Pith. sign in

REVIEW 2 major objections 5 minor 77 references

AdaThinkV claims a video model can learn, from paired think/answer rollouts, whether a question is worth explicit reasoning, reaching 40.79 mean accuracy with an average of 257 output tokens and beating adaptive baselines on the accuracy–to

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 17:17 UTC pith:RGHYZPNX

load-bearing objection Good adaptive-reasoning method, but unstated train/eval overlap (MMR-VBench, Charades-STA, ActivityNet, NExT-GQA) makes the headline numbers uninterpretable. the 2 major comments →

arxiv 2608.01980 v1 pith:RGHYZPNX submitted 2026-08-03 cs.CV cs.AI

AdaThinkV: Adaptive Thinking for Token-Efficient Video Reasoning

Jingqi Tian , Haoji Zhang , Lin Chen , Hongbo Jin , Haonan Xu , Tianrui Zhu , Xingming Shui , Shilin Ma
show 2 more authors
Wenjing Yang Yansong Tang
This is my paper
classification cs.CV cs.AI
keywords adaptive reasoningvideo reasoningchain-of-thoughtreinforcement learningtemporal groundingtoken efficiencymultimodal LLMrollout allocation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a video multimodal LLM can learn when explicit chain-of-thought reasoning is worth its token cost, without offline difficulty labels, confidence thresholds, or an external router. It trains the policy with two matched rollout branches per prompt — forced THINK and forced ANSWER — and estimates the net utility of reasoning as accuracy gain minus a length penalty (ThinkGain). For prompts where all rollouts are unsuccessful and rewards show no variation, it keeps expanding the same group rather than discarding it (VRPO). A sympathetic reading: if correct, this gives a single-model adaptive reasoning mechanism that improves both accuracy and token efficiency, pushing the Pareto frontier of video reasoning.

Core claim

On the paper's own terms, the central discovery is that the decision of whether to reason explicitly can be learned from controllable paired rollouts at the prompt level and expressed as a utility estimate, rather than inferred from confidence or routed externally. ThinkGain computes, for each prompt, the difference in mean normalized accuracy between THINK and ANSWER branches minus λ times the normalized length difference; a dead zone turns this estimate into a mode target that supervises autonomous marker selection, while continuation tokens are optimized by group-relative policy gradients with forced markers masked. The paper further claims that retaining and progressively expanding all-u

What carries the argument

The central object is ThinkGain, a prompt-level utility estimator Δ̂(x) = mean accuracy of THINK rollouts minus mean accuracy of ANSWER rollouts minus λ(length difference)/B, with a dead zone τ producing mode targets {THINK, ANSWER, ∅}. It converts paired, position-controlled rollouts into supervision: L_mode trains the opening mode marker <think>/<answer>, while the continuation policy is trained with forced markers masked from the loss. The companion mechanism, VRPO, classifies each rollout group as retain/expand/discard based on task success contrast C_suc and accuracy reward dispersion σ_acc, expanding zero-contrast groups up to G_max while preserving existing trajectories. Together they

Load-bearing premise

ThinkGain assumes that, for a given prompt, the mean accuracy difference between forced THINK and forced ANSWER rollouts, minus the length penalty, is an unbiased estimate of the true net utility of explicit reasoning; this requires the two branches to differ only in mode and finite-sample generation noise not to systematically bias the comparison.

