pith. sign in

arxiv: 2605.06094 · v4 · pith:E5JZWUAYnew · submitted 2026-05-07 · 💻 cs.CV · cs.AI

VISD: Enhancing Video Reasoning via Structured Self-Distillation

Hao Lin , Kunyang Lv , Xu Jiang , Jingqi Tian , Zhongjing Du , Jiayu Ding , Qiaoman Zhang , Hongbo Jin This is my paper

Pith reviewed 2026-05-25 06:05 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords self-distillationVideoLLMreinforcement learningvideo reasoningstructured feedbacktoken-level supervisionspatio-temporal grounding
0
0 comments X

The pith

VISD introduces structured self-distillation to supply diagnostically specific token-level supervision for VideoLLMs, stably combined with reinforcement learning.

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

The paper seeks to overcome sparse sequence-level rewards and unstructured self-distillation in VideoLLM training for complex reasoning. It proposes VISD, which deploys a video-aware judge to decompose reasoning quality into answer correctness, logical consistency, and spatio-temporal grounding. These signals supply token-level guidance through a teacher policy. A direction-magnitude decoupling mechanism keeps the dense signals compatible with RL by letting rollout advantages set update direction while judge signals scale token magnitudes. Experiments report higher accuracy, better grounding, and nearly 2x faster convergence, which would matter for efficient training on long video sequences.

Core claim

VISD is a structured self-distillation framework that uses a video-aware judge model to decompose reasoning quality into multiple dimensions including answer correctness, logical consistency, and spatio-temporal grounding. This structured feedback guides a teacher policy for token-level supervision. To integrate the dense signals stably with RL, the framework introduces a direction-magnitude decoupling mechanism where rollout-level advantages determine update direction and the privileged signals modulate token-level update magnitudes. Curriculum scheduling and EMA-based teacher stabilization further support robust optimization over long sequences.

What carries the argument

Direction-magnitude decoupling mechanism that uses rollout advantages to set policy update direction while structured judge signals control token-level update magnitudes.

If this is right

  • VideoLLMs achieve higher answer accuracy on diverse reasoning benchmarks.
  • Spatio-temporal grounding quality in generated responses improves.
  • Training reaches target performance in nearly half the optimization steps.
  • Credit assignment over long reasoning trajectories becomes finer-grained and more semantically aligned.

Where Pith is reading between the lines

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

  • The decoupling approach could stabilize dense supervision in other long-sequence generative tasks outside video.
  • Swapping in a stronger judge model might increase the performance gains without changing the rest of the training setup.
  • Diagnostic decomposition of quality signals may prove useful for reducing instability when mixing dense and verifiable rewards in multimodal models.

Load-bearing premise

A video-aware judge model can reliably decompose reasoning quality into the dimensions of correctness, logical consistency, and spatio-temporal grounding to supply stable, non-biased privileged signals.

What would settle it

An ablation that removes the structured judge feedback and finds no remaining gains in answer accuracy, grounding quality, or convergence speed compared with standard RLVR would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.06094 by Hao Lin, Hongbo Jin, Jiayu Ding, Jingqi Tian, Kunyang Lv, Qiaoman Zhang, Xu Jiang, Zhongjing Du.

