Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

VideoRewardBench: Comprehensive Evaluation of Multimodal Reward Models for Video Understanding

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read VideoRewardBench, a 1,563-preference-pair benchmark across four video-understanding dimensions, shows that even the strongest multimodal reward models—Gemini-2.5-Pro at 63.6% and LLaVA-Critic-72B at 63.0%—pick the human-preferred answer bar

desk verdict A genuinely useful, much larger video reward-model benchmark, but the headline numbers need human calibration and a few reporting fixes before I'd trust the rankings. read the letter →

arxiv 2509.00484 v1 pith:CA6O5YWQ submitted 2025-08-30 cs.CV cs.AI

Zhihong Zhang , Xiaojian Huang , Jin Xu , Zhuodong Luo , Xinzhi Wang , Jiansheng Wei , Xuejin Chen This is my paper · ORCID
classification cs.CVcs.AI
keywords videorewardmodelsmultimodalunderstandingpreferencedatabenchmarkRLHFevaluationinference-timescalingframesamplingsafety
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper builds VideoRewardBench, a benchmark that grades how well multimodal reward models—the AI judges used to train, select, and evaluate video-language models—can tell a good answer from a bad one. It assembles 1,563 video-based preference pairs spanning perception, knowledge, reasoning, and safety, and runs 28 reward models across all three families: generative, discriminative, and semi-scalar. The headline result is that judging video answers is still unsolved: the best model, Gemini-2.5-Pro, agrees with human preferences only 63.6% of the time, and the strongest open model, LLaVA-Critic-72B, reaches 63.0%, while most open models stay below 60%. The paper also reports two practical findings: sampling multiple judgments at higher temperature and majority-voting lifts most model types at inference time, and feeding more video frames helps critic-trained generative judges far more than semi-scalar ones. The benchmark's purpose is to give the field a hard, contamination-free yardstick for video judging ability.

What carries the argument

The benchmark's construction pipeline is the load-bearing mechanism. Prompts are harvested from ten public video benchmarks, then filtered through staged quality gates: videos over 10 minutes are dropped; a strong model must fail the question without the video and Qwen2-VL-7B-Instruct must fail it with the video; for knowledge and reasoning, a leading proprietary model must answer some but not all of ten tries correctly; for safety, only prompts with an attack success rate above 50% survive. Responses come from multiple LVLMs, labeled as chosen/rejected by three annotators with majority voting and ties discarded, producing 1,563 (video, question, chosen response, rejected response) triplets.

What would settle it

Take a random subset of the 1,563 preference pairs, have a fresh panel of annotators re-judge them, and measure agreement with the released labels: if a substantial share flip, the benchmark's ground truth is unstable and the reported ranking is not. A second check: rerun the multi-stage filtering pipeline with a different filter model than Qwen2-VL-7B-Instruct; if the 28-model accuracy ranking changes materially, the headline gaps depend on one filter choice rather than on genuine judging difficulty.

Watch

Extended reading notes

Core claim

VideoRewardBench claims to be the first benchmark that evaluates multimodal reward models (MRMs) in the video domain at scale and across holistic dimensions: long-form and short-form perception, knowledge, reasoning, and safety. It contains 1,563 human-annotated preference triplets drawn from ten existing video datasets, and on it the paper evaluates 28 MRMs covering generative models (both plain LVLMs and critic-trained judges), discriminative scalar scorers, and semi-scalar critic-plus-score models. The central result is a ceiling on current judging ability: the strongest proprietary model, Gemini-2.5-Pro, reaches 63.6% overall accuracy, the leading open model LLaVA-Critic-72B reaches 63.0

Load-bearing premise

The preference labels are a fair ground truth: the automated filters that remove 'too easy' questions must not systematically distort which model families look good, and the human judgments—reported without an inter-annotator agreement statistic—must be trustworthy.