What would settle it

A concrete check: construct a set of video questions where the same prompt is presented with both modes many times, and compute ThinkGain's Δ̂ against an oracle utility measured by a much larger sample or ground-truth reasoning necessity labels. If prompts with positive oracle utility frequently have negative Δ̂ (or vice versa) beyond the dead zone, the estimator is biased. A second check: run VRPO with an independently resampled update group and compare accuracy; if optional-stopping reuse inflates gains, the improvement should shrink or vanish.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A video MLLM trained this way can select its response mode and generate the answer in one sequence, requiring no preliminary answer or router at inference.
  • Because ThinkGain explicitly prices decoding length, the model outputs fewer tokens on easy questions while retaining (and even improving) accuracy on hard ones; thinking rate scales with difficulty tertiles from 25.6% to 69.1%.
  • VRPO recovers useful RL signal from all-wrong, low-dispersion prompts, adding accuracy beyond what fixed or replacement-based rollout allocation gives at comparable compute.
  • Under the paper's unified protocol, AdaThinkV ranks first or near-first on general video understanding, temporal grounding, and streaming benchmarks, not just the six-setting reasoning suite.
  • The same framework extends, per the paper's conclusion, to image reasoning and other multimodal tasks and to generative/rubric reward models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the load-bearing equivalence of paired rollouts holds, the approach could be applied to any verifiable-reward task where two response formats can be forced, making 'when to think' a learned property rather than a hand-tuned confidence rule; one could test it on math text tasks or audio-visual QA.
  • The ThinkGain estimator and VRPO's optional stopping interact: expansion stops based on reward dispersion, and the same group is reused for both target estimation and policy update, so a testable diagnostic would compare against a fully independent resampled update group on larger scale to see whether selection bias grows with expansion depth.
  • The paper's per-prompt length penalty λ and dead zone τ are fixed hyperparameters; an adaptive schedule for λ could let the model learn a per-domain token budget without retuning.
  • Because the marker is just one token (<think>) vs three (<answer>) in the tokenizer, one could examine whether mode selection is partly an artifact of token-count priors; a tokenizer-invariant control would strengthen the utility interpretation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. AdaThinkV is an adaptive video-reasoning framework that learns a per-prompt choice between explicit chain-of-thought reasoning and direct answering. During RL it samples forced THINK/ANSWER rollouts, estimates prompt-level net utility via ThinkGain (Eq. 1), and uses a dead-zone mode target (Eq. 2) for mode supervision plus a mode-consistency bonus in the group-relative policy loss. VRPO selectively expands all-unsuccessful, low-dispersion rollout groups instead of discarding them, and at inference the model generates the mode marker and continuation in one autoregressive sequence. On a six-setting video reasoning suite the method reports 40.79 average accuracy at 257.20 output tokens, outperforming VideoAuto-R1-Qwen3 by 2.98 points with 22.7% fewer tokens, and it also reports gains on temporal grounding and streaming video benchmarks.

Significance. The method is practical and reasonably novel: it avoids offline difficulty labels, confidence thresholds, and external routers, and it performs mode selection and response generation in a single pass. The evaluation is broad, includes a unified scoring pipeline, reports a bootstrap confidence interval for the headline comparison, and contains an ablation separating the ThinkGain and VRPO effects. These are genuine strengths. However, the central empirical claim is undermined by an apparent train/eval overlap in the RL data distribution, which must be resolved before the reported numbers can be taken as held-out generalization results.

major comments (2)
  1. [Experimental Setup, Fig. 4(c), Table 1, Table 3] Fig. 4(c) lists MMR-VBench (102 prompts) in the RL training data, and Table 1 evaluates MMRV and MMRV-CoT as two of the six settings that determine the headline average 40.79. The paper never states that evaluation prompts were excluded from RL training or deduplicated. If those 102 prompts are the same prompts scored in Table 1, those two settings measure memorization rather than generalization, and the reported 2.98-point gain over VideoAuto-R1-Qwen3 is not a valid held-out comparison. The same issue affects Table 3: Charades-STA, ActivityNet-TVG, and NExT-GQA all appear in the RL data distribution in Fig. 4(c) and are then used as evaluation benchmarks. Please report exact overlap counts after deduplication or re-evaluate on disjoint splits; if any overlap exists, the affected tables must be recomputed and the claims revised.
  2. [Method: VRPO, Eqs. (4)–(5) and 'Optimization relative to each group'] The paper states that the retained group is reused to estimate the ThinkGain target and to compute the policy update, and it acknowledges that 'this reuse may introduce selection bias from optional stopping.' The main text does not report the promised diagnostics (expansion-stage curves and the independently resampled update-group comparison). Because VRPO contributes an additional 2.37 points in Table 6 and is a core methodological contribution, the optional-stopping bias is load-bearing. Please include these diagnostics in the main text or give the key numbers and a bound on the bias; a pointer to the supplement is not sufficient for a claim this central.
minor comments (5)
  1. [Table 1] The abbreviations MMRV and MMRV-CoT are not defined in the caption or table note; the text cites MMR-VBench, but the two evaluation settings should be expanded at first use.
  2. [Fig. 4(c)] Dataset abbreviations (ViRL, ThinkLite-Hard, TimeR1, STI-Bench, TVBench-QA) are not defined in the caption. Also clarify whether counts refer to prompts or video-QA pairs.
  3. [Abstract / Evaluation protocol] The phrase 'unified suite' is used loosely: Table 1 has six settings, Table 2 uses five additional benchmarks, and Table 4 uses two streaming benchmarks. Clarify that the headline 40.79 refers only to the Table 1 suite.
  4. [Temporal grounding protocol] The evaluation protocol says all models 'share inputs, parsers, and scoring rules,' but several temporal grounding baselines (e.g., TimeChat, TimeMarker) were originally trained with different frame-sampling or input templates. A sentence on how their native templates are combined with the shared protocol would remove ambiguity.
  5. [Table 5] The RL-only setting has 92.4% format compliance while SFT-only has 99.1%; a brief explanation of why RL-only degrades format compliance would be useful.

