Pith. sign in

REVIEW 3 major objections 7 minor 72 references

SCOUT: Self-Checking and Recovery-Aware Tool-Thought Agents for Ultra-Long Egocentric Video Reasoning

T0 review · 3 major / 7 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A tool-using agent that can abandon a wrong video region records state-of-the-art results on ultra-long egocentric question answering.

desk verdict A coherent engineering contribution that delivers a plausible training recipe for recovery-aware tool agents, but the self-checking formalism is vacuous, the synthetic-to-real transfer is unverified, and the headline gains over Ego-R1 are size-confounded. read the letter →

arxiv 2608.07959 v1 pith:NGD5ERU3 submitted 2026-08-08 cs.AI

Keyang Zhong , Kuo Wang , Peng Liu , Quanlong Zheng , Junlin Xie , Zhijia Liang , Yanhao Zhang , Guanbin Li This is my paper · ORCID
classification cs.AI
keywords ultra-longegocentricvideoChain-of-Tool-Thoughtagenticsearchrecovery-awarereasoningself-checkingmechanismuncertainty-prioritizedreinforcementlearningturn-levelcreditassignmentsyntheticCoTTtrajectories
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

SCOUT argues that the main obstacle to answering questions over hours of egocentric video is not context length but the inability of tool-using agents to abandon an early, wrong temporal hypothesis. The paper introduces a self-checking Chain-of-Tool-Thought policy that, after each tool observation, chooses between zooming into the current region and switching to a different region entirely. Trained with an uncertainty-prioritized variant of GRPO, turn-level temporal-alignment rewards, and synthetically injected recovery traces, SCOUT reaches 47.6% on EgoLifeQA and 49.0% on Ego-R1 Bench, beating the strongest prior agentic baseline by 9.1 and 6.0 points while staying competitive on shorter-horizon benchmarks. If correct, the work shows that recovery behavior and granular credit assignment, not bigger context windows, are what unlock ultra-long video reasoning.

What carries the argument

The central object is the recovery transition set $T(S_t) = \{S' \mid S' \subseteq S_t \text{ or } S' \sim S_t\}$, where $S_t$ is the candidate temporal region at turn $t$: the policy may zoom in ($S' \subseteq S_t$) or switch to a non-subset region ($S' \sim S_t$), refusing the monotonic constraint $S_{t+1} \subseteq S_t$ of standard Chain-of-Tool-Thought. This set carries the argument by converting "recovery" from a rhetorical goal into a concrete action the policy can be trained to choose. The training machinery is UPS-GRPO: at each post-tool state it samples $m$ continuations and keeps only the highest-uncertainty one, with uncertainty scored by average token log-likelihood, then rescales trajectory-level advantages by $\rho_{i,t} = 1 + \operatorname{sign}(A^{\mathrm{traj}}_i)\tanh(a^{\mathrm{turn}}_{i,t})$, where $a^{\mathrm{turn}}$ is the normalized turn-level reward from temporal alignment (segment IoU or point-wise grounding). The synthetic RA-CoTT data pipeline supplies the recovery behavior via error injection and rewritten reasoning.

What would settle it

Collect tool observations from SCOUT running on EgoLifeQA and re-train the same policy on RA-CoTT trajectories with the recovery-aware rewriting stage removed, then compare accuracy; if accuracy does not drop, the recovery text is not the causal ingredient. Alternatively, run the trained model with the first retrieval call intentionally pointed at a plausible but wrong region; if the model's region-switch rate is uncorrelated with final accuracy, the self-checking mechanism is not doing the claimed work.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that replacing irreversible monotonic zoom-in with an explicit self-checking transition — at every turn either refine the current temporal region or switch to a non-subset region — lets a 7B tool-using policy locate sparse evidence in multi-day egocentric video far more reliably, and that this behavior can be trained with uncertainty-prioritized selection and turn-level reward modulation rather than dense supervision. The evidence is the benchmark table: on EgoLifeQA and Ego-R1 Bench, SCOUT-7B reaches 47.6% and 49.0%, improving over the prior best agent Ego-R1 by +9.1 and +6.0 points, while retaining competitive 63.0% on Video-MME(long) and 35.8% on HourVideo. The paper attributes the gain to the recovery-aware search policy, and its ablations show that removing the turn-level reward or uncertainty-prioritized selection degrades EgoLifeQA and Ego-R1 Bench, while additive reward shaping drops on two of the three main benchmarks.

Load-bearing premise

The load-bearing premise is that the synthetic error-and-recovery trajectories made by large proprietary models look like the real tool feedback the agent encounters at test time; if those revised recovery traces do not match actual tool observations, the learned switching policy will not transfer.

