Pith. sign in

REVIEW 5 major objections 6 minor 66 references

CyberV: Cybernetics for Test-time Scaling in Video Understanding

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CyberV, a training-free cybernetic feedback loop, improves video MLLM accuracy on VideoMMMU by 8.3 points for Qwen2.5-VL-7B and 10.0 for Qwen2.5-VL-72B, with the 7B model surpassing GPT-4o.

desk verdict CyberV's headline gains are mostly Best-of-N with a scoring controller; the cybernetic key-frame loop adds about half a point and lacks a random baseline. read the letter →

arxiv 2506.07971 v1 pith:2NRXITEW submitted 2025-06-09 cs.CV

Jiahao Meng , Shuyang Sun , Yue Tan , Lu Qi , Yunhai Tong , Xiangtai Li , Longyin Wen This is my paper
classification cs.CV
keywords test-timescalingvideounderstandingcyberneticsattentiondriftmultimodallargelanguagemodelskeyframeselectionself-correctionfeedbackcontrol
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

The paper claims that frozen video-language models can be made more accurate at test time without any retraining, by wrapping them in a feedback loop borrowed from cybernetics. On the VideoMMMU benchmark, the loop lifts Qwen2.5-VL-7B by 8.3 points, InternVL3-8B by 5.5, and Qwen2.5-VL-72B by 10.0, with the small models overtaking GPT-4o and the large one approaching human experts. The central idea is that a model's own attention maps reveal which video segments it has lost track of during reasoning, and re-injecting those segments as key frames corrects the drift. If true, this offers a training-free route to stronger video understanding.

What carries the argument

Attention drift: the difference, averaged over attention heads in the final layer, between the attention the answer token pays to each video and subtitle segment under chain-of-thought prompting versus under the base response, written as Δ_video and Δ_sub. Negative values index where the model's perceptual grounding has decayed, and the controller's Inference Feedback module selects the top-k segments by largest decrease and injects the corresponding frames as key frames in the next round. The Score Forest — an ensemble of five scoring signals (softmax confidence, logit stability, repetition penalty, visual attention retention, normalized rank) aggregated into a confidence score with a threshold — decides whether to stop or to trigger the corrective action. These two pieces carry the argument.

What would settle it

Run the same two-round protocol on VideoMMMU but replace the attention-drift key-frame selection (Eq. 3) with the same number of frames drawn uniformly at random from the video and subtitle segments, while keeping the Score Forest threshold and Best-of-N structure untouched. If random key frames reproduce the reported 8.3-point gain for Qwen2.5-VL-7B, the attention-drift signal is not the cause of the improvement; if random frames lose the gain, the mechanism is confirmed.

Watch

Extended reading notes

Core claim

CyberV recasts video inference as a closed-loop control problem. Three components — the MLLM inference system (the plant), a sensor that monitors attention drift and answer predictions, and a controller that scores responses and decides whether to accept them or trigger self-correction — iterate at test time. The self-correction action is concrete: the controller computes the difference between attention distributions in a base response and a chain-of-thought response, picks the top-k video and subtitle segments whose attention decreased most, and injects the corresponding frames back into the input for a second round. The paper reports that this two-round protocol, without any parameter updates or auxiliary vision models, raises accuracy on VideoMMMU by 8.3% for Qwen2.5-VL-7B, 5.5% for InternVL3-8B, and 10.0% for Qwen2.5-VL-72B, surpassing GPT-4o with the small models and approaching human expert scores with the large one; gains of 1.1% are reported on VideoMME and WorldSense.

Load-bearing premise

The load-bearing premise is that the segments where attention drops most sharply between base and chain-of-thought inference are exactly the segments whose frames, once re-injected, correct the model's answer.

Editorial extensions

If this is right

  • If correct, a 7B open-source video model can outperform a proprietary frontier model like GPT-4o on knowledge-dense video questions without any training.
  • The approach also lifts general-purpose benchmarks VideoMME, WorldSense, and MVBench, so the mechanism is not limited to expert-knowledge videos.
  • The two-round loop with N=8 first-round paths and a single corrective second round appears sufficient; the paper reports that increasing N to 16 or 32 does not improve beyond N=8.
  • Naive chain-of-thought prompting degrades accuracy on general-purpose benchmarks, so the reported gains are not simply from 'thinking harder' but depend on the adaptive gating.
  • The framework works across Qwen2.5-VL scales and on InternVL3, suggesting portability across different frozen video MLLMs.