Circularity Check

2 steps flagged

Unstated overlap between RL training data and evaluation benchmarks makes the headline 40.79 and temporal-grounding results partly in-sample; the method itself is not definitionally circular.

specific steps
  1. fitted input called prediction [Fig. 4(c) (RL Data Distribution) / Table 1 (Main video reasoning results) / Section 'Experimental Setup: Training data']
    ""MMR-VBench102" (Fig. 4c: RL Data Distribution); "MMRV" and "MMRV-CoT" (Table 1)."

    The headline 40.79 is the average of six settings, including MMRV and MMRV-CoT. The RL training set contains 102 MMR-VBench prompts (Fig. 4c). The paper states no split, deduplication, or train-only caveat for any RL dataset, so these two settings are in-sample: the model was optimized on the same benchmark it is then scored on. The 2.98-point gain over VideoAuto-R1-Qwen3 on this average is therefore partly forced by training on the evaluation data, not a held-out prediction.

  2. fitted input called prediction [Fig. 4(c) (RL Data Distribution) / Table 3 (Temporal grounding benchmarks)]
    ""CharadesSTA 5103" and "NeXT-GQA 1379" (Fig. 4c); "Charades-STA, ActivityNet, NExT-GQA" (Table 3)."

    Table 3 reports R@0.5 and mIoU on Charades-STA, ActivityNet, and NExT-GQA, while Fig. 4(c) lists CharadesSTA, ActivityNet-TVG, and NeXT-GQA as RL training data. With no stated held-out split, these temporal-grounding results are training-set performance, so the claim of ranking first on all but one temporal-grounding metric is not an independent prediction. The comparison is in-sample for AdaThinkV and does not demonstrate generalization.

full rationale

The core learning rule is not circular: ThinkGain estimates utility from paired rollouts and supervises mode selection; VRPO is a rollout-allocation heuristic; neither reduces a claimed result to its inputs by definition. The paper's self-acknowledged optional-stopping bias in VRPO is a disclosed statistical limitation, not a hidden circular step. Self-citations in Related Work are not load-bearing. However, the evaluation chain contains a genuine fitted-input-as-prediction problem: two of the six settings in the headline average (MMRV, MMRV-CoT) and all three temporal-grounding benchmarks in Table 3 appear in the RL data distribution (Fig. 4c), with no stated train/test split. As written, those numbers are in-sample, and the headline 40.79 and the 2.98-point/22.7%-token claims are partly forced by training on the evaluation data. If the authors can show a disjoint split, the circularity would be resolved; the paper currently provides no such statement.

Axiom & Free-Parameter Ledger

11 free parameters · 3 axioms · 0 invented entities

The method rests on several hand-chosen hyperparameters (λ, B, τ, α, γ, δ_acc, η_task, G0, Gmax, ρ, clipping bounds) plus a key domain assumption that paired THINK/ANSWER rollouts provide an unbiased estimate of reasoning utility. No new physical or conceptual entities are introduced. The MMR-VBench overlap is an unstated data-handling assumption that affects the soundness of the headline number.