Editorial extensions

If this is right

  • No current video reward model—proprietary or open—is reliable enough to act as a sole judge of video response quality; at roughly 63% top accuracy, the best model picks the worse answer about a third of the time.
  • Video judging can be improved without retraining: sampling several judgments with temperature 1.0 and taking a majority vote raises accuracy for generative and semi-scalar MRMs, though not for discriminative scorers.
  • Reinforcement-learning-trained critics are not inherently better judges across modalities; some RL-trained models (R1-Reward at 39.0%, Flex-Judge at 34.6%) fall well below their own base models when moved from image/text to video.
  • Frame sampling strategy should follow model type: critic-trained generative judges gain roughly 11 accuracy points as frames increase from 1 to 64, while semi-scalar judges slightly decline with more frames.
  • Because the benchmark excludes publicly available video preference training data, it can serve as a contamination-controlled test for future video reward models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the consistent drop of image/text-trained RL critics on video suggests video judging may need video-native training signal rather than more reasoning time; the paper reports the drop but does not draw this prescription.
  • A practical deployment corollary the paper leaves implicit: practitioners can save compute by keeping frame counts low for semi-scalar judges without losing accuracy, and should spend extra frames only on critic-trained generative judges.
  • A testable extension: rerunning the multi-stage filter with a stronger or different filter model would show whether the 28-model rankings are stable or partly an artifact of the single fixed filter (Qwen2-VL-7B-Instruct).
  • The finding that temperature-1.0 sampling with majority voting flips previous negative results suggests other benchmark suites that reported failed inference-time scaling may have used sampling parameters that suppressed output diversity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces VideoRewardBench, a video-domain benchmark for evaluating multimodal reward models (MRMs). It contains 1,563 preference triplets (video-text prompt, chosen response, rejected response) spanning 1,482 unique videos and 1,559 questions, organized into four dimensions: perception (long/short form), knowledge, reasoning, and safety. The authors describe an AI-assisted, multi-stage curation pipeline, then evaluate 28 MRMs of three types (generative, discriminative, semi-scalar) under pairwise or pointwise protocols. Headline results are that even the best model (Gemini-2.5-Pro in Table 4) reaches only 63.6% overall accuracy, most open models fall below 60%, RL-trained critic MRMs generalize worse than SFT-trained ones in some cases, and inference-time scaling and frame-count effects vary by MRM type. The dataset and evaluation code are promised for release.

Significance. If the benchmark labels are reliable and the curation pipeline is unbiased, VideoRewardBench would be a useful and much-needed resource: it is substantially larger than prior video reward-model benchmarks, covers four dimensions including safety, and evaluates all three MRM paradigms. The paper also provides falsifiable, non-obvious findings about RL-trained MRMs and inference-time scaling, and it ships concrete evaluation protocols and model cards. However, the benchmark's evidential value depends critically on label quality and on the independence of the filtering procedure from the evaluated models. The current manuscript does not supply a human baseline or inter-annotator agreement, and it contains a direct contradiction between the abstract and the main results table, so the headline conclusions are not yet calibrated.