Reading between the lines

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

  • An inference beyond the paper: the reported gains are not yet isolated from the effect of merely seeing more frames — the paper does not compare attention-selected key frames against the same number of randomly sampled frames — so the attention-drift mechanism is the natural next ablation to test.
  • Beyond the paper, the confidence threshold could be tuned per question or per video domain; the paper uses fixed thresholds (e.g., τ=0.3 for VideoMMMU), and the appendix failure cases suggest low-confidence answers benefit from correction while high-confidence ones can be harmed by noisy key frames.
  • A testable extension of the paper's logic is to use the attention-drift signal as a compute scheduler, not only a key-frame selector: allocate the corrective second round only to questions where the drift is large, and skip it elsewhere to save test-time cost.
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 / 6 minor

Summary. The paper presents CyberV, a training-free test-time adaptive scaling framework for video understanding. It wraps a frozen video MLLM in a closed-loop control structure: a Best-of-N inference system, a Sensor that extracts signals including prediction labels and attention drift, and a Controller that aggregates these signals into a Score Forest, decides whether to stop or trigger a second round, and injects attention-guided key frames when confidence is low. Experiments on VideoMMMU show gains of +8.3% for Qwen2.5-VL-7B, +5.5% for InternVL3-8B, and +10.0% for Qwen2.5-VL-72B, plus modest gains on VideoMME, WorldSense, and MVBench. The code is released.

Significance. The strongest aspect of the paper is the empirical demonstration that a small frozen video MLLM can be improved substantially via test-time compute: the Best-of-N with score-based selection is simple and effective, and the paper provides multiple ablations and a code release. The novelty of using attention drift as a self-monitoring signal for key-frame injection is conceptually appealing. However, the paper's central attribution of the headline gains to the cybernetic feedback loop is not supported by its own ablation, and the lack of a random key-frame baseline and the under-specified scoring functions limit the strength of the claims. With revisions addressing these issues, the work could be a useful contribution to test-time scaling for video MLLMs.

major comments (5)
  1. [§4.3, Table 4a] The ablation shows that the one-round Score Forest (Best-of-N without second-round feedback) reaches 62.8%, while the full two-round system reaches 63.3%. The headline +8.3% over the 55.0% base is therefore mostly due to the Best-of-N selection and Score Forest, with the attention-drift key-frame loop contributing only +0.5 points. The abstract and introduction attribute the gain to the cybernetic feedback loop; this is not supported by the reported ablation. Please reframe the contribution accordingly or provide evidence (e.g., an ablation with equal compute budget) that the feedback loop is the main driver.
  2. [§3.4, Eq. (3), §4.3, Table 4b] The key-frame injection is justified as restoring attention to critical visual segments, but no baseline with randomly selected key frames (or another simple selection rule) is reported. Without such a control, the +0.5 (Table 4a) and +1.8 (Table 3a) gains attributed to attention drift could be explained simply by giving the model additional frames. Please include a random key-frame baseline under the same second-round protocol.
  3. [§3.3, Eq. (1), Appendix B] The attention drift is computed between the base response (temperature 0) and one CoT response, but the first round generates seven CoT variants. The paper does not specify which CoT path is used for the drift, nor whether it corresponds to the response selected by the Score Forest. Because the CoT outputs are stochastic (temperature 1.0) and no seed averaging is reported, the drift signal may be noisy and unrelated to the chosen answer. Please clarify the selection and consider aggregating drift over multiple CoT paths or over seeds.
  4. [Appendix B, Tables 1-2] The thresholds τ and numbers of paths N are set per benchmark (VideoMMMU: N=8, τ=0.3; VideoMME: simplest form, τ=0.5; WorldSense: N=16, τ=0.7; MVBench: N=8, τ=0.7) and results are reported on those same benchmarks. This constitutes test-set tuning, which weakens the claim of generalization. Please provide a validation protocol (e.g., fixed hyperparameters across benchmarks) or report sensitivity curves for τ and N on a held-out split.
  5. [§3.4, Eq. (2)] The five scoring mechanisms in the Score Forest are only described verbally ('softmax confidence', 'logit stability', 'binary repetition penalty', 'visual attention retention score', 'normalized rank score'), without exact formulas. Since the Score Forest is a core component, please provide precise definitions in an appendix, or at least point to the released code for each function. This is necessary for reproducibility.
minor comments (6)
  1. [Figure 1, right panel] The baseline for the '+CoT' and '+Wait' series is not clearly defined; please add a legend or clarify in the caption that the baseline is Qwen2.5-VL-7B without subtitles or with subtitles, as appropriate.
  2. [Table 1] The comparison with GPT-4o should explicitly note that GPT-4o is evaluated without subtitle input while the CyberV rows use subtitles; otherwise the comparison may be perceived as unfair.
  3. [§4.4, Table 5] The stability analysis does not report multiple runs or variance, and the disturbance implementation (random shift range) is not described precisely; please add details.
  4. [§1 and §4 (Implementation Details)] The phrase 'extra-model-free' conflicts with the use of Faster-Whisper Large-v3.1 for subtitle extraction; please clarify that the extra model is only for input preprocessing and not part of the inference loop.
  5. [Appendix C, Table 6] The phrase 'genral-purpose' should be corrected to 'general-purpose'.
  6. [§4.1] The main text does not report the computational cost of the two-round scheme (number of forward passes, latency); given the paper's focus on test-time scaling, a compute-accuracy trade-off table would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CyberV is an empirical test-time scaling system evaluated on external benchmarks; no self-citation chain or definitional reduction carries the argument.