Editorial extensions

If this is right

  • If the claim holds, ultra-long egocentric video QA no longer needs to fit more video into the context; a tool-using policy that can revisit hypotheses will locate sparse evidence even when the first retrieval is wrong.
  • The gains on EgoLifeQA and Ego-R1 Bench (47.6% and 49.0%) would establish recovery-aware search as the dominant factor over the prior agentic baseline, with the gap widening as the horizon grows from 41 minutes to 44 hours.
  • The UPS-GRPO recipe, uncertainty-prioritized branching plus multiplicative turn-level advantages, should make long tool-use RL more sample-efficient, since it reduces average turns and response length while improving accuracy.
  • The SFT+RL training recipe with synthetic recovery traces would transfer to other benchmarks with grounded temporal intervals, because the turn-level reward only requires alignment with ground-truth clue intervals.

Reading between the lines

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

  • A direct testable extension is to instrument the trained agent and measure how often it switches regions after an uninformative tool observation, and whether switching events precede correct answers; the paper reports no such behavioral analysis, only aggregate accuracy.
  • The uncertainty-prioritized selection rule, keeping only the highest-uncertainty continuation, is a general exploration principle that could apply to other multi-turn tool agents beyond video, such as web or code agents, where post-tool states are similarly decision-critical.
  • If recovery behavior is what matters, then end-to-end training on real tool feedback from a weaker policy may be a cheaper route than the three-stage synthetic pipeline, provided the deployed feedback distribution matches training.
  • The turn-level reward based on temporal overlap presupposes access to ground-truth clue intervals; in noisily annotated or unannotated video, the same advantage decomposition could use predicted intervals or self-consistency signals, which the paper does not explore.
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

3 major / 7 minor

Summary. The paper proposes SCOUT, a recovery-aware Chain-of-Tool-Thought agent for ultra-long egocentric video understanding. It replaces monotonic zoom-in search with a self-checking policy that can switch to a different temporal region, and introduces UPS-GRPO, which selects high-uncertainty continuation branches and rescales trajectory-level advantages with turn-level temporal-alignment rewards. A three-stage pipeline (GPT-4o coarse-to-fine generation, Gemini error injection, Gemini recovery reasoning rewrite) produces RA-CoTT for SFT, followed by UPS-GRPO fine-tuning. Experiments on EgoLifeQA, Ego-R1 Bench, Video-MME(long), and HourVideo report state-of-the-art results on the two egocentric benchmarks (47.6% and 49.0%) and competitive results on shorter-horizon settings, with ablations of the training stages and UPS components.

Significance. The empirical claim, if reproducible, is significant: it would be one of the first demonstrations that a tool-using policy trained with explicit recovery trajectories and turn-level grounding rewards can improve evidence localization over 44-hour egocentric streams, and the reported gains over Ego-R1 are large. The controlled UPS-GRPO ablation on a different base model (Table 4) is a useful contribution, and the training-dynamics analysis in Figure 5 supports the efficiency claims. However, the paper does not release code or data, the recovery mechanism is validated only through end-task accuracy, and the formal definition of self-checking is vacuous. Most importantly, the synthetic-to-real transfer of recovery trajectories is unverified, so the central claim that the gains come from a learned recovery-aware policy is not yet established.