major comments (5)
  1. [Abstract and §4.3, Table 4] The abstract states that 'even the top-performing model GPT-4o achieves only 57.0% overall accuracy' and that Qwen2.5-VL-72B is the state-of-the-art open-source model at 53.3%. Table 4 and Section 4.3 report the opposite: Gemini-2.5-Pro is best at 63.6%, Claude-3.7-Sonnet is 63.2%, and LLaVA-Critic-72B is 63.0%, while GPT-4o is 57.0% and Qwen2.5-VL-72B is 53.3%. The abstract's 'top-performing' claims are factually wrong and must be corrected, as they are the paper's headline.
  2. [§3.3.1, §3.3.4, §4.3; Table 4] No human baseline or inter-annotator agreement statistic is reported. The construction uses human majority votes for long-form perception and safety (three annotators), but the paper never reports agreement rates, annotator counts per item beyond 'three', or human accuracy on the final 1,563 pairs. Without this calibration, the statement that 63.6% represents 'moderate' or 'challenging' performance is uninterpretable: low observed accuracies could equally reflect label noise, ambiguous pairs, or selection bias in the multi-stage filtering. A human-judge accuracy on the final benchmark and an inter-annotator agreement measure (e.g., Fleiss' kappa) are needed to support the paper's difficulty claims.
  3. [§3.3.2, §3.3.3, §3.3.4] The preference labels are substantially model-assisted, and the manuscript does not quantify the resulting label noise. Short-form perception labels are taken directly from source-benchmark ground truth; knowledge/reasoning chosen/rejected responses are selected by matching a proprietary model's final answer to ground truth; safety labels depend on RJScore to classify responses as successfully or unsuccessfully attacked before human voting. Errors in any of these automatic steps propagate directly into the benchmark and cap achievable MRM accuracy. The paper should report a human-validation subset (e.g., re-label a random sample of each dimension) or otherwise quantify agreement between the constructed labels and independent human judgment.
  4. [§3.2.2] The multi-stage filtering uses Qwen2-VL-7B-Instruct to remove prompts it can answer correctly, and an unspecified 'strong model' to remove text-only-answerable prompts. Since several evaluated models are from the Qwen family (Qwen2.5-VL-7B/72B, Skywork-VL Reward based on Qwen2.5-VL-7B) or share training data with Qwen models, the filtering stage can differentially distort the prompt distribution for or against particular model families. The paper's claim of a fair, comprehensive comparison requires evidence that this filtering does not systematically favor or penalize specific MRM categories. At minimum, report per-source retention rates and an analysis of which model families are affected; ideally, ablate the Qwen-based filtering stage.
  5. [Table 4 and §4.2.3] All accuracies are reported as point estimates with no confidence intervals or significance tests. With 1,563 samples, the standard error of a 63% accuracy is about 1.2 percentage points, so differences such as Gemini-2.5-Pro (63.6%) versus Claude-3.7-Sonnet (63.2%) versus LLaVA-Critic-72B (63.0%) are within noise; the 'clear performance gap' narrative is overstated. Also, Phi-3.5-Vision's short-form perception accuracy of 7.75% (32/413) is far below any plausible chance level for binary/multiple-choice questions and suggests an evaluation or extraction error. The authors should re-check this result and report CIs for all cells.
minor comments (5)
  1. [§1 and Abstract] The abstract says the dataset and code are released at a URL, while Section 1 says 'We will release all data and evaluation code upon acceptance.' Please reconcile these statements.
  2. [§3.4 and Figure 2] The claim that the word-count-difference histogram 'indicates that our preference labels are primarily determined by response quality instead of response length' is too strong. A balanced length distribution is evidence against gross length bias, but it does not establish that labels are quality-driven rather than driven by other confounds. Please soften the wording.
  3. [§2.2] Minor typographical issues: 'REW ARDBENCH 2' should be 'RewardBench 2', and in Section 4.2 'convering' should be 'covering'.
  4. [Table 7] MiniCPM-o-2.6's row reads 'MiniCPM-o-2 6 HF' and should read 'MiniCPM-o-2.6'; also specify the exact video frame sampling for Gemini models ('1fps') more explicitly in the caption.
  5. [§5.1, Figures 3–4] Inference-time scaling plots are shown without error bars or multiple seeds. Given the small gains (e.g., Qwen2.5-VL-72B +0.9% from K=1 to K=9), it is unclear which differences are significant. Reporting bootstrap intervals or repeated runs would strengthen the analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark construction and evaluation are empirically grounded in external data and human annotation, with no fitted parameter or self-citation chain forcing the reported accuracies.

full rationale

VideoRewardBench is a benchmark-construction and evaluation paper, not a mathematical derivation. There is no equation in which a predicted quantity reduces to an input by construction, and no parameter is fitted to a subset of data and then renamed as a prediction. The preference labels come from external source benchmarks' ground-truth answers (short-form perception), human majority voting (long-form perception and safety), and ground-truth matching with human review of reasoning chains (knowledge/reasoning). The multi-stage filtering uses Qwen2-VL-7B-Instruct and RJScore to remove easy or non-discriminative prompts; this can introduce selection bias, but it does not determine the evaluated MRMs' judgments, and the reported accuracies are therefore not forced by the filter. No load-bearing self-citations are present: the authors do not cite their own prior work to justify a core premise, and no uniqueness theorem or ansatz is imported from the authors' other papers. The skeptical concern that no human baseline or inter-annotator agreement is reported is a validity/calibration issue, not a circularity issue; it does not demonstrate that any evaluation result is equivalent to its input. Following the reviewing rule, no passage in the manuscript asserts a limitation, missing support, missing reference, circular step, or omitted proof; the only near-limitation is the statement that data and code will be released upon acceptance, which is a release plan rather than a circularity admission. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The benchmark's validity rests on inherited ground truth, human annotation quality, the unbiasedness of AI-based filtering, and the comparability of evaluation protocols. These are domain assumptions rather than derived facts, and several are not empirically validated in the paper.