Figure 1
Figure 1. Figure 1: Comparison between RLVR method and VISD. (a) The RLVR method, e.g., GRPO, provides sparse sequence-level signals and fails to capture fine-grained spatial and temporal evidence. (b) VISD uses dense token-level updates, thereby improving convergence speed and overall performance. To effectively integrate structured supervision with reinforcement learning, we extend the direction– magnitude decoupling paradi… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VISD. VISD first uses a video-aware judge to generate structured privileged feedback for each student rollout, then replays the same completion with a feedback-conditioned teacher. The rollout-level reward advantage determines the update direction, while the teacher-student discrepancy modulates token-level update magnitudes for fine-grained credit assignment. R(x, y), which reflects task-level… view at source ↗
Figure 3
Figure 3. Figure 3: Training ablation curves. (a) Feedback conditioning compares VISD with and without judge feedback, showing that structured feedback leads to a stronger final reward trajectory. (b) Teacher update strategy compares current-policy, Sync-10, and EMA teachers, where EMA provides the most stable optimization and best final reward. (c) Top-K support versus sampled token compares two teacher-student credit formul… view at source ↗
Figure 4
Figure 4. Figure 4: Top-K support versus sampled-token training curves. We compare (a) total reward, (b) group mean reward, (c) answer accuracy, (d) temporal IoU, (e) spatial IoU, (f) temporal point, (g) temporal segment, (h) spatial grounding, and (i) format reward. Feedback conditioning view at source ↗
Figure 5
Figure 5. Figure 5: Component-level reward curves for feedback ablation. Each panel compares VISD with and without judge feedback for one reward component: (a) total reward, (b) group mean reward, (c) answer accuracy, (d) temporal IoU, (e) spatial IoU, (f) temporal point, (g) temporal segment, (h) spatial grounding, and (i) format. Consistent with view at source ↗
Figure 6
Figure 6. Figure 6: Teacher update stability diagnostics. We show (a) gradient norm and (b) response entropy for different teacher parameterizations. D.4 Training Details for Ablation Settings The ablation curves are plotted by training step. For the feedback comparison, the no-feedback setting uses the same reinforcement-learning and teacher-replay pipeline but removes the judge-generated 22 view at source ↗
Figure 7
Figure 7. Figure 7: Trajectory-dependent judge feedback. For the same video question, the judge provides different evaluations and diagnoses for different student rollouts. The feedback is used as privileged information for teacher replay rather than as a replacement for the reinforcement reward. 26 view at source ↗
Figure 8
Figure 8. Figure 8: Answer-only versus feedback-conditioned token credit. (a) and (b) visualize the same fixed student rollout; only the teacher context changes. Warmer and cooler token backgrounds indicate positive or negative teacher-student token evidence used to modulate policy-gradient magnitude. In the answer-only replay, the teacher is conditioned on verified answer-side information, so the token-credit pattern mainly … view at source ↗
Figure 9
Figure 9. Figure 9: Visualization. For spatial relation reasoning, VISD accurately localizes the queried child and identifies the object positioned above him, providing precise visual evidence while avoiding confusion with nearby objects. In contrast, related video reasoning models either give incorrect answers or rely on incomplete spatial grounding. 28 view at source ↗
Figure 10
Figure 10. Figure 10: Visualization. For temporal action reasoning, VISD grounds the panda and bucket across relevant frames and correctly infers that the panda is putting itself in the bucket. Competing models miss this action transition and produce incorrect answers. Question: who leans on the baby walker outdoors? Ground Truth Answer: a baby leans on a baby walker outdoors. Open-o3 Video: <think>The video shows a baby sitti… view at source ↗
Figure 11
Figure 11. Figure 11: Visualization. For outdoor spatial interaction reasoning, VISD accurately localizes both the baby and the baby walker, identifies the baby as the one leaning on the walker, and provides precise visual evidence grounded in the relevant temporal window. Other models instead confuse the person with the baby and produce an incorrect answer. 29 view at source ↗
Figure 12
Figure 12. Figure 12: Visualization. For temporal localization reasoning, VISD accurately grounds the person’s gaze direction across frames and identifies the correct time window when the person looks out of the window. Other models focus on earlier or incomplete head movements and produce incorrect intervals, whereas VISD captures the sustained window-looking action and matches the ground-truth period. Question: What changed … view at source ↗
Figure 13
Figure 13. Figure 13: Visualization. In object disappearance reasoning, VISD explicitly grounds the apple before and after the camera movement, correctly identifying that the apple has been removed from the table. Although related models partially recognize the same visual change in their reasoning process, they still produce an inconsistent final answer, highlighting VISD’s stronger alignment between visual grounding, reasoni… view at source ↗
Figure 14
Figure 14. Figure 14: Visualization. VISD not only answers correctly but also localizes the action within the correct temporal window, identifying the moment when the orange kitten kicks the white kitten. Although related video reasoning models produce the correct final answer, they rely on inaccurate temporal evidence. Question: During which time period does the person get up to turn off the light? Ground Truth Answer: 19.70s… view at source ↗
Figure 15
Figure 15. Figure 15: Visualization. For temporal action localization reasoning, VISD accurately grounds the key transition from the lit room to the dark room and identifies the correct time interval when the person gets up to turn off the light. In contrast, related video reasoning models focus on earlier incomplete movements, such as sitting up or approaching the switch, resulting in incorrect temporal predictions. 31 view at source ↗
Figure 16
Figure 16. Figure 16: Visualization. VISD correctly answers the question and localizes the lorry within the ground-truth temporal window. Although other models also produce the correct answer, they ground their reasoning in an incorrect time interval, showing less precise temporal localization. Question: Which activity do the two ghost cowboys enjoy while horse riding as depicted in the video? Ground Truth Answer: C Open-o3 Vi… view at source ↗
Figure 17
Figure 17. Figure 17: Visualization. In fine-grained activity reasoning, VISD accurately grounds the two ghost cowboys during horse riding and identifies their smiling, friendly interaction, correctly inferring that they enjoy singing rather than dancing, fighting, or quarrelling. In contrast, competing models are distracted by later or irrelevant visual cues and produce incorrect answers. 32 view at source ↗
read the original abstract