major comments (3)
  1. [Section 3.3] The recovery signal is entirely synthetic. Stage II inserts erroneous search segments that are randomly chosen and non-overlapping with the grounded intervals, and Stage III rewrites only the reasoning, not the tool observations, with Gemini-2.5-Pro producing the recovery text. The paper does not verify that these synthetic observation strings resemble the actual RAG/VideoSeg/FrameProbe outputs the policy encounters during RL or at test time, nor does it analyze rollouts to show that the trained model abandons genuinely wrong temporal regions. This is load-bearing because Table 2 already attributes most of the headline gain to SFT alone (EgoLifeQA 44.0 SFT-only vs. 47.6 full; Ego-R1 Bench 47.0 vs. 49.0), and the paper never ablates Stage III recovery rewrites while holding the extra SFT data fixed. The +9.1/+6.0 point gains over Ego-R1 could therefore come from the additional training corpus or from a shallow 'retry after any bad retrieval' heuristic rather than from learned recovery. A convincing test would compare SFT on Stage II trajectories (errors without recovery rewrites) against full RA-CoTT, and would count region switches in test rollouts and correlate them with final-answer correctness.
  2. [Section 3.2.1, Eq. (4)] The transition set T(S_t) = {S' | S' ⊆ S_t or S' ∼ S_t}, with S' ∼ S_t defined as 'S' not a subset of S_t', is a tautology: every possible next region satisfies the disjunction, so Eq. (4) imposes no constraint on the policy and does not formally distinguish SCOUT from monotonic zoom-in CoTT. Because self-checking is the paper's central conceptual contribution, this equation as written cannot support the claim that the policy is explicitly trained or architected for recovery. The authors should either remove the formal framing and present self-checking as an empirical design choice, or replace Eq. (4) with a substantive condition, for example one that depends on the content of the tool observation, retrieval confidence, or a temporal-overlap threshold with the current hypothesis.
  3. [Section 4.3, Table 4] The text states that UPS-GRPO 'consistently improves accuracy over standard GRPO,' but the same table reports Video-MME(long) 52.1 vs. Ego-R1's 53.0 (-0.9), with gains only on EgoLifeQA (+0.9) and Ego-R1 Bench (+5.0). This is a direct contradiction of the word 'consistently.' Please qualify the claim to 'improves on the two egocentric benchmarks' or provide multiple runs and statistical evidence; as written, the claim is not supported by the paper's own numbers.
minor comments (7)
  1. [Table 1] Several numbers are run together without separators (e.g., '42.038.5' for GPT-4o and '67.436.9' for Gemini-1.5-Pro in the HourVideo column); please fix the table formatting.
  2. [Table 1] SCOUT-7B (RL-only) is marked with tool-calling and self-checking checkmarks, yet the text says the RL-only variant 'cannot reliably invoke tools'; the table semantics should be clarified to distinguish architecture-level capabilities from trained behavior.
  3. [Figure 3] The caption for Figure 3(a) contains garbled text ('Policy Group mm m'); please clean up the annotation.
  4. [Section 3.3] The dataset counts are not fully explained: Stage I yields ~2k samples, Stage II 'expands' to ~6k, and the final RA-CoTT is ~8k; please clarify whether inserted turns create new trajectories or are applied in place, since the arithmetic is otherwise unclear.
  5. [Eq. (8)] The symbol I★ is used for both the set of ground-truth intervals and an individual interval; please use distinct symbols and explicitly define M(·,·) for both segment IoU and point-wise grounding.
  6. [References] TSPO appears twice as references [30] and [32] with the same authors and title; please consolidate into a single entry.
  7. [Reproducibility] The paper does not state whether code or data will be released; an availability statement would strengthen reproducibility, especially for the RA-CoTT pipeline and UPS-GRPO implementation.

Circularity Check

1 steps flagged · score 2.0 of 10

The empirical claims are externally evaluated and not circular; the only circularity is a formal self-definition in Eq. (4), where 'self-checking' is definitionally equated with a transition set that includes all possible regions.

  1. self definitional [Section 3.2.1 (Self-Checking Chain-Of-Tool-thought), Eq. (4)]
    "We define a unified transition set for the next-step search region: T(S_t)={S' | S'⊆S_t or S'∼S_t}, where S'⊆S_t denotes zoom-in refinement, and S'∼S_t denotes S' not a subset of S_t. Self-checking is achieved through the aforementioned non-monotonic transformation space, and the policy learns from recovered perceptual trajectories."

    By the law of excluded middle, every candidate next region S' is either a subset of S_t or not a subset of S_t, so T(S_t) is the set of all possible temporal regions. The paper then states that self-checking is 'achieved through' this transformation space, making the self-checking property definitionally equivalent to allowing non-monotonic moves. The formalism imposes no constraint that distinguishes a recovery-aware policy from any policy that may switch regions; the actual recovery behavior is learned from RA-CoTT data and RL rewards, which are external to Eq. (4). Thus the formal derivation of self-checking reduces to a tautology, even though the benchmark results do not.

full rationale

The paper's headline results are not circular: UPS-GRPO rewards use ground-truth temporal intervals and final-answer correctness as supervision, and the reported accuracies are measured on held-out public benchmarks (Video-MME(long), EgoLifeQA, Ego-R1 Bench, HourVideo). No fitted parameter is renamed as a prediction, and no evaluation metric is redefined by the method. The only circularity found is the formal self-definition in Sec. 3.2.1: Eq. (4) defines the transition set as all regions that are either subsets or non-subsets of the current region, which by excluded middle is the set of all regions, and then attributes 'self-checking' to that vacuous space. This is a definitional tautology rather than a source of the benchmark gains. Self-citations in the paper ([19], [36], [47], [63]) appear in related-work and infrastructure contexts and are not load-bearing for the central claim. The RA-CoTT synthetic-to-real transfer concern is a validity or robustness limitation about distribution match, not circularity.