full rationale

This is an empirical systems paper rather than a derivation, so the circularity tests apply differently. CyberV's components — Eq. 1 (attention drift), Eq. 2 (Score Forest aggregation), and Eq. 3 (top-k attention-decrease key frames) — are operational definitions with independent empirical content; the paper's central claim that these components improve accuracy is tested against external benchmarks (VideoMMMU, VideoMME, WorldSense, MVBench) rather than being assumed. No self-citations are used to justify the method, and no uniqueness theorem or ansatz is imported from the authors' prior work. The per-benchmark choices of N and tau (e.g., 'On the VideoMMMU dataset, we set N = 8 and τ = 0.3 in the first round') are a tuning/selection concern, not a definitional circularity: the reported accuracy is a measured outcome, not a quantity constructed from those hyperparameters. The ablation showing one-round Score Forest at 62.8 versus two-round at 63.3 indicates the attention-feedback loop contributes less than the headline +8.3, but that is an attribution/over-claim issue, not a case of a prediction reducing to its input by construction. No equation of the paper reduces to its own input by construction, so the honest finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The framework relies on several hand-chosen hyperparameters (tau, N, K, score weights) and on unvalidated domain assumptions about attention interpretability and the reliability of heuristic confidence signals. No new physical entities are introduced.

free parameters (4)
  • per-benchmark threshold tau = 0.3 (VideoMMMU), 0.5 (VideoMME), 0.7 (WorldSense, MVBench)
    The confidence threshold is set separately for each benchmark, likely to maximize accuracy on that evaluation set. This tuning is a form of fitting to the test data.
  • per-benchmark number of paths N = 8 (VideoMMMU), 2 (VideoMME), 16 (WorldSense), 8 (MVBench)
    N is chosen per benchmark, and N=8 on VideoMMMU is selected because it gives the best accuracy in the ablation study (Figure 7a), which is another instance of test-set fitting.
  • top-k key frame count K and cap = K=5 per part, cap=20
    The number of selected key frames is a hand-chosen hyperparameter, not derived from any principle.
  • Score Forest scoring functions = unspecified
    The paper lists five scoring signals (softmax confidence, logit stability, repetition penalty, attention retention, rank) but does not define the exact mapping from signals to [0,1] scores for each tree, leaving hidden degrees of freedom.
assumptions (3)
  • domain assumption Attention maps from the final decoder layer reflect the model's visual grounding.
    Used in the Sensor (Section 3.3) to compute attention drift. The appendix (Table 7b) shows minor sensitivity to layer choice, so this is an assumption.
  • domain assumption The difference in attention between base and CoT decoding isolates reasoning-induced attention loss, not decoding noise.
    Base decoding uses temperature 0, CoT uses temperature 1.0 with top-p sampling, so the attention difference may reflect sampling stochasticity rather than reasoning. This is unvalidated.
  • domain assumption The score heuristics (softmax confidence, logit stability, repetition, attention retention, rank) correlate with answer correctness.
    Section 3.4 defines the Score Forest but provides no calibration or evidence that these signals are reliable correctness estimators.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CyberV: Cybernetics for Test-time Scaling in Video Understanding." pith.science (2026). https://pith.science/paper/2NRXITEW