Training VideoLLMs for complex reasoning remains challenging due to sparse sequence level rewards and the lack of fine grained credit assignment over long, temporally grounded reasoning trajectories. While reinforcement learning with verifiable rewards (RLVR) provides reliable supervision, it fails to capture token level contributions, leading to inefficient learning. Conversely, existing self distillation methods offer dense supervision but lack structure and diagnostic specificity, and often interact unstably with reinforcement learning. In this work, we propose VISD, a structured self distillation framework that introduces diagnostically meaningful privileged information for video reasoning. VISD employs a video aware judge model to decompose reasoning quality into multiple dimensions, including answer correctness, logical consistency, and spatio-temporal grounding, and uses this structured feedback to guide a teacher policy for token level supervision. To stably integrate dense supervision with RL, we introduce a direction magnitude decoupling mechanism, where rollout level advantages computed from rewards determine update direction, while structured privileged signals modulate token level update magnitudes. This design enables semantically aligned and fine grained credit assignment, improving both reasoning faithfulness and training efficiency. Additionally, VISD incorporates curriculum scheduling and EMA based teacher stabilization to support robust optimization over long video sequences. Experiments on diverse benchmarks show that VISD consistently outperforms strong baselines, improving answer accuracy and spatio temporal grounding quality. Notably, VISD reaches these gains with nearly 2x faster convergence in optimization steps, highlighting the effectiveness of structured self supervision in improving both performance and sample efficiency for VideoLLMs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes VISD, a structured self-distillation framework for VideoLLMs. It introduces a video-aware judge model that decomposes reasoning quality into dimensions including answer correctness, logical consistency, and spatio-temporal grounding to supply token-level privileged signals. These are integrated with RL via a direction-magnitude decoupling mechanism (rollout advantages set update direction; privileged signals modulate magnitudes), plus curriculum scheduling and EMA-based teacher stabilization. The central empirical claim is consistent outperformance over strong baselines on diverse benchmarks, with gains in answer accuracy and spatio-temporal grounding quality, achieved at nearly 2x faster convergence in optimization steps.

Significance. If the reported gains hold under rigorous verification, the work would demonstrate a practical way to combine dense structured supervision with verifiable RL rewards for long-horizon video reasoning, addressing sparse credit assignment while maintaining stability. The direction-magnitude decoupling and judge-based decomposition are concrete contributions that could generalize beyond the specific VideoLLM setting.

major comments (1)
  1. [Abstract / Experiments] Abstract and Experiments section: the central claim of 'consistent outperformance' and 'nearly 2x faster convergence' is stated without any reported benchmarks, baselines, metrics (e.g., accuracy, grounding scores), number of runs, variance, or statistical tests. This absence prevents assessment of whether the judge model and decoupling mechanism actually produce the claimed gains.
minor comments (2)
  1. [Method] The description of the judge model as 'video-aware' is introduced without a formal definition or architecture diagram; a concrete specification (e.g., input format, output tokenization) would clarify how the multi-dimensional signals are generated.
  2. [Method] Notation for the direction-magnitude decoupling (advantages vs. magnitudes) is described in prose but would benefit from an explicit equation or pseudocode block to show the precise update rule.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The single major comment highlights an important presentational issue regarding the reporting of empirical results. We address it directly below.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments section: the central claim of 'consistent outperformance' and 'nearly 2x faster convergence' is stated without any reported benchmarks, baselines, metrics (e.g., accuracy, grounding scores), number of runs, variance, or statistical tests. This absence prevents assessment of whether the judge model and decoupling mechanism actually produce the claimed gains.