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

The central claims depend on a small set of unreported hyperparameters (m, N), explicit design limits (7 turns, 5 retries), and four assumptions about data quality, uncertainty proxies, and availability of ground-truth temporal intervals. No new physical or formal entities are introduced; SCOUT, UPS-GRPO, and RA-CoTT are artifacts, not postulates.

free parameters (4)
  • candidate continuation count m = not reported
    Eq. 5 samples m continuations per active prefix; m sets the exploration budget for UPS-GRPO and is never given a value.
  • GRPO group size N = not reported
    Eq. 9 normalizes turn-level rewards within groups of size N; N controls the variance of the advantage estimates.
  • max tool-use turns = 7
    Stage I filters trajectories so no sequence exceeds seven tool-use turns, bounding the search horizon the policy learns.
  • generation retry limit = 5
    Stage I repeats GPT-4o trajectory generation up to five times until the final answer is correct, shaping the SFT data distribution.
assumptions (4)
  • domain assumption Ground-truth evidence intervals I* are available for every training QA
    Eq. 8 defines the turn-level reward as max alignment between the predicted region and I*; without these intervals, the turn-level advantage cannot be evaluated.
  • ad hoc to paper Negative mean token log-likelihood is a valid uncertainty score for post-tool states
    Eq. 6-7 define the uncertainty-prioritized selection; the paper gives no evidence that low average likelihood identifies decision-critical or recoverable states.
  • domain assumption GPT-4o and Gemini-2.5-Pro generated CoTT trajectories are faithful training supervision
    Section 3.3 constructs RA-CoTT entirely from proprietary model generation and synthetic error insertion, with only correctness and temporal-overlap filters.
  • ad hoc to paper The transition space T(S_t) in Eq. 4 captures self-checking
    T(S_t) includes every possible next region, so the formal space is vacuous; recovery behavior must come from the synthetic data, not from the definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCOUT: Self-Checking and Recovery-Aware Tool-Thought Agents for Ultra-Long Egocentric Video Reasoning." pith.science (2026). https://pith.science/paper/NGD5ERU3

@misc{pith2026260807959,
  author       = {Pith},
  title        = {Pith review of: SCOUT: Self-Checking and Recovery-Aware Tool-Thought Agents for Ultra-Long Egocentric Video Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NGD5ERU3}},
  note         = {Machine review of arXiv:2608.07959}
}
read the original abstract

Ultra-long egocentric video understanding requires reasoning over temporally sparse evidence distributed across hours or days, challenging current multimodal models with limited context and the grounding of key video segments. While Chain-of-Tool-Thought (CoTT) agent systems enable iterative retrieval and inspection, they suffer from error propagation due to rigid zoom-in strategies that lack recovery mechanisms. In this work, we address these challenges through SCOUT (Self-Checking Chain-Of-Tool-thought), a recovery-aware agentic framework introducing an adaptive policy that evaluates intermediate tool observations and dynamically trades off exploitation (zoom-in) and exploration (region switching), enabling robust multi-hop reasoning over extremely long horizons. However, training such multi-turn tool-using agents remains challenging, as existing RL methods rely on sparse outcome-level rewards and lack supervision over extended decision trajectories, resulting in suboptimal credit assignment for long-horizon reasoning. To address this, we develop UPS-GRPO, an uncertainty-prioritized policy optimization method that concentrates exploration on high-uncertainty post-tool states while preserving sample efficiency. We further introduce a turn-level advantage decomposition that integrates outcome rewards with tool-grounded temporal alignment rewards for improved credit assignment. Experiments show that SCOUT achieves state-of-the-art results on ultra-long egocentric benchmarks, while remaining competitive on shorter-horizon long-video settings.

Figures

Figures reproduced from arXiv: 2608.07959 by the authors.

Figure 1
Figure 1. Comparison of temporal search strategies for ultra [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) An illustration of the process of our proposed self-checking SCOUT. (b) Our three-stage training pipeline, consisting [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. UPS-GRPO. (a) Uncertainty-Prioritized Selection: [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Recovery-aware CoTT data construction pipeline. Training trajectories are constructed through (1) constrained [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Training dynamics of our method and ablations. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

72 extracted references · 20 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al . 2025. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631(2025)

    arXiv 2025

  3. [3]

    Hadzic, Taran Kota, Jimming He, Cristobal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li

    Keshigeyan Chandrasegaran, Agrim Gupta, Lea M. Hadzic, Taran Kota, Jimming He, Cristobal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li

    work page

  4. [4]

    Guo Chen, Yicheng Liu, Yifei Huang, Yuping He, Baoqi Pei, Jilan Xu, Yali Wang, Tong Lu, and Limin Wang. 2024. Cg-bench: Clue-grounded question answering benchmark for long video understanding.arXiv preprint arXiv:2412.12075(2024)

    arXiv 2024

  5. [5]

    Qirui Chen, Shangzhe Di, and Weidi Xie. 2025. Grounded Multi-Hop VideoQA in Long-Form Egocentric Videos.Proceedings of the AAAI Conference on Artificial Intelligence39, 2 (Apr. 2025), 2159–2167. doi:10.1609/aaai.v39i2.32214

    doi:10.1609/aaai.v39i2.32214 2025

  6. [6]