@misc{pith2026250607971,
  author       = {Pith},
  title        = {Pith review of: CyberV: Cybernetics for Test-time Scaling in Video Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2NRXITEW}},
  note         = {Machine review of arXiv:2506.07971}
}
read the original abstract

Current Multimodal Large Language Models (MLLMs) may struggle with understanding long or complex videos due to computational demands at test time, lack of robustness, and limited accuracy, primarily stemming from their feed-forward processing nature. These limitations could be more severe for models with fewer parameters. To address these limitations, we propose a novel framework inspired by cybernetic principles, redesigning video MLLMs as adaptive systems capable of self-monitoring, self-correction, and dynamic resource allocation during inference. Our approach, CyberV, introduces a cybernetic loop consisting of an MLLM Inference System, a Sensor, and a Controller. Specifically, the sensor monitors forward processes of the MLLM and collects intermediate interpretations, such as attention drift, then the controller determines when and how to trigger self-correction and generate feedback to guide the next round. This test-time adaptive scaling framework enhances frozen MLLMs without requiring retraining or additional components. Experiments demonstrate significant improvements: CyberV boosts Qwen2.5-VL-7B by 8.3% and InternVL3-8B by 5.5% on VideoMMMU, surpassing the competitive proprietary model GPT-4o. When applied to Qwen2.5-VL-72B, it yields a 10.0% improvement, achieving performance even comparable to human experts. Furthermore, our method demonstrates consistent gains on general-purpose benchmarks, such as VideoMME and WorldSense, highlighting its effectiveness and generalization capabilities in making MLLMs more robust and accurate for dynamic video understanding. The code is released at https://github.com/marinero4972/CyberV.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 17 canonical work pages

  1. [1]

    Critique-out-loud reward models

    ZacharyAnkner, MansheejPaul, BrandonCui, JonathanDChang, andPrithvirajAmmanabrolu. Critique-out-loud reward models. arXiv preprint arXiv:2408.11791, 2024

    arXiv 2024

  2. [2]

    Claude Team

    Anthropic. Claude Team. Introducing Claude 3.5 Sonnet.https://www.anthropic.com/claude/sonnet, 2024

    work page 2024

  3. [3]

    An introduction to cybernetics

    William Ross Ashby. An introduction to cybernetics. 1956

    1956

  4. [4]

    Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report.arXiv preprint arXiv:2309.16609, 2023

    arXiv 2023

  5. [5]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  6. [6]

    Forest-of-thought: Scaling test-time compute for enhancing llm reasoning.arXiv preprint arXiv:2412.09078, 2024

    Zhenni Bi, Kai Han, Chuanjian Liu, Yehui Tang, and Yunhe Wang. Forest-of-thought: Scaling test-time compute for enhancing llm reasoning.arXiv preprint arXiv:2412.09078, 2024

    arXiv 2024

  7. [7]

    Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling.arXiv preprint arXiv:2407.21787, 2024

    arXiv 2024

  8. [8]

    On the importance of being emergent.Constructivist Foundations, 5(2):89, March 2010

    Peter Cariani. On the importance of being emergent.Constructivist Foundations, 5(2):89, March 2010. URL https://constructivist.info/5/2/089. Retrieved 13 August 2012

    work page 2010

Show all 66 references
  1. [9]

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271, 2024

    2024 arXiv

  2. [10]

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites

    Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024

    2024

  3. [11]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vis...