free parameters (4)
  • Video duration cutoff = 10 minutes
    Prompts with videos longer than 10 minutes are discarded in Section 3.2.2; this ad hoc cutoff affects which videos enter the benchmark and could bias toward short-video tasks.
  • Safety attack success rate threshold = >50%
    In Section 3.3.4, only safety prompts with attack success rate above 50% across six models are retained; this choice shapes the difficulty distribution of the safety dimension.
  • Filtering model for easy-question removal = Qwen2-VL-7B-Instruct
    Section 3.2.2 uses Qwen2-VL-7B-Instruct to identify and discard prompts answerable without video or by a small model; the choice of this specific model is ad hoc and may introduce family-specific bias.
  • Number of sampled responses for knowledge/reasoning = 10
    Section 3.3.3 samples 10 responses per prompt and discards prompts where all or none are correct; the number 10 is arbitrary and affects which prompts are kept.
assumptions (5)
  • domain assumption Ground-truth labels in the ten source benchmarks (MVBench, VideoHallucer, MMWorld, MMVU, Video-MMMU, Video-MME, MMBench-Video, VSI-Bench, VCGBench-Diverse, Video-SafetyBench) are correct.
    The benchmark inherits labels from these datasets and uses them as chosen responses or filtering criteria (Sections 3.2.1, 3.3.2, 3.3.3).
  • domain assumption Human annotators can reliably identify the better of two responses, and majority voting with three annotators is sufficient.
    Preference labels for long-form perception, knowledge, reasoning, and safety depend on human annotation with no reported inter-annotator agreement statistic (Sections 3.3.1-3.3.4).
  • domain assumption Qwen2-VL-7B-Instruct's ability to answer a question is a valid proxy for question easiness.
    Multi-stage filtering removes prompts this model answers correctly, but no validation is given that this proxy does not distort the difficulty or topic distribution of the benchmark (Section 3.2.2).
  • domain assumption The RLHF/DPO/reward-model training data for the evaluated MRMs does not overlap with the source videos and questions in VideoRewardBench.
    The paper states it excludes publicly available video preference data used for training (Section 1), but contamination from the source benchmarks themselves (e.g., MVBench being in an MRM's training corpus) is not analyzed.
  • domain assumption Pairwise ranking (for generative MRMs) and pointwise scoring (for discriminative/semi-scalar MRMs) provide comparable and valid measures of reward-model quality.
    The evaluation protocol in Sections 4.2.1-4.2.2 assumes these two inference paradigms can be compared on identical preference pairs without introducing protocol-specific bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VideoRewardBench: Comprehensive Evaluation of Multimodal Reward Models for Video Understanding." pith.science (2026). https://pith.science/paper/CA6O5YWQ

@misc{pith2026250900484,
  author       = {Pith},
  title        = {Pith review of: VideoRewardBench: Comprehensive Evaluation of Multimodal Reward Models for Video Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CA6O5YWQ}},
  note         = {Machine review of arXiv:2509.00484}
}
read the original abstract

Multimodal reward models (MRMs) play a crucial role in the training, inference, and evaluation of Large Vision Language Models (LVLMs) by assessing response quality. However, existing benchmarks for evaluating MRMs in the video domain suffer from a limited number and diversity of questions, a lack of comprehensive evaluation dimensions, and inadequate evaluation of diverse types of MRMs. To address these gaps, we introduce VideoRewardBench, the first comprehensive benchmark covering four core aspects of video understanding: perception, knowledge, reasoning, and safety. Through our AI-assisted data pipeline, we curate a high-quality preference dataset of 1,563 annotated samples, including 1,482 unique videos and 1,559 distinct questions--15 times the number found in the most question-rich prior benchmark. Each sample is a triplet consisting of a video-text prompt, a chosen response, and a rejected response. We also conduct a comprehensive evaluation across 28 multimodal reward models spanning three categories: generative, discriminative, and semi-scalar. Results show that even the top-performing model GPT-4o achieves only 57.0% overall accuracy, and the state-of-the-art open-source model Qwen2.5-VL-72B reaches merely 53.3%. Our analysis further reveals three key insights: (i) MRMs trained with reinforcement learning (RL) do not necessarily exhibit stronger cross-modal generalization than those trained without RL; (ii) except for discriminative MRMs, other types of MRMs across varying model capacities can benefit from inference-time scaling; and (iii) variations in input video frame count have different effects on different types of MRMs. We believe VideoRewardBench offers a challenging and valuable benchmark for advancing the evaluation and development of MRMs in the video domain.