    Authors: We agree that the current abstract and experiments section do not provide the specific quantitative details needed for independent assessment. The manuscript text supplied to the review process contained only high-level claims without the supporting tables, exact benchmark names, baseline comparisons, metric values, run counts, standard deviations, or significance tests. In the revised version we will (1) revise the abstract to include the key numerical results (e.g., accuracy deltas and convergence-step ratios on each dataset), (2) expand the experiments section with full tables listing all baselines, metrics (answer accuracy, spatio-temporal grounding scores), number of random seeds, variance, and statistical tests, and (3) explicitly link the reported gains to the judge-model dimensions and direction-magnitude decoupling. These additions will make the empirical support verifiable. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained via empirical procedure

full rationale

The paper introduces VISD as a training framework combining a video-aware judge for multi-dimensional feedback with a direction-magnitude decoupling mechanism for RL integration. No equations, fitted parameters, or self-citations are presented that would reduce any claimed prediction or result to an input quantity by construction. The abstract and described method rely on procedural definitions and benchmark experiments for validation, with no load-bearing self-referential steps or uniqueness theorems invoked. This is the standard case of an empirical method paper whose central claims remain independently testable.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review; no explicit free parameters, mathematical axioms, or external benchmarks are stated. The video-aware judge model is introduced as part of the proposed system.

invented entities (1)
  • video-aware judge model no independent evidence
    purpose: decompose reasoning quality into multiple diagnostic dimensions for token-level supervision
    Presented as a new component of VISD; no independent evidence or external validation is mentioned in the abstract.

pith-pipeline@v0.9.0 · 5811 in / 1195 out tokens · 40330 ms · 2026-05-25T06:05:41.663696+00:00 · methodology

Review history (4 revisions) →

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. A Brief Overview: On-Policy Self-Distillation In Large Language Models

    cs.HC 2026-05 unverdicted novelty 2.0

    OPSD lets a single LLM distill its own reasoning by sampling trajectories from the student role while granting the teacher role privileged access to verified solutions, reducing memory needs versus separate-model dist...

  2. A Brief Overview: On-Policy Self-Distillation In Large Language Models

    cs.HC 2026-05 unverdicted novelty 2.0

    This overview paper explains the conceptual foundations and design principles of On-Policy Self-Distillation for large language models from a beginner's perspective.

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages · cited by 1 Pith paper · 25 internal anchors

  1. [1]

    On-policy distillation of language models: Learning from self-generated mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. InThe twelfth international conference on learning representations, 2024

    work page 2024

  2. [2]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Datasets and recipes for video temporal grounding via reinforcement learning

    Ruizhe Chen, Tianze Luo, Zhiting Fan, Heqing Zou, Zhaopeng Feng, Guiyang Xie, Hansheng Zhang, Zhuochen Wang, Zuozhu Liu, and Zhang Huaijian. Datasets and recipes for video temporal grounding via reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 983–992, 2025

    work page 2025

  4. [4]

    Longvila: Scaling long-context visual language models for long videos

    Yukang Chen, Fuzhao Xue, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. InThe Thirteenth International Conference on Learning Representa- tions, 2024

    work page 2024

  5. [5]

    Scaling rl to long videos.arXiv preprint arXiv:2507.07966, 2025

    Yukang Chen, Wei Huang, Baifeng Shi, Qinghao Hu, Hanrong Ye, Ligeng Zhu, Zhijian Liu, Pavlo Molchanov, Jan Kautz, Xiaojuan Qi, et al. Scaling rl to long videos.arXiv preprint arXiv:2507.07966, 2025

    work page arXiv 2025

  6. [6]

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shen- glong 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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    V-star: Bench- marking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495, 2025