    Yuxiao Chen, Jue Wang, Zhikang Zhang, Jingru Yi, Xu Zhang, Yang Zou, Zhaowei Cai, Jianbo Yuan, Xinyu Li, Hao Yang, and Davide Modolo. 2026. Learning Compact Video Representations for Efficient Long-form Video Understanding in Large Multimodal Models. arXiv:2602.17869 [cs.CV] https://arxiv.org/abs/2602. 17869

    work page arXiv 2026

  7. [7]

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

    arXiv 2025

  8. [8]

    Yifeng Ding, Hung Le, Songyang Han, Kangrui Ruan, Zhenghui Jin, Varun Kumar, Zijian Wang, and Anoop Deoras. 2025. Empowering Multi-Turn Tool-Integrated Reasoning with Group Turn Policy Optimization. arXiv:2511.14846 [cs.LG] https://arxiv.org/abs/2511.14846

    arXiv 2025

Show all 72 references
  1. [9]

    Yang Ding, Yizhen Zhang, Xin Lai, Ruihang Chu, and Yujiu Yang. 2025. Video- Zoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning. arXiv preprint arXiv:2512.22315(2025)

    2025

  2. [10]

    Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, Guorui Zhou, Yutao Zhu, Ji-Rong Wen, and Zhicheng Dou. 2025. Agentic Reinforced Policy Optimization. arXiv:2507.19849 [cs.CL] https://arxiv.org/a...

    2025 arXiv

  3. [11]

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

    2025 arXiv

  4. [12]

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. 2025. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. InCVPR

    2025

  5. [13]

    Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrish- nan, Fiona Ryan, Jayant Sharma, Michael Wray, Meng...

    2022 arXiv

  6. [14]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda ...

    2025

  7. [15]

    Dongfu Jiang, Yi Lu, Zhuofeng Li, Zhiheng Lyu, Ping Nie, Haozhe Wang, Alex Su, Hui Chen, Kai Zou, Chao Du, Tianyu Pang, and Wenhu Chen. 2025. VerlTool: Towards Holistic Agentic Reinforcement Learning with Tool Use. arXiv:2509.01055 [cs.AI] https://arxiv.org/abs/2509.01055

    2025

  8. [16]

    Bin Li, Jiahao Li, Yan Lu, Xiaoyi Zhang, Houqiang Li, Zongyu Guo, and Zhaoyang Jia. 2025. Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding. arXiv:2505.18079 [cs.CV] https://arxiv.org/abs/2505.18079

    2025

  9. [17]

    Chenglin Li, Qianglong Chen, Feng Han, Yikun Wang, Xingxi Yin, Yan Gong, Ruilin Li, Yin Zhang, and Jiaqi Wang. 2026. VideoThinker: Building Agentic VideoLLMs with LLM-Guided Tool Reasoning. arXiv:2601.15724 [cs.CV] https: //arxiv.org/abs/2601.15724

    2026 arXiv

  10. [18]

    Jiaxiang Li, Siliang Zeng, Hoi-To Wai, Chenliang Li, Alfredo Garcia, and Mingyi Hong. 2024. Getting More Juice Out of the SFT Data: Reward Learning from Human Demonstration Improves SFT for LLM Alignment. InAdvances in Neural Information Processing Systems, A. Globerson, L. Ma...

    2024 doi

  11. [19]

    Zhijia Liang, Jiaming Li, Weikai Chen, Yanhao Zhang, Haonan Lu, and Guanbin Li. 2026. OASIS: On-Demand Hierarchical Event Memory for Streaming Video Reasoning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2821–2831

    2026

  12. [20]

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s Verify Step by Step. arXiv:2305.20050 [cs.LG] https://arxiv.org/abs/2305.20050

    2023 arXiv

  13. [21]

    Jingyang Lin, Jialian Wu, Jiang Liu, Ximeng Sun, Ze Wang, Xiaodong Yu, Jiebo Luo, Zicheng Liu, and Emad Barsoum. 2026. VideoSeek: Long-Horizon Video Agent with Tool-Guided Seeking.arXiv preprint arXiv:2603.20185(2026)

    2026

  14. [22]

    Sachit Menon, Ahmet Iscen, Arsha Nagrani, Tobias Weyand, Carl Vondrick, and Cordelia Schmid. 2025. CAViAR: Critic-Augmented Video Agentic Reasoning. arXiv:2509.07680 [cs.CV] https://arxiv.org/abs/2509.07680

    2025 arXiv

  15. [23]

    Junwen Pan, Qizhe Zhang, Rui Zhang, Ming Lu, Xin Wan, Yuan Zhang, Chang Liu, and Qi She. 2025. TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning.arXiv preprint arXiv:2511.05489(2025)

    2025

  16. [24]

