs/2505.21479v1 |
pre-decision (right) reasoning. Pre-decision reasoning appears to increase the proportion of deonto- logical rationales compared to consequentialist ones. Decide-then-Reason Reason-then-Decide0.20.30.40.50.60.70.80.91.0Utility Pre-Decision Reasoning Increases Utility Figure 11: Plots the UTILITY per model for Decide- t... | https://arxiv.org/abs/2505.21479v1 |
all models consistency versus their util- ity in reasoning first scenarios. We find that consistency is highly correlated with utility, which intuitively makes sense—as for a model to have high utility, it cannot be making inconsistent decisions that result it in choosing to save an individual over a group just as a re... | https://arxiv.org/abs/2505.21479v1 |
arXiv:2505.21523v1 [cs.CL] 23 May 2025More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models Chengzhi Liu*1,3, Zhongxing Xu*1, Qingyue Wei2, Juncheng Wu1, James Zou2, Xin Eric Wang1,3, Yuyin Zhou1, Sheng Liu2 1UC Santa Cruz,2Stanford University,3UC Santa Barbara chengzhi@ucsb.edu, ... | https://arxiv.org/abs/2505.21523v1 |
of outputs from a reasoning model and a non-reasoning model on a perception task. Red highlights indicate visual hallucination. Multimodal reasoning models are generally more prone to amplifying hallucinations during the reasoning process compared to their non-reasoning counterparts. (b)Performance of different models ... | https://arxiv.org/abs/2505.21523v1 |
deficit remains consistent at both the 3 B and 7 B scales, demonstrating that the elevated hallucination rate stems from the reasoning paradigm itself rather than model size. Bingo Score MMVP MMEval-Pro VMCBenchMMHALU3.64 3.33 77.6 67.836.0 3.523.58 28.7 24.0 60.2 66.973.0 76.02.78 2.27 Bingo Score MMVP MMEval-Pro VMCB... | https://arxiv.org/abs/2505.21523v1 |
of these explicitly mention cups. However, since it‘s a juice stall, it’s likely that cups would be present for serving the juice. Even though the image doesn‘t show any cups directly, th e context strongly suggests their presence. </think>\n\n<Answer> Yes, there is a cup in the image.</Answer> (a) (b)Figure 4: Two com... | https://arxiv.org/abs/2505.21523v1 |
further diminished attention to visual tokens, with increased focus directed toward the end of the user instruction. This suggests that longer reasoning chains tend to further exacerbate the degradation of the model’s visual grounding, potentially leading to an increase in hallucinations. Is there a gray wall in the im... | https://arxiv.org/abs/2505.21523v1 |
the difference between the long and short embeddings, denoted as dℓ, which captures the variation in the model’s representation resulting from different reasoning chain lengths. To adjust the hidden representation based on this direction, We introduce a parameter α∈[−0.15,0.15]to dynamically control the reasoning lengt... | https://arxiv.org/abs/2505.21523v1 |
increased hallucination, yet both scenarios could yield the same average score. To capture this evolving balance, in the next section, we propose an AUC-style metric that summarizes the balance curve between reasoning and perception fidelity across various reasoning lengths. This provides a more faithful and holistic m... | https://arxiv.org/abs/2505.21523v1 |
perceptual hallucination. To compute the area under this curve, we first sort the pairs in ascending order of reasoning performance RT. Let 7 the sorted indices be denoted as T(0), T(1), . . . , T(n−1), such that RT(0)≤RT(1)≤ ··· ≤ RT(n−1). To ensure comparability across models, both RTandHTare min-max normalized to th... | https://arxiv.org/abs/2505.21523v1 |
compared to SFT+RL. This phenomenon suggests that although SFT helps the model learn reasoning formats, it may introduce rigid imitation reasoning paths, limiting the model’s adaptability to dynamic tasks and ultimately resulting in redundant reasoning. In contrast, RL encourages the model to generate more adaptive rea... | https://arxiv.org/abs/2505.21523v1 |
of physics and commonsense reasoning from visual inputs. Reinforcement Learning in MLLMs. Recent approaches enhance the reasoning capabilities of multimodal large models by incorporating chain-of-thought supervision during supervised fine- tuning or reinforcement learning [ 55,51,41,34,43,45]. Methods like RLHF-V [ 47]... | https://arxiv.org/abs/2505.21523v1 |
Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, 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 L... | https://arxiv.org/abs/2505.21523v1 |
Yusheng Zhao, Bohan Wu, Ye Yuan, Haozhe Zhao, Zhihui Guo, Yichi Zhang, et al. Mmevalpro: Calibrating multimodal benchmarks towards trustworthy and efficient evaluation. arXiv preprint arXiv:2407.00468 , 2024. [12] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Neng... | https://arxiv.org/abs/2505.21523v1 |
Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling, 2025. [26] OpenAI. Learning to reason with LLMs. 2024. [27] Runqi Qiao, Qiuna Tan, Guanting Dong, Minhui Wu, Chong Sun, Xiaoshuai Song, Zhuoma Gongque, Shanglin Lei, Zhe Wei, Miaoxuan Zhang, Runfeng Qiao, Yifan Zhang,... | https://arxiv.org/abs/2505.21523v1 |
Yu, Peng Zhang, Hao Jiang, et al. Fast-slow thinking for large vision-language model reasoning. arXiv preprint arXiv:2504.18458 , 2025. [41] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-o1: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440 , 2024. [42] Guowei Xu... | https://arxiv.org/abs/2505.21523v1 |
arXiv:2505.21527v1 [eess.AS] 23 May 2025VietASR: Achieving Industry-level Vietnamese ASR with 50-hour labeled data and Large-Scale Speech Pretraining Jianheng Zhuo1,2, Yifan Yang1, Yiwen Shao2, Yong Xu2, Dong Yu2, Kai Yu1, Xie Chen1,† 1X-LANCE Lab, School of Computer Science, MoE Key Lab of Artificial Intelligence, Sha... | https://arxiv.org/abs/2505.21527v1 |
1 billion parameters, posing challenges for de- ployment and accessibility. As for self-supervised models with low-resource language capabilities, they often have overwhelm- ing numbers of parameters and use waveforms as the front-end, making them difficult to deploy as practical speech recognition models. GigaSpeech 2... | https://arxiv.org/abs/2505.21527v1 |
[14–20] and training efficiency [21–23]. To better align pre-training with the downstream ASR task, sev- eral studies [14, 17–19] have explored ASR-biased HuBERT, where the learning target is derived from a supervised model rather than being purely self-supervised. PBERT [14] intro- duces phoneme-level alignment with a... | https://arxiv.org/abs/2505.21527v1 |
MLM pre- training is applied to the high-dimensional features. However, Zipformer uses Fbank as the model input, which is a multi- dimensional feature that already contains local information, so we apply masking to the Fbank features directly as in [22]. 3.1.3. Loss Function Since the original HuBERT prediction functio... | https://arxiv.org/abs/2505.21527v1 |
Experiments 4.1. Experimental Setups Dataset We collect approximately 70,000 hours of unlabeled audio from YouTube and apply FunASR V AD [29] for auto- matic segmentation. 50-hour audio from YouTube is adopted for fine-tuning, which is manually transcribed by professionals. Model evaluation is conducted on three public... | https://arxiv.org/abs/2505.21527v1 |
Google USM [8], MMS L1107 [9], Azure Speech CLI, and the 68 M and 152 M model from Gi- gaSpeech 2.0 [10]. We adopted Azure Speech CLI of version 1.37, and Chirp Speech-to-Text v2 model For Google USM. The results are presented in Table 3. Compared to the non- pretrained Zipformer, VietASR achieves significant improve- ... | https://arxiv.org/abs/2505.21527v1 |
VietASR can reduce the average WER by 8.0% when compared with Hu- BERT. With only a small additional cost for training the labelextractor, VietASR achieves a significant performance improve- ment over HuBERT. Table 4: Comparison between HuBERT codebook and VietASR codebook. The average WER (Avg) is weighted by the word... | https://arxiv.org/abs/2505.21527v1 |
al. , “Recent advances in end-to-end automatic speech recognition,” APSIPA Transactions on Signal and Information Processing , 2022. [4] A. Radford, J. W. Kim, T. Xu et al. , “Robust speech recognition via large-scale weak supervision,” in Proc. ICML , 2023. [5] K. C. Puvvada, P. ˙Zelasko, H. Huang et al. , “Less is mo... | https://arxiv.org/abs/2505.21527v1 |
J. Zhuo, Z. Jin et al. , “k2SSL: A faster and better frame- work for self-supervised speech representation learning,” arXiv preprint arXiv:2411.17100 , 2024. [24] A. Waswani, N. Shazeer, N. Parmar et al. , “Attention is all you need,” in Proc. NeurIPS , 2017. [25] A. Gulati, J. Qin, C.-C. Chiu et al. , “Conformer: Conv... | https://arxiv.org/abs/2505.21527v1 |
How Much Do Large Language Models Know about Human Motion? A Case Study in 3D Avatar Control Kunhang Li1Jason Naradowsky1Yansong Feng2Yusuke Miyao1,3 1The University of Tokyo2Peking University3NII LLMC {kunhangli, narad, yusuke}@is.s.u-tokyo.ac.jp fengyansong@pku.edu.cn Abstract We explore Large Language Models (LLMs)’... | https://arxiv.org/abs/2505.21531v1 |
instructions with full coverage of basic movement 1https://unity.com/ 1arXiv:2505.21531v1 [cs.CV] 23 May 2025 primitives and balanced body part usage, and evalu- ate both commercial (e.g., Claude 3.5 Sonnet) and open-source (e.g., Llama-3.1-70B) LLMs through three complementary approaches: (1) human eval- uation of hig... | https://arxiv.org/abs/2505.21531v1 |
et al. (2024) and Huang et al. (2024a) show that LLMs can directly generate keyframe coordinates to be interpolated as motions. We aim to investigate LLMs’ knowledge of human motion by designing a hierarchical framework that grounds LLM responses into 3D avatar animations, provid- ing clear visual verification of their... | https://arxiv.org/abs/2505.21531v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.