Pith. sign in

REVIEW 5 major objections 5 minor 58 references

Video-CoT: A Comprehensive Dataset for Spatiotemporal Understanding of Videos Based on Chain-of-Thought

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

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

Pith's one-line read A chain-of-thought video dataset lets a 3B vision-language model beat larger open-source models at spatiotemporal reasoning.

desk verdict A useful dataset paper whose headline claim is internally contradicted by its own Table 1; fix that before anything else. read the letter →

arxiv 2506.08817 v3 pith:PL42M4VB submitted 2025-06-10 cs.CV

Shuyi Zhang , Xiaoshuai Hao , Yingbo Tang , Lingfeng Zhang , Pengwei Wang , Zhongyuan Wang , Hongxuan Ma , Shanghang Zhang This is my paper
classification cs.CV
keywords videounderstandingchain-of-thoughtspatiotemporalreasoningvision-languagemodelstemporallocalizationgroundingdatasetbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces Video-CoT, a dataset of 192,000 fine-grained spatiotemporal question-answer pairs plus 23,000 chain-of-thought annotated samples built from existing video datasets. The dataset spans six tasks grouped into spatio-temporal localization and captioning, grounding, and reference. The paper's central claim is that fine-tuning a vision-language model on these reasoning chains improves spatiotemporal understanding more than training on final answers alone. Evidence includes a 3-billion-parameter Qwen2.5-VL fine-tuned with CoT-SFT reaching 19.7 tIoU on temporal video localization, surpassing the open-source 7-billion-parameter models the authors compare against. The accompanying 4,500-item benchmark reveals that current VLMs, including closed-source models, perform poorly on all six tasks.

What carries the argument

The load-bearing mechanism is the CoT-SFT objective, which trains the model to output a reasoning chain $R$ before the final answer $A$, using a two-term loss: log-likelihood of each reasoning step conditioned on the video and question, plus a weighted log-likelihood of the answer conditioned on the completed reasoning chain. The reasoning chains themselves were produced by a large vision-language model inside <think> and </think> tags, kept only when their final answers passed task-metric thresholds, and then manually curated. A curriculum learning strategy progressively increases reasoning chain complexity from simple 2-step scenarios to 5-step chains. This setup lets a small 3B model serve as a testbed showing that the reasoning text, not just the answer, carries much of the spatiotemporal task gain.

What would settle it

Have independent annotators score every intermediate reasoning step in the retained chains against the video facts. If a substantial fraction of chains contain wrong intermediate statements while CoT-SFT still improves scores, the gains are attributable to answer filtering or output format rather than to the reasoning content; if removing chains with flawed steps eliminates the improvement, the reasoning text itself is load-bearing.

Watch

Extended reading notes

Core claim

Video-CoT is organized into three key components: spatio-temporal localization and captioning, spatio-temporal grounding, and spatio-temporal reference. Chain-of-thought samples are generated by Qwen2.5-VL-72B-Instruct, filtered using task-specific thresholds derived from average task accuracy and prior best results, and then manually reviewed. Fine-tuning Qwen2.5-VL-3B with CoT-SFT produces consistent gains over both the untreated baseline and answer-only supervised fine-tuning, with the largest improvements on temporally oriented metrics: temporal video localization tIoU rises from 4.4 to 19.7, and temporal video reference MENTOR rises from 14.8 to 21.1. The paper argues that these results show chain-of-thought supervision specifically strengthens a model's capacity for long-term spatiotemporal reasoning.

Load-bearing premise

The paper's conclusions depend on the assumption that the 23,000 reasoning chains kept after automatic answer verification and manual review are correct reasoning targets, even though the automatic check only validates final answers and the manual review is not quantified.

Editorial extensions

If this is right

  • CoT-SFT on Video-CoT raises Qwen2.5-VL-3B temporal video localization tIoU from 4.4 to 19.7, above every open-source 7B model the paper reports.
  • Chain-of-thought supervision helps most on temporally oriented metrics, while answer-supervised fine-tuning also improves all tasks but by smaller margins.
  • Current VLMs, including closed-source GPT-4o and Gemini-1.5-pro, score low on fine-grained spatiotemporal tasks, with the best STVG sIoU reaching only 9.9.
  • The released benchmark's 4,500 disjoint question-answer pairs provide a reusable testbed for evaluating spatiotemporal reasoning across six subtasks.
  • Fine-tuning on structured reasoning chains improves spatial relationship exact match from 10.0 to 16.9, suggesting the dataset also strengthens spatial reference reasoning, not only temporal localization.