    Jihao Qiu, Lingxi Xie, Xinyue Huo, Qi Tian, and Qixiang Ye. 2026. LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding. arXiv:2602.20913 [cs.CV] https://arxiv.org/abs/2602.20913

    2026 arXiv

  17. [25]

    Aniket Rege, Arka Sadhu, Yuliang Li, Kejie Li, Ramya Korlakai Vinayak, Yuning Chai, Yong Jae Lee, and Hyo Jin Kim. 2026. Agentic Very Long Video Under- standing. arXiv:2601.18157 [cs.CV] https://arxiv.org/abs/2601.18157

    2026 arXiv

  18. [26]

    Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bor- des, et al. 2024. Longvu: Spatiotemporal adaptive compression for long video- language understanding.arXiv preprint arXiv:2410.17434(2024)

    2024 arXiv

  19. [27]

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. HybridFlow: A Flexible and Efficient RLHF Framework.arXiv preprint arXiv: 2409.19256(2024)

    2024 arXiv

  20. [28]

    Yudi Shi, Shangzhe Di, Qirui Chen, Qinian Wang, Jiayin Cai, Xiaolong Jiang, Yao Hu, and Weidi Xie. 2026. Weaver: End-to-End Agentic System Training for Video Interleaved Reasoning. arXiv:2602.05829 [cs.CV] https://arxiv.org/abs/2602.05829

    2026

  21. [29]

    Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao. 2024. Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding. arXiv:2409.14485 [cs.CV] https: //arxiv.org/abs/2409.14485

    2024 arXiv

  22. [30]

    Hao Sun, Ye Yuan, Jinglin Xu, Hongbo Sun, Xin Wei, Sanping Zhou, Xu- chong Zhang, Canhui Tang, and Zifan Han. 2025. TSPO: Temporal Sam- pling Policy Optimization for Long-form Video Language Understanding. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Zhong et al. arXiv...

    2025

  23. [31]

    Shitong Sun, Ke Han, Yukai Huang, Weitong Cai, and Jifei Song. 2026. Ego- Graph: Temporal Knowledge Graph for Egocentric Video Understanding. arXiv:2602.23709 [cs.CV] https://arxiv.org/abs/2602.23709

    2026

  24. [32]

    Canhui Tang, Zifan Han, Hongbo Sun, Sanping Zhou, Xuchong Zhang, Xin Wei, Ye Yuan, Huayu Zhang, Jinglin Xu, and Hao Sun. 2026. TSPO: Temporal Sampling Policy Optimization for Long-form Video Language Understanding.Proceedings of the AAAI Conference on Artificial Intelligence40...

    2026 doi

  25. [33]

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

    2024 arXiv

  26. [34]

    Qwen Team. 2024. Qwen2.5: A Party of Foundation Models. https://qwenlm. github.io/blog/qwen2.5/

    2024

  27. [35]

    Hanlin Wang, Chak Tou Leong, Jiashuo Wang, Jian Wang, and Wenjie Li

  28. [36]

    Kuo Wang, Quanlong Zheng, Junlin Xie, Yanhao Zhang, Jinguo Luo, Haonan Lu, Liang Lin, Fan Zhou, and Guanbin Li. 2025. Free-moref: Instantly multiplexing context perception capabilities of video-mllms within single inference. In2025 IEEE/CVF International Conference on Computer...

    2025

  29. [37]

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. 2025. InternVL3.5: Ad- vancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency. arXiv preprint arXiv:2508.18265(2025)

    2025 arXiv

  30. [38]

    Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. 2024. Videoa- gent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision. Springer, 58–76

    2024

  31. [39]

    Zheng Wang, Haoran Chen, Haoxuan Qin, Zhipeng Wei, Tianwen Qian, and Cong Bai. 2026. Think, Then Verify: A Hypothesis-Verification Multi-Agent Framework for Long Video Understanding. arXiv:2603.04977 [cs.CV] https: //arxiv.org/abs/2603.04977

    2026

  32. [40]

    Ryoo, and Juan Carlos Niebles

    Ziyang Wang, Honglu Zhou, Shijie Wang, Junnan Li, Caiming Xiong, Silvio Savarese, Mohit Bansal, Michael S. Ryoo, and Juan Carlos Niebles. 2025. Active Video Perception: Iterative Evidence Seeking for Agentic Long Video Under- standing. arXiv:2512.05774 [cs.CV] https://arxiv.or...

    2025 arXiv

  33. [41]

    Quan Wei, Siliang Zeng, Chenliang Li, William Brown, Oana Frunza, Wei Deng, Anderson Schneider, Yuriy Nevmyvaka, Yang Katie Zhao, Alfredo Garcia, and Mingyi Hong. 2025. Reinforcing Multi-Turn Reasoning in LLM Agents via Turn- Level Reward Design. arXiv:2505.11821 [cs.LG] https...