    Zixu Cheng, Jian Hu, Ziquan Liu, Chenyang Si, Wei Li, and Shaogang Gong. V-star: Bench- marking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495, 2025

    work page arXiv 2025

  8. [8]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Video-R1: Reinforcing Video Reasoning in MLLMs

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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding

    Chaoyou Fu, Haozhi Yuan, Yuhao Dong, Yi-Fan Zhang, Yunhang Shen, Xiaoxing Hu, Xueying Li, Jinsen Su, Chengwu Long, Xiaoyao Xie, et al. Video-mme-v2: Towards the next stage in benchmarks for comprehensive video understanding.arXiv preprint arXiv:2604.05015, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  11. [11]

    Born again neural networks

    Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. InInternational conference on machine learning, pages 1607–1616. PMLR, 2018

    work page 2018

  12. [12]

    Tall: Temporal activity localization via language query

    Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. InProceedings of the IEEE international conference on computer vision, pages 5267–5275, 2017

    work page 2017

  13. [13]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Trace: Temporal grounding video llm via causal event modeling.arXiv preprint arXiv:2410.05643, 2024

    Yongxin Guo, Jingyu Liu, Mingda Li, Qingbin Liu, Xi Chen, and Xiaoying Tang. Trace: Temporal grounding video llm via causal event modeling.arXiv preprint arXiv:2410.05643, 2024

    work page arXiv 2024

  15. [15]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. 2015

    work page 2015

  16. [16]

    WorldSense: Evaluating Real-world Omnimodal Understanding for Multimodal LLMs

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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Video-MMMU: Evaluating Knowledge Acquisition from Multi-Discipline Professional Videos

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Reinforcement Learning via Self-Distillation

    Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, et al. Reinforcement learning via self-distillation.arXiv preprint arXiv:2601.20802, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    GPT-4o System Card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Visioncoach: Reinforcing grounded video reasoning via visual-perception prompting.arXiv preprint arXiv:2603.14659, 2026

    Daeun Lee, Shoubin Yu, Yue Zhang, and Mohit Bansal. Visioncoach: Reinforcing grounded video reasoning via visual-perception prompting.arXiv preprint arXiv:2603.14659, 2026

    work page arXiv 2026

  21. [21]

    Unifying group-relative and self-distillation policy optimization via sample routing.arXiv preprint arXiv:2604.02288, 2026

    Gengsheng Li, Tianyu Yang, Junfeng Fang, Mingyang Song, Mao Zheng, Haiyun Guo, Dan Zhang, Jinqiao Wang, and Tat-Seng Chua. Unifying group-relative and self-distillation policy optimization via sample routing.arXiv preprint arXiv:2604.02288, 2026

    work page arXiv 2026

  22. [22]

    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

    work page 2024

  23. [23]

    VideoChat-R1: Enhancing Spatio-Temporal Perception via Reinforcement Fine-Tuning

    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 reinforce- ment fine-tuning.arXiv preprint arXiv:2504.06958, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe

    Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu, et al. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Self-hinting language models enhance reinforcement learning.arXiv preprint arXiv:2602.03143, 2026

    Baohao Liao, Hanze Dong, Xinxing Xu, Christof Monz, and Jiang Bian. Self-hinting language models enhance reinforcement learning.arXiv preprint arXiv:2602.03143, 2026

    work page arXiv 2026

  26. [26]

    St-llm: Large language models are effective temporal learners

    Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2024

    work page 2024

  27. [27]

    Oryx mllm: On- demand spatial-temporal understanding at arbitrary resolution.arXiv preprint arXiv:2409.12961, 2024

    Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On- demand spatial-temporal understanding at arbitrary resolution.arXiv preprint arXiv:2409.12961, 2024

    work page arXiv 2024

  28. [28]

    Fipo: Eliciting deep reasoning with future-kl influenced policy optimization.arXiv preprint arXiv:2603.19835, 2026

    Chiyu Ma, Shuo Yang, Kexin Huang, Jinda Lu, Haoming Meng, Shangshang Wang, Bolin Ding, Soroush V osoughi, Guoyin Wang, and Jingren Zhou. Fipo: Eliciting deep reasoning with future-kl influenced policy optimization.arXiv preprint arXiv:2603.19835, 2026

    work page arXiv 2026

  29. [29]

    Video-chatgpt: Towards detailed video understanding via large vision and language models