Figures

Figures reproduced from arXiv: 2509.00484 by the authors.

Figure 1
Figure 1. Comparing mainstream multimodal reward models on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution of the word count difference between the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Inference-time scaling results for different MRMs. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Further results on inference-time scaling. (a) Left: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Effect of input frame count on six MRMs. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: An example from the long-form perception dimension. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: An example from the short-form perception dimension. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: An example from the knowledge dimension. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: An example from the reason dimension [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: An example from the safety dimension. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Social Caption: Evaluating Social Understanding in Multimodal Models

    cs.CL 2026-01 conditional novelty 6.0 of 10

    A three-part benchmark for multimodal LLMs separates question-answering about social videos from holistic and directed scene description, and shows AI judges can approximate human quality ratings.

Reference graph

Works this paper leans on

57 extracted references · 56 canonical work pages · cited by 1 Pith paper

  1. [1]

    Phi-3 technical report: A highly capable lan- guage model locally on your phone, 2024

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadal- lah, and et al. Phi-3 technical report: A highly capable lan- guage model locally on your phone, 2024. 6

    work page 2024

  2. [2]

    Claude-3.7-sonnet

    Anthropic. Claude-3.7-sonnet. https : / / www . anthropic . com / news / claude - 3 - 7 - sonnet,

    work page

  3. [3]

    Qwen2.5-vl technical report, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2025. 6

    work page 2025

  4. [4]

    Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark

    Dongping Chen, Ruoxi Chen, Shilin Zhang, Yaochen Wang, Yinuo Liu, Huichi Zhou, Qihui Zhang, Yao Wan, Pan Zhou, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. In Forty- first International Conference on Machine Learning , 2024. 3

    work page 2024

  5. [5]

    From captions to rewards (carevl): Leveraging large language model experts for en- hanced reward modeling in large vision-language models,

    Muzhi Dai, Jiashuo Sun, Zhiyuan Zhao, Shixuan Liu, Rui Li, Junyu Gao, and Xuelong Li. From captions to rewards (carevl): Leveraging large language model experts for en- hanced reward modeling in large vision-language models,

    work page

  6. [6]

    Mmbench-video: A long-form multi-shot benchmark for holistic video under- standing

    Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video under- standing. Advances in Neural Information Processing Sys- tems, 37:89098–89124, 2024. 3

    work page 2024

  7. [7]

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in 9 video analysis

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in 9 video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108–24118, 2025. 3

    work page 2025

  8. [8]

    Gemini 2.5 flash, 2025

    Google DeepMind. Gemini 2.5 flash, 2025. High-efficiency multimodal model with adaptive thinking. 6

    work page 2025

Show all 57 references
  1. [9]

    Gemini 2.5 pro, 2025

    Google DeepMind. Gemini 2.5 pro, 2025. Advanced multi- modal reasoning model for complex tasks. 6

    2025

  2. [10]

    The llama 3 herd of models, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Ab- hinav Pandey, and et al. The llama 3 herd of models, 2024. 1

    2024

  3. [11]

    Mmworld: Towards multi- discipline multi-faceted world model evaluation in videos,

    Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wan- rong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, Kevin Lin, William Yang Wang, Lijuan Wang, and Xin Eric Wang. Mmworld: Towards multi- discipline multi-faceted world model evaluation in videos,

  4. [12]

    Video-mmmu: Evaluating knowledge acquisition from multi-discipline pro- fessional videos, 2025

    Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Yuanhan Zhang, Xiang Yue, Bo Li, and Ziwei Liu. Video-mmmu: Evaluating knowledge acquisition from multi-discipline pro- fessional videos, 2025. 3

    2025

  5. [13]

    Flex-judge: Think once, judge anywhere, 2025

    Jongwoo Ko, Sungnyun Kim, Sungwoo Cho, and Se-Young Yun. Flex-judge: Think once, judge anywhere, 2025. 6

    2025

  6. [14]

    Smith, and Hannaneh Hajishirzi

    Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. Rewardbench: Evaluating reward models for language modeling, 2024. 3, 6

    2024

  7. [15]