    2025

  34. [42]

    Siwei Wen, Zhangcheng Wang, Xingjian Zhang, Lei Huang, and Wenjun Wu

  35. [43]

    Fei Wu, Zhenrong Zhang, Qikai Chang, Jianshu Zhang, Quan Liu, and Jun Du

  36. [44]

    Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. 2024. LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding. arXiv:2407.15754 [cs.CV] https://arxiv.org/abs/2407.15754

    2024 arXiv

  37. [45]

    Zhengxian Wu, Chuanrui Zhang, Shen’Ao Jiang, Hangrui Xu, Zirui Liao, Luyuan Zhang, Li Huaqiu, Peng Jiao, and Haoqian Wang. 2026. Language-guided and motion-aware gait representation for generalizable recognition. InProceedings of the AAAI Conference on Artificial Intelligence,...

    2026

  38. [46]

    Zhengxian Wu, Chuanrui Zhang, Hangrui Xu, Peng Jiao, and Haoqian Wang

  39. [47]

    arXiv:2601.03823 [cs.CL] https://arxiv.org/abs/2601.03823

    Step Potential Advantage Estimation: Harnessing Intermediate Confidence and Correctness for Efficient Mathematical Reasoning. arXiv:2601.03823 [cs.CL] https://arxiv.org/abs/2601.03823

  40. [48]

    Hangrui Xu, Zhengxian Wu, Chuanrui Zhang, Zhuohong Chen, Zhifang Liu, Peng Jiao, and Haoqian Wang. 2026. Psgait: Gait recognition using parsing skeleton. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 10427–10431

    2026

  41. [49]

    Chen Yang, Ran Le, Yun Xing, Zhenwei An, Zongchao Chen, Wayne Xin Zhao, Yang Song, and Tao Zhang. 2025. ToolMind Technical Report: A Large-Scale, Reasoning-Enhanced Tool-Use Dataset. arXiv:2511.15718 [cs.CL] https://arxiv. org/abs/2511.15718

    2025

  42. [50]

    Jingkang Yang, Shuai Liu, Hongming Guo, Yuhao Dong, Xiamengwei Zhang, Sicheng Zhang, Pengyun Wang, Zitang Zhou, Binzhu Xie, Ziyue Wang, Bei Ouyang, Zhengyu Lin, Marco Cominelli, Zhongang Cai, Yuanhan Zhang, Peiyuan Zhang, Fangzhou Hong, Joerg Widmer, Francesco Gringoli, Lei Ya...

    2025

  43. [51]

    In2025 IEEE International Conference on Multimedia and Expo (ICME)

    DAGait: Generalized skeleton-guided data alignment for gait recognition. In2025 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6

  44. [52]

    Junlin Xie, Quanlong Zheng, Ruifei Zhang, Kuo Wang, Yanhao Zhang, Jinguo Luo, Haonan Lu, Xiang Wan, and Guanbin Li. 2026. StreamRAG: Enhancing Real- Time Video Understanding with Retrieval Augmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2026

  45. [53]

    Jinhui Ye, Zihan Wang, Haosen Sun, Keshigeyan Chandrasegaran, Zane Durante, Cristobal Eyzaguirre, Yonatan Bisk, Juan Carlos Niebles, Ehsan Adeli, Li Fei-Fei, Jiajun Wu, and Manling Li. 2025. T*: Re-thinking Temporal Search for Long-Form Video Understanding. arXiv:2504.02259 [c...

    2025 arXiv

  46. [54]

    Junkeun Yi, Damon Mosk-Aoyama, Baihe Huang, Ritu Gala, Charles Wang, Sugam Dipak Devare, Khushi Bhardwaj, Abhibha Gupta, Oleksii Kuchaiev, Jiantao Jiao, Jian Zhang, and Venkat Srinivasan. 2026. PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost. arXiv:2603.21383 ...

    2026

  47. [55]

    Xinlei Yin, Xiulian Peng, Xiao Li, Zhiwei Xiong, and Yan Lu. 2026. Hierarchical Long Video Understanding with Audiovisual Entity Cohesion and Agentic Search. arXiv:2601.13719 [cs.AI] https://arxiv.org/abs/2601.13719

    2026

  48. [56]

    Jingkang Yang, Ziwei Liu, Hao Zhang, Penghao Wu, Ruiqi Wang, Yuhao Dong, Hongyuan Zhu, Shulin Tian, Hongming Guo, and Xiuying Wang. 2025. Ego-R1: Chain-of-Tool-Thought for Ultra-Long Egocentric Video Reasoning. arXiv:2506.13654 [cs.CV] https://arxiv.org/abs/2506.13654

    2025 arXiv

  49. [57]