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12585–12602, 2024. 11

    work page 2024

  30. [30]

    Open-o3 video: Grounded video reasoning with explicit spatio-temporal evidence.arXiv preprint arXiv:2510.20579, 2025

    Jiahao Meng, Xiangtai Li, Haochen Wang, Yue Tan, Tao Zhang, Lingdong Kong, Yunhai Tong, Anran Wang, Zhiyang Teng, Yujing Wang, et al. Open-o3 video: Grounded video reasoning with explicit spatio-temporal evidence.arXiv preprint arXiv:2510.20579, 2025

    work page arXiv 2025

  31. [31]

    SpaceR: Reinforcing MLLMs in Video Spatial Reasoning

    Kun Ouyang, Yuanxin Liu, Haoning Wu, Yi Liu, Hao Zhou, Jie Zhou, Fandong Meng, and Xu Sun. Spacer: Reinforcing mllms in video spatial reasoning.arXiv preprint arXiv:2504.01805, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Simko: Simple pass@ k policy optimization.arXiv preprint arXiv:2510.14807, 2025

    Ruotian Peng, Yi Ren, Zhouliang Yu, Weiyang Liu, and Yandong Wen. Simko: Simple pass@ k policy optimization.arXiv preprint arXiv:2510.14807, 2025

    work page arXiv 2025

  33. [33]

    Policy Distillation

    Andrei A Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, V olodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy distillation.arXiv preprint arXiv:1511.06295, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  34. [34]

    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, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Self-Distillation Enables Continual Learning

    Idan Shenfeld, Mehul Damani, Jonas Hübotter, and Pulkit Agrawal. Self-distillation enables continual learning.arXiv preprint arXiv:2601.19897, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  36. [36]

    Video-xl: Extra-long vision language model for hour-scale video understanding

    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. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26160–26169, 2025

    work page 2025

  37. [37]

    PPLLaVA: Varied Video Sequence Understanding With Prompt Guidance

    Shangkun Sun, Ruyang Liu, Haoran Tang, Yixiao Ge, Haibo Lu, Jiankun Yang, and Chen Li. Ppllava: Varied video sequence understanding with prompt guidance.arXiv preprint arXiv:2411.02327, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [38]

    Mvp: Enhancing video large language models via self-supervised masked video prediction.arXiv preprint arXiv:2601.03781, 2026

    Xiaokun Sun, Zezhong Wu, Zewen Ding, and Linli Xu. Mvp: Enhancing video large language models via self-supervised masked video prediction.arXiv preprint arXiv:2601.03781, 2026

    work page arXiv 2026

  39. [39]

    Traceable evidence enhanced visual grounded reasoning: Evaluation and methodology.arXiv preprint arXiv:2507.07999, 2025

    Haochen Wang, Xiangtai Li, Zilong Huang, Anran Wang, Jiacong Wang, Tao Zhang, Jiani Zheng, Sule Bai, Zijian Kang, Jiashi Feng, et al. Traceable evidence enhanced visual grounded reasoning: Evaluation and methodology.arXiv preprint arXiv:2507.07999, 2025

    work page arXiv 2025

  40. [40]

    Videorft: Incentivizing video reasoning capability in mllms via reinforced fine-tuning.arXiv preprint arXiv:2505.12434, 2025

    Qi Wang, Yanrui Yu, Ye Yuan, Rui Mao, and Tianfei Zhou. Videorft: Incentivizing video reasoning capability in mllms via reinforced fine-tuning.arXiv preprint arXiv:2505.12434, 2025

    work page arXiv 2025

  41. [41]

    Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding

    Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, et al. Time-r1: Post-training large vision language model for temporal video grounding.arXiv preprint arXiv:2503.13377, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  42. [42]

    InternVideo2.5: Empowering Video MLLMs with Long and Rich Context Modeling

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  43. [43]

    Video-rts: Rethinking reinforcement learning and test-time scaling for efficient and enhanced video reasoning

    Ziyang Wang, Jaehong Yoon, Shoubin Yu, Md Mohaiminul Islam, Gedas Bertasius, and Mohit Bansal. Video-rts: Rethinking reinforcement learning and test-time scaling for efficient and enhanced video reasoning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 28114–28128, 2025

    work page 2025

  44. [44]