    Vhelm: A holistic evaluation of vision language models

    Tony Lee, Haoqin Tu, Chi Heem Wong, Wenhao Zheng, Yiyang Zhou, Yifan Mai, Josselin Roberts, Michihiro Ya- sunaga, Huaxiu Yao, Cihang Xie, et al. Vhelm: A holistic evaluation of vision language models. Advances in Neural Information Processing Systems, 37:140632–140666, 2024. 3

    2024

  8. [16]

    Llava-onevision: Easy visual task transfer, 2024

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Zi- wei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 6

    2024

  9. [17]

    Aria: An open multimodal native mixture-of-experts model, 2025

    Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Fan Zhou, Chengen Huang, Yanpeng Li, Chongyan Zhu, Xiaoyi Ren, Chao Li, Yifan Ye, Peng Liu, Lihuan Zhang, Hanshu Yan, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of...

    2025

  10. [18]

    Mvbench: A comprehensive multi-modal video understand- ing benchmark

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understand- ing benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195–...

    2024

  11. [19]

    Vl-rewardbench: A challenging benchmark for vision-language generative reward models

    Lei Li, Yuancheng Wei, Zhihui Xie, Xuqing Yang, Yifan Song, Peiyi Wang, Chenxin An, Tianyu Liu, Sujian Li, Bill Yuchen Lin, et al. Vl-rewardbench: A challenging benchmark for vision-language generative reward models. In Proceedings of the Computer Vision and Pattern Recognitio...

    2025

  12. [20]

    Holistic evaluation of language models, 2023

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, and et al. Holistic evaluation of language models, 2023. 3

    2023

  13. [21]

    Video- safetybench: A benchmark for safety evaluation of video lvlms, 2025

    Xuannan Liu, Zekun Li, Zheqi He, Peipei Li, Shuhan Xia, Xing Cui, Huaibo Huang, Xi Yang, and Ran He. Video- safetybench: A benchmark for safety evaluation of video lvlms, 2025. 4, 5

    2025

  14. [22]

    Rm-bench: Benchmarking reward models of lan- guage models with subtlety and style, 2024

    Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Rm-bench: Benchmarking reward models of lan- guage models with subtlety and style, 2024. 3

    2024

  15. [23]

    Videogpt+: Integrating image and video encoders for enhanced video understanding, 2024

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fa- had Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding, 2024. 3

    2024

  16. [24]

    Smith, Hannaneh Hajishirzi, and Nathan Lambert

    Saumya Malik, Valentina Pyatkin, Sander Land, Jacob Mor- rison, Noah A. Smith, Hannaneh Hajishirzi, and Nathan Lambert. Rewardbench 2: Advancing reward model eval- uation, 2025. 3

    2025

  17. [25]

    Hello gpt-4o

    OpenAI. Hello gpt-4o. https : / / openai . com / index/hello-gpt-4o/, 2024. 6

    2024

  18. [26]

    Gpt-4o mini: advancing cost-efficient intel- ligence

    OpenAI. Gpt-4o mini: advancing cost-efficient intel- ligence. https://openai.com/index/gpt-4o-mini-advancing- cost-efficient-intelligence/, 2024. 6

    2024

  19. [27]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Ad- vances in neural information processing systems, 35:...

    2022

  20. [28]

    Vibe-eval: A hard eval- uation suite for measuring progress of multimodal language models, 2024

    Piotr Padlewski, Max Bain, Matthew Henderson, Zhongkai Zhu, Nishant Relan, Hai Pham, Donovan Ong, Kaloyan Aleksiev, Aitor Ormazabal, Samuel Phua, Ethan Yeo, Eu- genie Lamprecht, Qi Liu, Yuqi Wang, Eric Chen, Deyu Fu, Lei Li, Che Zheng, Cyprien de Masson d’Autume, Dani Yo- gata...

    2024

  21. [29]

    Renjie Pi, Felix Bai, Qibin Chen, Simon Wang, Jiulong Shan, Kieran Liu, and Meng Cao. Mr. judge: Multimodal reasoner as a judge, 2025. 3

    2025

  22. [30]

    Shu Pu, Yaochen Wang, Dongping Chen, Yuhang Chen, Guohao Wang, Qi Qin, Zhongyi Zhang, Zhiyuan Zhang, Ze- tong Zhou, Shuang Gong, Yi Gui, Yao Wan, and Philip S. Yu. Judge anything: Mllm as a judge across any modality,

  23. [31]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741, 2023. 1

    2023

  24. [32]

    Scaling llm test-time compute optimally can be more effec- tive than scaling model parameters, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effec- tive than scaling model parameters, 2024. 2

    2024

  25. [33]

    Aligning large mul- timodal models with factually augmented rlhf

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large mul- timodal models with factually augmented rlhf. In Annual Meeting of the Association for Computational Linguistics ,

  26. [34]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin 10 Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Jun- yang Lin. Qwen2-vl: Enhancing vision-language model...