free parameters (11)
  • lambda (length penalty weight) = 0.1
    Balances accuracy gain against output length in ThinkGain Eq. (1).
  • B (length normalization constant) = 512
    Normalizes token counts to a comparable scale in ThinkGain Eq. (1).
  • tau (dead zone margin) = 0.05
    Suppresses small estimated utility differences before converting to a mode target in Eq. (2).
  • alpha (mode-consistency bonus weight) = 0.3
    Weight for the bonus g_i when the response mode matches the target in Eq. (4).
  • gamma (mode supervision weight) = 0.1
    Scales L_mode relative to L_RL in the total objective Eq. (7).
  • delta_acc (accuracy reward dispersion threshold) = 0.05
    Threshold for deciding when a group has sufficient reward variation to train on in VRPO Eq. (5).
  • eta_task (continuous reward success threshold) = 0.5
    Threshold for binarizing continuous rewards; the paper claims it is fixed by the task success criterion rather than tuned on test sets.
  • G0 (initial rollout group size) = 8
    Initial group size for VRPO; expansion adds four rollouts per mode per stage.
  • Gmax (maximum expanded group size) = 32
    Cap on VRPO expansion.
  • rho (THINK branch fraction) = 0.5
    Balanced forced-mode rollout allocation in ThinkGain.
  • epsilon_low and epsilon_high (PPO clipping bounds) = 0.20, 0.28
    Clipping bounds for the importance ratio in the RL objective.
axioms (3)
  • standard math Group-relative RL and PPO clipping machinery from GRPO/DAPO is valid for this setting.
    The paper uses DAPO-style group-relative advantages and clipped importance ratios without a value model or KL penalty.
  • domain assumption Within-prompt matched THINK/ANSWER rollouts remove between-prompt difficulty variation and yield an unbiased estimate of reasoning utility.
    Stated in the Method section: 'their within-prompt comparison removes between-prompt difficulty variation while retaining finite-sample generation noise.' This is load-bearing for ThinkGain.
  • domain assumption The RL training data and the evaluation benchmarks are disjoint, or overlapping names like MMR-VBench use separate splits.
    The paper does not state this, but the RL data distribution in Fig. 4c includes MMR-VBench 102 while Table 1 reports MMR-VBench accuracy. The claim that evaluation is unbiased requires this assumption.

pith-pipeline@v1.3.0-daily-deepseek · 15060 in / 12305 out tokens · 127174 ms · 2026-08-04T17:17:52.371355+00:00 · methodology

0 comments
read the original abstract

Chain-of-thought (CoT) reasoning can improve performance on difficult video questions but often wastes decoding tokens on simple ones. We study whether a video multimodal large language model can adapt its reasoning effort to each question. We propose AdaThinkV, an adaptive framework for video reasoning that learns whether to reason explicitly without offline difficulty labels, manually tuned confidence thresholds, or an external router. During reinforcement learning, AdaThinkV samples matched rollouts in explicit reasoning and direct answering modes for each prompt. ThinkGain estimates the prompt-level utility of explicit reasoning by balancing its accuracy gain against additional response length, providing supervision for both conditional response generation and autonomous mode selection. For difficult prompts, limited rollout exploration can yield groups in which every response is unsuccessful and accuracy rewards show little variation, providing insufficient signal for learning. We therefore introduce Variance Recovery Policy Optimization (VRPO), which retains and progressively expands these groups to recover informative signals from prompts that are difficult yet solvable. At inference, AdaThinkV selects a response mode and generates the response in a single autoregressive sequence. Across a unified suite of video reasoning evaluations, AdaThinkV achieves a mean accuracy of 40.79 with an average of 257.20 output tokens, outperforming the strongest evaluated adaptive baseline by 2.98 points while using 22.7% fewer tokens. Project page: https://trilarflagz.github.io/AdaThinkV/

Figures

Figures reproduced from arXiv: 2608.01980 by Haoji Zhang, Haonan Xu, Hongbo Jin, Jingqi Tian, Lin Chen, Shilin Ma, Tianrui Zhu, Wenjing Yang, Xingming Shui, Yansong Tang.

Figure 1
Figure 1. Figure 1: Why does response mode selection matter? Excessive reasoning can overcomplicate simple questions, while confidence [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AdaThinkV RL training and inference pipeline. ThinkGain constructs mode supervision from prompt-level [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Rollout allocation comparison. GRPO uses fixed [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training data construction pipeline and data distribution. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy–token trade-off. AdaThinkV advances [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

77 extracted references · 16 linked inside Pith

  1. [1]

    2022 , eprint =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. 2022 , eprint =

    2022

  2. [2]

    2023 , eprint =

    Multimodal Chain-of-Thought Reasoning in Language Models , author =. 2023 , eprint =

    2023

  3. [3]