    Video-ktr: Reinforcing video reasoning via key token attribution.arXiv preprint arXiv:2601.19686, 2026

    Ziyue Wang, Sheng Jin, Zhongrong Zuo, Jiawei Wu, Han Qiu, Qi She, Hao Zhang, and Xudong Jiang. Video-ktr: Reinforcing video reasoning via key token attribution.arXiv preprint arXiv:2601.19686, 2026

    work page arXiv 2026

  45. [45]

    Videochat-r1

    Ziang Yan, Xinhao Li, Yinan He, Zhengrong Yue, Xiangyu Zeng, Yali Wang, Yu Qiao, Limin Wang, and Yi Wang. Videochat-r1. 5: Visual test-time scaling to reinforce multimodal reasoning by iterative perception.arXiv preprint arXiv:2509.21100, 2025. 12

    work page arXiv 2025

  46. [46]

    Self-Distilled RLVR

    Chenxu Yang, Chuanyu Qin, Qingyi Si, Minghui Chen, Naibin Gu, Dingyu Yao, Zheng Lin, Weiping Wang, Jiaqi Wang, and Nan Duan. Self-distilled rlvr.arXiv preprint arXiv:2604.03128, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  47. [47]

    LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling

    Zuhao Yang, Sudong Wang, Kaichen Zhang, Keming Wu, Sicong Leng, Yifan Zhang, Bo Li, Chengwei Qin, Shijian Lu, Xingxuan Li, et al. Longvt: Incentivizing" thinking with long videos" via native tool calling.arXiv preprint arXiv:2511.20785, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  48. [48]

    Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning

    Xiangyu Zeng, Zhiqiu Zhang, Yuhan Zhu, Xinhao Li, Zikang Wang, Changlian Ma, Qingyu Zhang, Zizheng Huang, Kun Ouyang, Tianxiang Jiang, et al. Video-o3: Native interleaved clue seeking for long video multi-hop reasoning.arXiv preprint arXiv:2601.23224, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  49. [49]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  50. [50]

    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. InProceedings of the 2023 conference on empirical methods in natural language processing: system demonstrations, pages 543–553, 2023

    work page 2023

  51. [51]

    Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv preprint arXiv:2508.04416, 2025

    Haoji Zhang, Xin Gu, Jiawen Li, Chixiang Ma, Sule Bai, Chubin Zhang, Bowen Zhang, Zhichao Zhou, Dongliang He, and Yansong Tang. Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv preprint arXiv:2508.04416, 2025

    work page arXiv 2025

  52. [52]

    Llava- video: Video instruction tuning with synthetic data.Transactions on Machine Learning Research, 2025

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Llava- video: Video instruction tuning with synthetic data.Transactions on Machine Learning Research, 2025

    work page 2025

  53. [53]

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models.arXiv preprint arXiv:2601.18734, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  54. [54]

    Group Sequence Policy Optimization

    Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, et al. Group sequence policy optimization.arXiv preprint arXiv:2507.18071, 2025. 13 Appendix Contents A Extended Related Work 15 B Limitations and Future Works 16 C Implementation and Evaluation Details 16 C.1 Algorithm Details . . . . . . ....

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [55]

    Say whether the student final answer is fully correct, partly correct, or wrong

    Answer diagnosis. Say whether the student final answer is fully correct, partly correct, or wrong. Briefly name the problematic part. For natural-language answers, focus on semantic meaning rather than exact wording . For structured answers, say which part is wrong or missing

    work page

  56. [56]

    Judge whether the student’s reasoning broadly supports the final answer

    Reasoning-versus-answer consistency. Judge whether the student’s reasoning broadly supports the final answer. If the reasoning points to one event, object, text clue, time range, or spatial reference but the final answer states another, say so. If the reasoning is too weak, too broad, or too incomplete to justify the final answer, say that clearly

    work page

  57. [57]

    feedback

    High-level error cause. When the response is not fully correct, pick the single most likely high-level cause and mention it briefly. Prefer one of these categories: the reasoning focused on the wrong event, time span, or object; the reasoning was too broad or lacked enough evidence to support such a specific answer; the reasoning was mostly on the right t...

    work page