    2024

  27. [35]

    Visualprm: An effective process reward model for multimodal reasoning, 2025

    Weiyun Wang, Zhangwei Gao, Lianjie Chen, Zhe Chen, Jin- guo Zhu, Xiangyu Zhao, Yangzhou Liu, Yue Cao, Shenglong Ye, Xizhou Zhu, Lewei Lu, Haodong Duan, Yu Qiao, Jifeng Dai, and Wenhai Wang. Visualprm: An effective process reward model for multimodal reasoning, 2025. 2

    2025

  28. [36]

    Skywork-vl re- ward: An effective reward model for multimodal understand- ing and reasoning, 2025

    Xiaokun Wang, Peiyu Wang, Jiangbo Pei, Wei Shen, Yi Peng, Yunzhuo Hao, Weijie Qiu, Ai Jian, Tianyidan Xie, Xuchen Song, Yang Liu, and Yahui Zhou. Skywork-vl re- ward: An effective reward model for multimodal understand- ing and reasoning, 2025. 3, 6

    2025

  29. [37]

    Videohallucer: Evaluating intrinsic and extrinsic hallucinations in large video-language models,

    Yuxuan Wang, Yueqian Wang, Dongyan Zhao, Cihang Xie, and Zilong Zheng. Videohallucer: Evaluating intrinsic and extrinsic hallucinations in large video-language models,

  30. [38]

    Internvideo2.5: Empowering video mllms with long and rich context modeling, 2025

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi- angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, Min Dou, Kai Chen, Wenhai Wang, Yu Qiao, Yali Wang, and Limin Wang. Internvideo2.5: Empowering video mllms with long and rich context modeling, 2025. 6

    2025

  31. [39]

    Unified multimodal chain-of-thought reward model through reinforcement fine- tuning, 2025

    Yibin Wang, Zhimin Li, Yuhang Zang, Chunyu Wang, Qinglin Lu, Cheng Jin, and Jiaqi Wang. Unified multimodal chain-of-thought reward model through reinforcement fine- tuning, 2025. 3, 6

    2025

  32. [40]

    Unified reward model for multimodal understanding and generation, 2025

    Yibin Wang, Yuhang Zang, Hao Li, Cheng Jin, and Jiaqi Wang. Unified reward model for multimodal understanding and generation, 2025. 3, 6

    2025

  33. [41]

    reword- bench: Benchmarking and improving the robustness of re- ward models with transformed inputs, 2025

    Zhaofeng Wu, Michihiro Yasunaga, Andrew Cohen, Yoon Kim, Asli Celikyilmaz, and Marjan Ghazvininejad. reword- bench: Benchmarking and improving the robustness of re- ward models with transformed inputs, 2025. 3

    2025

  34. [42]

    Llava- critic: Learning to evaluate multimodal models

    Tianyi Xiong, Xiyao Wang, Dong Guo, Qinghao Ye, Haoqi Fan, Quanquan Gu, Heng Huang, and Chunyuan Li. Llava- critic: Learning to evaluate multimodal models. In Proceed- ings of the Computer Vision and Pattern Recognition Con- ference, pages 13618–13628, 2025. 3, 6

    2025

  35. [43]

    Thinking in space: How mul- timodal large language models see, remember, and recall spaces

    Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How mul- timodal large language models see, remember, and recall spaces. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 10632–10643, 2025. 3

    2025

  36. [44]

    Minicpm-v: A gpt-4v level mllm on your phone, 2024

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. M...