    2024

  4. [12]

    Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms

    Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. URLhttps:/...

    2024 arXiv

  5. [13]

    Video-of-thought: step-by-step video reasoning from perception to cognition

    Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong Li Lee, and Wynne Hsu. Video-of-thought: step-by-step video reasoning from perception to cognition. InICML, pages 13109–13125, 2024

    2024

  6. [14]

    Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776, 2025

    Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776, 2025

    2025 arXiv

  7. [15]

    Do i know this entity? knowledge awareness and hallucinations in language models.arXiv preprint arXiv:2411.14257, 2024

    Javier Ferrando, Oscar Obeso, Senthooran Rajamanoharan, and Neel Nanda. Do i know this entity? knowledge awareness and hallucinations in language models.arXiv preprint arXiv:2411.14257, 2024

    2024 arXiv

  8. [16]

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

    Chaoyou Fu, Yuhan Dai, Yondong 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 video analysis. InCVPR, 2025

    2025

  9. [17]

    The boat/helmsman

    Stephen Gage. The boat/helmsman. Technoetic Arts, 5(1):15–24, January 2007. ISSN 1477-965X. doi: 10.1386/tear.5.1.15_1

    2007 doi

  10. [18]

    Stream of search (sos): Learning to search in language

    Kanishk Gandhi, Denise Lee, Gabriel Grand, Muxin Liu, Winson Cheng, Archit Sharma, and Noah D Goodman. Stream of search (sos): Learning to search in language. InCOLM, 2024

    2024

  11. [19]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 11

    2025 arXiv

  12. [20]

    Logic-in-frames: Dynamic keyframe search via visual semantic-logical verification for long video understanding

    Weiyu Guo, Ziyang Chen, Shaoguang Wang, Jianxiang He, Yijie Xu, Jinhui Ye, Ying Sun, and Hui Xiong. Logic-in-frames: Dynamic keyframe search via visual semantic-logical verification for long video understanding. arXiv preprint arXiv:2503.13139, 2025

    2025 arXiv

  13. [21]

    Free video-llm: Prompt-guided visual perception for efficient training-free video llms.arXiv preprint arXiv:2410.10441, 2024

    Kai Han, Jianyuan Guo, Yehui Tang, Wei He, Enhua Wu, and Yunhe Wang. Free video-llm: Prompt-guided visual perception for efficient training-free video llms.arXiv preprint arXiv:2410.10441, 2024

    2024 arXiv

  14. [22]

    Worldsense: Evaluating real-world omnimodal understanding for multimodal llms.arXiv preprint arXiv:2502.04326, 2025

    Jack Hong, Shilin Yan, Jiayin Cai, Xiaolong Jiang, Yao Hu, and Weidi Xie. Worldsense: Evaluating real-world omnimodal understanding for multimodal llms.arXiv preprint arXiv:2502.04326, 2025

    2025 arXiv

  15. [23]

    Following clues, approaching the truth: Explainable micro-video rumor detection via chain-of-thought reasoning

    Rongpei Hong, Jian Lang, Jin Xu, Zhangtao Cheng, Ting Zhong, and Fan Zhou. Following clues, approaching the truth: Explainable micro-video rumor detection via chain-of-thought reasoning. InTHE WEB CONFERENCE, 2025

    2025

  16. [24]

    Cos: Chain-of-shot prompting for long video understanding

    Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video understanding. arXiv preprint arXiv:2502.06428, 2025

    2025 arXiv

  17. [25]

    Video-mmmu: Evaluating knowledge acquisition from multi-discipline professional videos.arXiv preprint arXiv:2501.13826, 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 professional videos.arXiv preprint arXiv:2501.13826, 2025

    2025 arXiv

  18. [26]

    Tsao, and Anima Anandkumar

    Yujia Huang, James Gornet, Sihui Dai, Zhiding Yu, Tan Nguyen, Doris Y. Tsao, and Anima Anandkumar. Neural networks with recurrent generative feedback.https://doi.org/10.48550/arXiv.2007.09200, Jul 2020. First submitted on 17 Jul 2020, latest version 10 Nov 2020

    2007 doi arXiv

  19. [27]