    2024 , eprint =

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author =. 2024 , eprint =

    2024

  4. [4]

    2024 , eprint =

    TempCompass: Do Video LLMs Really Understand Videos? , author =. 2024 , eprint =

    2024

  5. [5]

    2024 , eprint =

    Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis , author =. 2024 , eprint =

    2024

  6. [6]

    2023 , eprint =

    MVBench: A Comprehensive Multi-modal Video Understanding Benchmark , author =. 2023 , eprint =

    2023

  7. [7]

    2025 , eprint =

    Video-R1: Reinforcing Video Reasoning in MLLMs , author =. 2025 , eprint =

    2025

  8. [8]

    2025 , eprint =

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author =. 2025 , eprint =

    2025

  9. [9]

    2025 , eprint =

    VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? , author =. 2025 , eprint =

    2025

  10. [10]

    2025 , eprint =

    Video-R2: Reinforcing Consistent and Grounded Reasoning in Multimodal Language Models , author =. 2025 , eprint =

    2025

  11. [11]

    2025 , eprint =

    Rethinking Chain-of-Thought Reasoning for Videos , author =. 2025 , eprint =

    2025

  12. [12]

    2026 , eprint =

    VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice , author =. 2026 , eprint =

    2026

  13. [13]

    2026 , eprint =

    Switch-Reasoner: Learn When to Think in Multitask Mixtures via Reinforcement Learning , author =. 2026 , eprint =

    2026

  14. [14]

    2025 , eprint =

    VideoChat-R1.5: Visual Test-Time Scaling to Reinforce Multimodal Reasoning by Iterative Perception , author =. 2025 , eprint =

    2025

  15. [15]

    2022 , eprint =

    Holistic Evaluation of Language Models , author =. 2022 , eprint =

    2022

  16. [16]

    2024 , eprint =

    VHELM: A Holistic Evaluation of Vision Language Models , author =. 2024 , eprint =

    2024

  17. [17]

    2025 , eprint =

    OckBench: Measuring the Efficiency of LLM Reasoning , author =. 2025 , eprint =

    2025

  18. [18]

    2025 , eprint =

    Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models , author =. 2025 , eprint =

    2025

  19. [19]

    2025 , eprint =

    How Well do LLMs Compress Their Own Chain-of-Thought? A Token Complexity Approach , author =. 2025 , eprint =

    2025

  20. [20]

    2026 , eprint =

    Train Less, Learn More: Adaptive Efficient Rollout Optimization for Group-Based Reinforcement Learning , author =. 2026 , eprint =

    2026

  21. [21]

    2026 , eprint =

    Adaptive Rollout Allocation for Online Reinforcement Learning with Verifiable Rewards , author =. 2026 , eprint =

    2026

  22. [22]

    2026 , eprint =

    Cross-Epoch Adaptive Rollout Optimization for RL Post-Training , author =. 2026 , eprint =

    2026

  23. [23]

    2025 , eprint =

    No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping , author =. 2025 , eprint =

    2025

  24. [24]

    2025 , eprint =

    Act Only When It Pays: Efficient Reinforcement Learning for LLM Reasoning via Selective Rollouts , author =. 2025 , eprint =

    2025

  25. [25]

    2026 , eprint =

    Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training , author =. 2026 , eprint =

    2026

  26. [26]

    2026 , eprint =

    Selective Off-Policy Reference Tuning with Plan Guidance , author =. 2026 , eprint =

    2026

  27. [27]

    Advances in Neural Information Processing Systems , year =

    When Thinking Drifts: Evidential Grounding for Robust Video Reasoning , author =. Advances in Neural Information Processing Systems , year =. 2510.06077 , archivePrefix =

    arXiv

  28. [28]

    Learning When to Think: Shaping Adaptive Reasoning in

    Tu, Songjun and Lin, Jiahao and Zhang, Qichao and Tian, Xiangyu and Li, Linjing and Lan, Xiangyuan and Zhao, Dongbin , year =. Learning When to Think: Shaping Adaptive Reasoning in. 2505.10832 , archivePrefix =

    arXiv

  29. [29]

    2505.13417 , archivePrefix =

    Zhang, Jiajie and Lin, Nianyi and Hou, Lei and Feng, Ling and Li, Juanzi , year =. 2505.13417 , archivePrefix =

    Pith/arXiv arXiv

  30. [30]