Reading between the lines

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

  • If the answer-only filter is what drives the observed gains, then ablating the <think> reasoning text while keeping the same final answers should yield similar improvements; this is directly testable with the released dataset.
  • The reasoning chains generated by a 72B model could serve as synthetic rationales for other video models, but the unquantified manual review means quality may vary across tasks and video lengths.
  • A natural extension is to verify intermediate reasoning steps with detection models or rule-based checkable facts, turning CoT-SFT into verifiable-reasoning training rather than imitation of a large model's output style.
  • The per-frame bounding boxes and timestamps in the dataset could support joint temporal grounding and trajectory prediction objectives beyond question answering, 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

5 major / 5 minor

Summary. The paper introduces Video-CoT, a dataset of 192,000 fine-grained spatiotemporal question-answer pairs and 23,000 Chain-of-Thought (CoT) annotated samples across six video-understanding tasks (TVL, VC, SRR, TVR, SVG, STVG), together with a held-out benchmark of 4,500 video QA pairs. The authors generate CoT annotations with Qwen2.5-VL-72B-Instruct, filter them via task-specific metrics and manual review, and then fine-tune Qwen2.5-VL-3B using both answer-only SFT (Ans-SFT) and CoT-supervised SFT (CoT-SFT). They report substantial gains for CoT-SFT over the 3B baseline and claim that the 3B CoT-SFT model surpasses all open-source 7B models listed in their benchmark on the TVL task.

Significance. If the dataset and benchmark are reliable, Video-CoT would be a valuable community resource: it addresses a real gap in fine-grained spatiotemporal reasoning, covers six tasks with tailored metrics, and is publicly released. The systematic evaluation of current VLMs—particularly the low STVG scores—is a useful and credible finding that aligns with the literature. The paper also makes a reproducible methodology contribution by pairing a large QA dataset with structured CoT annotations. However, the central demonstration of CoT-SFT's benefit is undermined by an internal numeric contradiction and by missing statistical detail, so the significance is currently conditional on correcting these issues.

major comments (5)
  1. [Abstract and Section 5.2 vs. Table 1] The abstract states that 'a 3B Qwen2.5-VL reaches 19.7 tIoU on TVL, surpassing the open-source 7B models listed in Table 1.' This is directly contradicted by Table 1, where Qwen2.5-VL-7B reports a TVL tIoU of 20.0, which is strictly greater than 19.7. Section 5.2 hedges the claim to 'all other open-source 7B models,' but the 7B Qwen2.5-VL model is itself an open-source 7B model listed in Table 1, so the hedged version is also false. Since this comparison is the headline evidence for the effectiveness of CoT-SFT, the claim must be corrected, and the comparison should be rerun or re-reported with proper statistical support.
  2. [Section 3.2] The CoT filtering thresholds are not specified. The text says thresholds are established 'based on the average accuracy of individual tasks and the best-performing results from prior studies,' but no concrete threshold values, task-wise criteria, or reference results are given. Without these numbers, the filtering procedure is not reproducible and the claim that retained CoT samples are 'high-quality' cannot be independently assessed. The authors should report the exact thresholds per task and the number of samples retained at each filtering stage.
  3. [Section 3.2 and Figure 4] The quality-control pipeline verifies only the final answers against ground-truth annotations; it does not validate the correctness or faithfulness of the intermediate reasoning chains. The manual expert review is mentioned but not quantified (e.g., number of experts, number of samples reviewed, inter-annotator agreement). Because the dataset's core contribution is the CoT annotations, the lack of validation of the reasoning text itself is a load-bearing gap: CoT-SFT may simply imitate the generator's (Qwen2.5-VL-72B) reasoning patterns, including potential hallucinations or biases. The authors should either provide a quantitative assessment of reasoning-chain quality or clearly state this as a limitation.
  4. [Section 4.2, Eq. (5)] The CoT-SFT loss in Eq. (5) contains an unconstrained hyperparameter λ that balances reasoning-step and final-answer likelihoods. No value or selection procedure for λ is reported anywhere in the paper, and the curriculum learning strategy described in the text (progressing from 2-step to 5-step chains) is not specified in terms of how steps are counted, how data are scheduled, or what hyperparameters are used. These omissions prevent replication of the proposed method and make it impossible to attribute the reported gains to the CoT supervision rather than to λ tuning or curriculum details.
  5. [Tables 1 and 2] All benchmark numbers are reported without error bars, standard deviations, or significance tests, and no information is given about the number of evaluation seeds or runs. This is particularly problematic for the flagship TVL comparison, where the 3B CoT-SFT model scores 19.7 and the 7B Qwen2.5-VL baseline scores 20.0; a 0.3 tIoU gap from a single run cannot support a meaningful superiority claim. The authors should provide multi-seed results with measures of variance, or at least state explicitly that the differences are not statistically tested.
minor comments (5)
  1. [Abstract and affiliations] There are several typographical errors, including 'spa-tiotemporal' in the abstract and 'Artifcial' in the author affiliation; these should be corrected.
  2. [Section 5.1] The MENTOR metric is used for VC and TVR but is never defined or cited. Please provide a definition or reference, and state whether it is a learned metric or a rule-based score.
  3. [Table 2] The STVG row for Video-CoT-SFT shows '(5.6↑)' and '(3.5↑)' without the leading plus signs used in the rest of the table, which is inconsistent formatting.
  4. [Section 3.2 and Section 5.1] The benchmark is described as 'each task featuring 750 images' in the abstract and introduction, but the tasks are video-based and Section 5.1 states there are 4,500 video question-answer pairs in total. Please clarify whether the 750 is per-task videos or QA pairs, and make the wording consistent throughout.
  5. [References] Several references have formatting issues, for example [28] lists author names out of order and [50] uses incomplete author initials; please ensure all entries follow the venue's style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central CoT-SFT result is measured on a held-out benchmark, not reduced to fitted inputs or self-citation.