    Memory-space visual prompting for efficient vision-language fine-tuning.arXiv preprint arXiv:2405.05615, 2024

    Shibo Jie, Yehui Tang, Ning Ding, Zhi-Hong Deng, Kai Han, and Yunhe Wang. Memory-space visual prompting for efficient vision-language fine-tuning.arXiv preprint arXiv:2405.05615, 2024

    2024 arXiv

  20. [28]

    A simple model of inference scaling laws.arXiv preprint arXiv:2410.16377, 2024

    Noam Levi. A simple model of inference scaling laws.arXiv preprint arXiv:2410.16377, 2024

    2024 arXiv

  21. [29]

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    2024 arXiv

  22. [30]

    Aria: An open multimodal native mixture-of-experts model.arXiv preprint arXiv:2410.05993, 2024

    Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Fan Zhou, Chengen Huang, Yanpeng Li, et al. Aria: An open multimodal native mixture-of-experts model.arXiv preprint arXiv:2410.05993, 2024

    2024 arXiv

  23. [31]

    Mvbench: A comprehensive multi-modal video understanding 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 understanding benchmark. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195–22206, 2024

    2024

  24. [32]

    Mvbench: A comprehensive multi-modal video understanding 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 understanding benchmark. InCVPR, pages 22195–22206, 2024

    2024

  25. [33]

    Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning.arXiv preprint arXiv:2504.06958, 2025

    Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning.arXiv preprint arXiv:2504.06958, 2025

    2025 arXiv

  26. [34]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InICLR, 2024

    2024

  27. [35]

    Vila: On pre-training for visual language models

    Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. InCVPR, pages 26689–26699, 2024

    2024

  28. [36]

    Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling.arXiv preprint arXiv:2502.06703, 2025

    Runze Liu, Junqi Gao, Jian Zhao, Kaiyan Zhang, Xiu Li, Biqing Qi, Wanli Ouyang, and Bowen Zhou. Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling.arXiv preprint arXiv:2502.06703, 2025

    2025 arXiv

  29. [37]

    Feast your eyes: Mixture-of- resolution adaptation for multimodal large language models.arXiv preprint arXiv:2403.03003, 2024

    Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of- resolution adaptation for multimodal large language models.arXiv preprint arXiv:2403.03003, 2024

    2024 arXiv

  30. [38]

    Mllm-selector: Necessity and diversity-driven high-value data selection for enhanced visual instruction tuning.arXiv preprint arXiv:2503.20502, 2025

    Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Jiayi Ji, Jie Lou, Debing Zhang, and Rongrong Ji. Mllm-selector: Necessity and diversity-driven high-value data selection for enhanced visual instruction tuning.arXiv preprint arXiv:2503.20502, 2025

    2025 arXiv

  31. [39]

    McCulloch and Walter Pitts

    Warren S. McCulloch and Walter Pitts. A logical calculus of the ideas immanent in nervous activity.The Bulletin of Mathematical Biophysics, 5(4):115–133, December 1943. ISSN 0007-4985. doi: 10.1007/BF02478259. 12

    1943 doi

  32. [40]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025

    2025 arXiv

  33. [41]

    Hello gpt4-o.https://openai.com/index/hello-gpt-4o/, 2024

    OpenAI. Hello gpt4-o.https://openai.com/index/hello-gpt-4o/, 2024

    2024

  34. [42]

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

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. InNeurIPS, 2023

    2023

  35. [43]

    Eve: Efficient multimodal vision language models with elastic visual experts.arXiv preprint arXiv:2501.04322, 2025

    Miao Rang, Zhenni Bi, Chuanjian Liu, Yehui Tang, Kai Han, and Yunhe Wang. Eve: Efficient multimodal vision language models with elastic visual experts.arXiv preprint arXiv:2501.04322, 2025

    2025 arXiv

  36. [44]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  37. [45]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  38. [46]

    Long-vita: Scaling large multi-modal models to 1 million tokens with leading short-context accuray.arXiv preprint arXiv:2502.05177, 2025