    2505.13379 , archivePrefix =

    Fang, Gongfan and Ma, Xinyin and Wang, Xinchao , year =. 2505.13379 , archivePrefix =

    Pith/arXiv arXiv

  31. [31]

    International Conference on Learning Representations , year =

    Adaptive Thinking: Large Language Models Know When to Think in Latent Space , author =. International Conference on Learning Representations , year =

  32. [32]

    When to Think, When to Speak: Learning Disclosure Policies for

    Wei, Jiaqi and Guo, Xuehang and Yu, Pengfei and Zhang, Xiang and Ouyang, Wanli and Sun, Siqi and Wang, Qingyun and You, Chenyu , year =. When to Think, When to Speak: Learning Disclosure Policies for. 2605.03314 , archivePrefix =

    Pith/arXiv arXiv

  33. [33]

    2025 , eprint =

    Video Reasoning without Training , author =. 2025 , eprint =

    2025

  34. [34]

    2025 , eprint =

    Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition , author =. 2025 , eprint =

    2025

  35. [35]

    Findings of the Association for Computational Linguistics: EMNLP , year =

    Video-Skill-CoT: Skill-based Chain-of-Thoughts for Domain-Adaptive Video Reasoning , author =. Findings of the Association for Computational Linguistics: EMNLP , year =

  36. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Flexible Frame Selection for Efficient Video Reasoning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  37. [37]

    Yu, Shoubin and Yoon, Jaehong and Bansal, Mohit , booktitle =

  38. [38]

    Reinforcement Learning Tuning for

    Li, Hongyu and Han, Songhao and Liao, Yue and Luo, Junfeng and Gao, Jialin and Yan, Shuicheng and Liu, Si , year =. Reinforcement Learning Tuning for. 2506.01908 , archivePrefix =

    Pith/arXiv arXiv

  39. [39]

    2025 , eprint =

    Wang, Ye and Wang, Ziheng and Xu, Boshen and Du, Yang and Lin, Kejun and Xiao, Zihan and Yue, Zihao and Ju, Jianzhong and Zhang, Liang and Yang, Dingyi and Fang, Xiangnan and He, Zewen and Luo, Zhenbo and Wang, Wenxuan and Lin, Junqi and Luan, Jian and Jin, Qin , booktitle =. 2025 , eprint =

    2025

  40. [40]

    2025 , eprint =

    Thinking With Videos: Multimodal Tool-Augmented Reinforcement Learning for Long Video Reasoning , author =. 2025 , eprint =

    2025

  41. [41]

    Chen, Yukang and Huang, Wei and Shi, Baifeng and Hu, Qinghao and Ye, Hanrong and Zhu, Ligeng and Liu, Zhijian and Molchanov, Pavlo and Kautz, Jan and Qi, Xiaojuan and Liu, Sifei and Yin, Hongxu and Lu, Yao and Han, Song , year =. Scaling. 2507.07966 , archivePrefix =

    arXiv

  42. [42]

    2025 , eprint =

    Wang, Ziyang and Yoon, Jaehong and Yu, Shoubin and Islam, Md Mohaiminul and Bertasius, Gedas and Bansal, Mohit , booktitle =. 2025 , eprint =

    2025

  43. [43]

    2511.21631 , archivePrefix =

    Bai, Shuai and Cai, Yuxuan and Chen, Ruizhe and Chen, Keqin and Chen, Xionghui and others , year =. 2511.21631 , archivePrefix =

    Pith/arXiv arXiv

  44. [44]

    2407.15754 , archivePrefix =

    Wu, Haoning and Li, Dongxu and Chen, Bei and Li, Junnan , year =. 2407.15754 , archivePrefix =

    Pith/arXiv arXiv

  45. [45]

    2411.03628 , archivePrefix =

    Lin, Junming and Fang, Zheng and Chen, Chi and Wan, Zihao and Luo, Fuwen and Li, Peng and Liu, Yang and Sun, Maosong , year =. 2411.03628 , archivePrefix =

    Pith/arXiv arXiv

  46. [46]

    Online Video Understanding:

    Huang, Zhenpeng and Li, Xinhao and Li, Jiaqi and Wang, Jing and Zeng, Xiangyu and Liang, Cheng and Wu, Tao and Chen, Xi and Li, Liang and Wang, Limin , booktitle =. Online Video Understanding:. 2025 , eprint =