    2024

  37. [45]

    Multimodal rewardbench: Holistic evalua- tion of reward models for vision language models, 2025

    Michihiro Yasunaga, Luke Zettlemoyer, and Marjan Ghazvininejad. Multimodal rewardbench: Holistic evalua- tion of reward models for vision language models, 2025. 3, 5

    2025

  38. [46]

    mplug- owl3: Towards long image-sequence understanding in multi- modal large language models, 2024

    Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug- owl3: Towards long image-sequence understanding in multi- modal large language models, 2024. 6

    2024

  39. [47]

    Internlm-xcomposer2.5-reward: A simple yet effec- tive multi-modal reward model, 2025

    Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Ziyu Liu, Shengyuan Ding, Shenxi Wu, Yubo Ma, Haodong Duan, Wenwei Zhang, Kai Chen, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer2.5-reward: A simple yet effec- tive multi-modal reward model, 2025. 1, 2, 3, 6

    2025

  40. [48]

    Video instruction tuning with synthetic data, 2024

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Zi- wei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 6

    2024

  41. [49]

    R1-reward: Train- ing multimodal reward model through stable reinforcement learning, 2025

    Yi-Fan Zhang, Xingyu Lu, Xiao Hu, Chaoyou Fu, Bin Wen, Tianke Zhang, Changyi Liu, Kaiyu Jiang, Kaibing Chen, Kaiyu Tang, Haojie Ding, Jiankang Chen, Fan Yang, Zhang Zhang, Tingting Gao, and Liang Wang. R1-reward: Train- ing multimodal reward model through stable reinforcement ...

    2025

  42. [50]

    Mm-rlhf: The next step forward in mul- timodal llm alignment, 2025

    Yi-Fan Zhang, Tao Yu, Haochen Tian, Chaoyou Fu, Peiyan Li, Jianshu Zeng, Wulin Xie, Yang Shi, Huanyu Zhang, Junkang Wu, Xue Wang, Yibo Hu, Bin Wen, Fan Yang, Zhang Zhang, Tingting Gao, Di Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Mm-rlhf: The next step forward in mul- timod...

    2025

  43. [51]

    Mmvu: Measuring expert-level multi- discipline video understanding

    Yilun Zhao, Haowei Zhang, Lujing Xie, Tongyan Hu, Guo Gan, Yitao Long, Zhiyuan Hu, Weiyuan Chen, Chuhan Li, Zhijian Xu, et al. Mmvu: Measuring expert-level multi- discipline video understanding. In Proceedings of the Com- puter Vision and Pattern Recognition Conference , pages...

    2025

  44. [52]

    Generative rlhf-v: Learning principles from multi- modal human preference, 2025

    Jiayi Zhou, Jiaming Ji, Boyuan Chen, Jiapeng Sun, Wenqi Chen, Donghai Hong, Sirui Han, Yike Guo, and Yaodong Yang. Generative rlhf-v: Learning principles from multi- modal human preference, 2025. 3

    2025

  45. [53]

    Input Frames

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shen- glong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, Hao Li, Jiahao Wang, Nianchen Deng, Songze L...

    2025

  46. [54]

    When placed in water, there is a violent reaction

    First, analyze the video phenomena: - In the video, the metal reacts vigorously with water. When placed in water, there is a violent reaction. - The reaction of potassium (\\(K\\)) with water is very violent. The chemical equation is \\(2K + 2H_2O=2KOH + H_2\\uparrow\\), and t...

  47. [55]

    - Silver (\\(Ag\\)):\n - Silver is a very unreactive metal and does not react with water under normal conditions

    Then, check other options: - Iron (\\(Fe\\)):\n - Iron reacts with water only at high - temperature steam condition (\\(3Fe + 4H_2O(g)\\stackrel{high - temperature}{=\\!=\\!=}Fe_3O_4+4H_2\\)), and it does not react with cold water or hot water (liquid water) as violently as sh...

  48. [56]

    - Iron (Fe) reacts with steam (not cold water easily in a simple setup like this video) and silver (Ag) is a noble - metal that does not react with water under normal conditions

    Recall the properties of metals in water: - Lithium (Li), sodium (Na), and potassium (K) are alkali metals that react with water. - Iron (Fe) reacts with steam (not cold water easily in a simple setup like this video) and silver (Ag) is a noble - metal that does not react with...

  49. [57]

    Also, when phenolphthalein is added (the pink - colour change indicates a basic solution), which is consistent with the reaction of alkali metals with water

    Analyze the reaction in the video: - In the video, the metal reacts with water, and there are signs of a chemical reaction (such as possible gas evolution). Also, when phenolphthalein is added (the pink - colour change indicates a basic solution), which is consistent with the ...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.