    Yunhang Shen, Chaoyou Fu, Shaoqi Dong, Xiong Wang, Peixian Chen, Mengdan Zhang, Haoyu Cao, Ke Li, Xiawu Zheng, Yan Zhang, et al. Long-vita: Scaling large multi-modal models to 1 million tokens with leading short-context accuray.arXiv preprint arXiv:2502.05177, 2025

    2025

  39. [47]

    Video-xl: Extra-long vision language model for hour-scale video understanding.arXiv preprint arXiv:2409.14485, 2024

    Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding.arXiv preprint arXiv:2409.14485, 2024

    2024 arXiv

  40. [48]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314, 2024

    2024 arXiv

  41. [49]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530, 2024

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530, 2024

    2024 arXiv

  42. [50]

    Solving math word problems with process-and outcome-based feedback.arXiv preprint arXiv:2211.14275, 2022

    Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. Solving math word problems with process-and outcome-based feedback.arXiv preprint arXiv:2211.14275, 2022

    2022 arXiv

  43. [51]

    Cybernetics: Circular causal and feedback mechanisms in biological and social systems

    Heinz von Foerster. Cybernetics: Circular causal and feedback mechanisms in biological and social systems. In Transactions of the Seventh Conference. Josiah Macy Jr. Foundation, 1952

    1952

  44. [52]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  45. [53]

    Visionllm: Large language model is also an open-ended decoder for vision-centric tasks

    Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. In NeurIPS, 2023

    2023

  46. [54]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. InICLR, 2023

    2023

  47. [55]

    Internvideo2

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling.arXiv preprint arXiv:2501.12386, 2025

    2025 arXiv

  48. [56]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. InNeurIPS, volume 35, pages 24824–24837, 2022

    2022

  49. [57]

    Cybernetics or Control and Communication in the Animal and the Machine

    Norbert Wiener. Cybernetics or Control and Communication in the Animal and the Machine. MIT press, 1948

    1948

  50. [58]

    Controlmllm: Training-free visual prompt learning for multimodal large language models

    Mingrui Wu, Xinyue Cai, Jiayi Ji, Jiale Li, Oucheng Huang, Gen Luo, Hao Fei, Guannan Jiang, Xiaoshuai Sun, and Rongrong Ji. Controlmllm: Training-free visual prompt learning for multimodal large language models. Advances in Neural Information Processing Systems, 37:45206–45234...

    2024

  51. [59]

    Deepseek-prover: Advancing theorem proving in llms through large-scale synthetic data.arXiv preprint arXiv:2405.14333, 2024

    Huajian Xin, Daya Guo, Zhihong Shao, Zhizhou Ren, Qihao Zhu, Bo Liu, Chong Ruan, Wenda Li, and Xiaodan Liang. Deepseek-prover: Advancing theorem proving in llms through large-scale synthetic data.arXiv preprint arXiv:2405.14333, 2024

    2024 arXiv

  52. [60]

    Videollama 3: Frontier multimodal foundation models for image and video understanding

    Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025

    2025 arXiv

  53. [61]

    Video-llama: An instruction-tuned audio-visual language model for video understanding

    Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP (Demos), 2023

    2023

  54. [62]

    Long context transfer from language to vision.arXiv preprint arXiv:2406.16852, 2024

    Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision.arXiv preprint arXiv:2406.16852, 2024

    2024 arXiv

  55. [63]

    Adarefiner: Refining decisions of language models with adaptive feedback

    Wanpeng Zhang and Zongqing Lu. Adarefiner: Refining decisions of language models with adaptive feedback. https://doi.org/10.48550/arXiv.2309.17176, Sep 2023. First submitted on 29 Sep 2023, last revised 3 May 2024

    2023 doi arXiv

  56. [64]

    Tinyllava-video-r1: Towards smaller lmms for video reasoning

    Xingjian Zhang, Siwei Wen, Wenjun Wu, and Lei Huang. Tinyllava-video-r1: Towards smaller lmms for video reasoning. arXiv preprint arXiv:2504.09641, 2025

    2025 arXiv

  57. [65]

    Video instruction tuning with synthetic data.arXiv preprint arXiv:2410.02713, 2024

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data.arXiv preprint arXiv:2410.02713, 2024

    2024 arXiv

  58. [66]

    Add key frames

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 14 Appendix A Over...

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.