    2025

  47. [47]

    2025 , eprint =

    Li, Yifei and Niu, Junbo and Miao, Ziyang and Ge, Chunjiang and Zhou, Yuanhang and He, Qihao and Dong, Xiaoyi and Duan, Haodong and Ding, Shuangrui and Qian, Rui and Zhang, Pan and Zang, Yuhang and Cao, Yuhang and He, Conghui and Wang, Jiaqi , booktitle =. 2025 , eprint =

    2025

  48. [48]

    2017 , pages =

    Gao, Jiyang and Sun, Chen and Yang, Zhenheng and Nevatia, Ram , booktitle =. 2017 , pages =

    2017

  49. [49]

    Proceedings of the IEEE International Conference on Computer Vision , year =

    Dense-Captioning Events in Videos , author =. Proceedings of the IEEE International Conference on Computer Vision , year =

  50. [50]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Can I Trust Your Answer? Visually Grounded Video Question Answering , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =. 2309.01327 , archivePrefix =

    Pith/arXiv arXiv

  51. [51]

    2026 , eprint =

    What Should a Streaming Video Model Remember? , author =. 2026 , eprint =

    2026

  52. [52]

    2026 , eprint =

    Rasheed, Hanoona and Shaker, Abdelrahman and Tang, Anqi and Maaz, Muhammad and Yang, Ming-Hsuan and Khan, Salman and Khan, Fahad Shahbaz , booktitle =. 2026 , eprint =

    2026

  53. [53]

    2026 , eprint =

    Zhu, Kejian and Jin, Zhuoran and Yuan, Hongbang and Li, Jiachun and Tu, Shangqing and Cao, Pengfei and Chen, Yubo and Liu, Kang and Zhao, Jun , booktitle =. 2026 , eprint =

    2026

  54. [54]

    2510.08559 , archivePrefix =

    Deng, Andong and Yang, Taojiannan and Yu, Shoubin and Spencer, Lincoln and Bansal, Mohit and Chen, Chen and Yeung-Levy, Serena and Wang, Xiaohan , year =. 2510.08559 , archivePrefix =

    arXiv

  55. [55]

    2025 , eprint =

    Zhao, Yilun and Xie, Lujing and Zhang, Haowei and Gan, Guo and Long, Yitao and Hu, Zhiyuan and Hu, Tongyan and Chen, Weiyuan and Li, Chuhan and Song, Junyang and Xu, Zhijian and Wang, Chengye and Pan, Weifeng and Shangguan, Ziyao and Tang, Xiangru and Liang, Zhenwen and Liu, Yixin and Zhao, Chen and Cohan, Arman , booktitle =. 2025 , eprint =

    2025

  56. [56]

    2501.13826 , archivePrefix =

    Hu, Kairui and Wu, Penghao and Pu, Fanyi and Xiao, Wang and Zhang, Yuanhan and Yue, Xiang and Li, Bo and Liu, Ziwei , year =. 2501.13826 , archivePrefix =

    Pith/arXiv arXiv

  57. [57]

    Transactions on Machine Learning Research , year =

    A Shortcut-Aware Video-QA Benchmark for Physical Understanding via Minimal Video Pairs , author =. Transactions on Machine Learning Research , year =. 2506.09987 , archivePrefix =

    Pith/arXiv arXiv

  58. [58]

    2026 , eprint =

    A Simple Baseline for Streaming Video Understanding , author =. 2026 , eprint =

    2026

  59. [59]

    Jin, Hongbo and Zhu, Mingnan and Tian, Jingqi and Jiang, Xu and Du, Zhongjing and Tang, Haoran and Xie, Siyi and Zhang, Qiaoman and Ding, Jiayu , eprint =

  60. [60]

    2512.18741 , archivePrefix =

    Memorize-and-Generate: Towards Long-Term Consistency in Real-Time Video Generation , author =. 2512.18741 , archivePrefix =

    arXiv

  61. [61]

    Lin, Hao and Lv, Kunyang and Jiang, Xu and Tian, Jingqi and Du, Zhongjing and Ding, Jiayu and Zhang, Qiaoman and Jin, Hongbo , eprint =

  62. [62]

    Jin, Hongbo and Zhu, Rongpeng and Du, Zhongjing and Jiang, Xu and Tian, Jingqi and Zhang, Qiaoman and Ding, Jiayu , eprint =