    Thinking with Long Videos

    Zuhao Yang, Sudong Wang, Kaichen Zhang, Keming Wu, Sicong Leng, Yifan Zhang, Bo Li, Chengwei Qin, Shijian Lu, Xingxuan Li, and Lidong Bing. 2025. LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling. arXiv preprint arXiv:2511.20785(2025)

    2025 arXiv

  50. [58]

    Congzhi Zhang, Zhibin Wang, Yinchao Ma, Jiawei Peng, Yihan Wang, Qiang Zhou, Jun Song, and Bo Zheng. 2025. ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis. arXiv preprint arXiv:2509.23652(2025)

    2025

  51. [59]

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chun- yuan Li. 2025. LLaVA-Video: Video Instruction Tuning With Synthetic Data. arXiv:2410.02713 [cs.CV] https://arxiv.org/abs/2410.02713

    2025 arXiv

  52. [60]

    Yabo Zhang, Yihan Zeng, Qingyun Li, Zhen Hu, Kavin Han, and Wangmeng Zuo

  53. [61]

    Yufei Yin, Qianke Meng, Minghao Chen, Jiajun Ding, Zhenwei Shao, and Zhou Yu

  54. [62]

    arXiv:2512.12360 [cs.CV] https://arxiv.org/abs/2512.12360

    VideoARM: Agentic Reasoning over Hierarchical Memory for Long-Form Video Understanding. arXiv:2512.12360 [cs.CV] https://arxiv.org/abs/2512.12360

  55. [63]

    Xiangyu Zeng, Zhiqiu Zhang, Yuhan Zhu, Xinhao Li, Zikang Wang, Changlian Ma, Qingyu Zhang, Zizheng Huang, Kun Ouyang, Tianxiang Jiang, Ziang Yan, Yi Wang, Hongjie Zhang, Yali Wang, and Limin Wang. 2026. Video- o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reason...

    2026 arXiv

  56. [64]

    Mayol-Cuevas, Fan Zhang, Weizhe Lin, and Junx- iao Shen

    Wenqi Zhou, Kai Cao, Hao Zheng, Yunze Liu, Xinyi Zheng, Miao Liu, Per Ola Kristensson, Walterio W. Mayol-Cuevas, Fan Zhang, Weizhe Lin, and Junx- iao Shen. 2025. X-LeBench: A Benchmark for Extremely Long Egocentric Video Understanding. InFindings of the Association for Computa...

    2025 doi

  57. [65]

    Heqing Zou, Tianze Luo, Guiyang Xie, Victor Xiao Jie Zhang, Fengmao Lv, Guangcong Wang, Junyang Chen, Zhuochen Wang, Hansheng Zhang, and Huai- jian Zhang. 2025. HLV-1K: A Large-scale Hour-Long Video Benchmark for Time-Specific Long Video Understanding. arXiv:2501.01645 [cs.CV]...

    2025 arXiv

  58. [67]

    arXiv:2509.12867 [cs.CL] https://arxiv.org/abs/2509.12867

    Tool-R1: Sample-Efficient Reinforcement Learning for Agentic Tool Use. arXiv:2509.12867 [cs.CL] https://arxiv.org/abs/2509.12867

  59. [68]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: Sy...

    2024 arXiv

  60. [69]

    Haitian Zhong, Jixiu Zhai, Lei Song, Jiang Bian, Qiang Liu, and Tieniu Tan. 2026. RC-GRPO: Reward-Conditioned Group Relative Policy Optimization for Multi- Turn Tool Calling Agents. arXiv:2602.03025 [cs.AI] https://arxiv.org/abs/2602. 03025

    2026

  61. [70]

    Keyang Zhong, Junlin Xie, Hefeng Wu, Haofeng Li, and Guanbin Li. 2026. Col- laborative Multi-Agent Scripts Generation for Enhancing Imperfect-Information Reasoning in Murder Mystery Games. InFindings of the Association for Computa- tional Linguistics: ACL 2026. 15796–15819

    2026

  62. [2024]

    InAdvances in Neural Information Processing Systems, Vol

    HourVideo: 1-Hour Video-Language Understanding. InAdvances in Neural Information Processing Systems, Vol. 37

  63. [2025]

    arXiv:2505.20732 [cs.CL] https://arxiv.org/abs/2505.20732

    SPA-RL: Reinforcing LLM Agents via Stepwise Progress Attribution. arXiv:2505.20732 [cs.CL] https://arxiv.org/abs/2505.20732

    arXiv

  64. [2026]

    arXiv:2602.15329 [cs.CV] https://arxiv

    EventMemAgent: Hierarchical Event-Centric Memory for Online Video Understanding with Adaptive Tool Use. arXiv:2602.15329 [cs.CV] https://arxiv. org/abs/2602.15329

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.