  63. [63]

    Tian, Jingqi and Du, Yiheng and Zhang, Haoji and Wang, Yuji and Lee, Isaac Ning and Bai, Xulong and Zhu, Tianrui and Niu, Jingxuan and Tang, Yansong , eprint =

  64. [64]

    2502.13923 , archivePrefix =

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and Zhong, Humen and Zhu, Yuanzhi and Yang, Mingkun and Li, Zhaohai and Wan, Jianqiang and Wang, Pengfei and Ding, Wei and Fu, Zheren and Xu, Yiheng and Ye, Jiabo and Zhang, Xi and Xie, Tianbao and Cheng, Z...

    Pith/arXiv arXiv

  65. [65]

    2505.12434 , archivePrefix =

    Wang, Qi and Yu, Yanrui and Yuan, Ye and Mao, Rui and Zhou, Tianfei , year =. 2505.12434 , archivePrefix =

    arXiv

  66. [66]

    2509.24786 , archivePrefix =

    Fu, Shenghao and Yang, Qize and Li, Yuan-Ming and Wei, Xihan and Xie, Xiaohua and Zheng, Wei-Shi , year =. 2509.24786 , archivePrefix =

    arXiv

  67. [67]

    2312.02051 , archivePrefix =

    Ren, Shuhuai and Yao, Linli and Li, Shicheng and Sun, Xu and Hou, Lu , year =. 2312.02051 , archivePrefix =

    Pith/arXiv arXiv

  68. [68]

    2410.19702 , archivePrefix =

    Zeng, Xiangyu and Li, Kunchang and Wang, Chenting and Li, Xinhao and Jiang, Tianxiang and Yan, Ziang and Li, Songze and Shi, Yansong and Yue, Zhengrong and Wang, Yi and Wang, Yali and Qiao, Yu and Wang, Limin , year =. 2410.19702 , archivePrefix =

    Pith/arXiv arXiv

  69. [69]

    2411.18211 , archivePrefix =

    Chen, Shimin and Lan, Xiaohan and Yuan, Yitian and Jie, Zequn and Ma, Lin , year =. 2411.18211 , archivePrefix =

    Pith/arXiv arXiv

  70. [70]

    2509.24871 , archivePrefix =

    Zeng, Xiangyu and Qiu, Kefan and Zhang, Qingyu and Li, Xinhao and Wang, Jing and Li, Jiaxin and Yan, Ziang and Tian, Kun and Tian, Meng and Zhao, Xinhai and Wang, Yi and Wang, Limin , year =. 2509.24871 , archivePrefix =

    arXiv

  71. [71]

    2601.14724 , archivePrefix =

    Zhang, Haowei and Yang, Shudong and Fu, Jinlan and Ng, See-Kiong and Qiu, Xipeng , year =. 2601.14724 , archivePrefix =

    Pith/arXiv arXiv

  72. [72]

    2026 , eprint =

    Segment-Aligned Policy Optimization for Multi-Modal Reasoning , author =. 2026 , eprint =

    2026

  73. [73]

    Cores, Daniel and Dorkenwald, Michael and Mucientes, Manuel and Snoek, Cees G. M. and Asano, Yuki M. , year =. 2410.07752 , archivePrefix =

    Pith/arXiv arXiv

  74. [74]

    2025 , eprint =

    Li, Yun and Zhang, Yiming and Lin, Tao and Liu, XiangRui and Cai, Wenxiao and Liu, Zheng and Zhao, Bo , booktitle =. 2025 , eprint =

    2025

  75. [75]

    2025 , eprint =

    Wang, Xiyao and Yang, Zhengyuan and Feng, Chao and Lu, Hongjin and Li, Linjie and Lin, Chung-Ching and Lin, Kevin and Huang, Furong and Wang, Lijuan , booktitle =. 2025 , eprint =

    2025

  76. [76]

    2511.16334 , archivePrefix =

    Zhang, Kaichen and Wu, Keming and Yang, Zuhao and Li, Bo and Hu, Kairui and Wang, Bin and Liu, Ziwei and Li, Xingxuan and Bing, Lidong , year =. 2511.16334 , archivePrefix =

    arXiv

  77. [77]

    2017 , eprint =

    Proximal Policy Optimization Algorithms , author =. 2017 , eprint =

    2017