full rationale

The paper's core empirical claim is that fine-tuning Qwen2.5-VL-3B on the Video-CoT chain-of-thought annotations improves spatiotemporal reasoning, with CoT-SFT reaching 19.7 tIoU on TVL. This is a held-out evaluation: Section 5.1 states the 4,500 benchmark pairs are 'entirely distinct from the training dataset,' so the improvement is not forced by the construction of the training set. The closest circularity-adjacent mechanism is the CoT filtering in Section 3.2, where thresholds are set from task accuracy and prior best results and only samples whose answers exceed them are retained. That filtering selects training examples, but it does not determine the held-out benchmark scores; generalization from filtered training data to a disjoint test set is an independent empirical measurement, not a quantity equal by construction to the threshold. The manual review step is described but not quantified, and the CoT chains are generated by Qwen2.5-VL-72B, which raises a data-quality limitation but not a circular derivation. The paper also cites several prior works by its own authors, but these are contextual or application references, not load-bearing justifications for the dataset's validity or the CoT-SFT result. The abstract's claim that the 3B model 'surpasses the open-source 7B models listed in Table 1' is internally inconsistent with Table 1, where Qwen2.5-VL-7B scores 20.0 tIoU on TVL versus 19.7 for the 3B CoT-SFT model; however, this is a factual/correctness inconsistency, not a circularity of derivation. Overall, no prediction or claimed result reduces by construction to its inputs, so the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central resource rests on reusing existing annotated datasets and on unverified model-generated reasoning chains; no new physical or conceptual entities are introduced. The unspecified loss weight and filtering thresholds are the main free parameters.

free parameters (2)
  • lambda (CoT-SFT loss weight) = not reported
    In Eq. 5, lambda balances the reasoning-chain loss and the final-answer loss; its value is not given, so the training objective is not fully specified.
  • CoT filtering thresholds = not reported
    Section 3.2 says thresholds are based on average task accuracy and best prior results, but exact values are omitted, making the dataset construction non-reproducible.
assumptions (3)
  • domain assumption Source datasets (VTimeLLM-stage2, MSR-VTT, MSVD, WebVid, HCSTVG-V1, VidSTG) contain accurate annotations that can be reused as ground truth.
    The dataset construction in Section 3.2 inherits the quality of these public datasets without validating them.
  • ad hoc to paper Qwen2.5-VL-72B-Instruct produces correct reasoning chains that are suitable training targets after filtering.
    The CoT samples are generated by this model (Section 3.2) and the paper provides no independent verification of reasoning-chain correctness beyond final-answer metrics.
  • domain assumption Task-specific metrics (tIoU, MENTOR, EM, sIoU) capture the quality of spatiotemporal reasoning.
    Section 5.1 applies these metrics as the evaluation and filtering criteria, but their adequacy for reasoning quality is not justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Video-CoT: A Comprehensive Dataset for Spatiotemporal Understanding of Videos Based on Chain-of-Thought." pith.science (2026). https://pith.science/paper/PL42M4VB

@misc{pith2026250608817,
  author       = {Pith},
  title        = {Pith review of: Video-CoT: A Comprehensive Dataset for Spatiotemporal Understanding of Videos Based on Chain-of-Thought},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PL42M4VB}},
  note         = {Machine review of arXiv:2506.08817}
}
read the original abstract

Video content comprehension is essential for various applications, ranging from video analysis to interactive systems. Despite advancements in large-scale vision-language models (VLMs), these models often struggle to capture the nuanced, spatiotemporal details essential for thorough video analysis. To address this gap, we introduce Video-CoT, a groundbreaking dataset designed to enhance spatiotemporal understanding using Chain-of-Thought (CoT) methodologies. Video-CoT contains 192,000 fine-grained spa-tiotemporal question-answer pairs and 23,000 high-quality CoT-annotated samples, providing a solid foundation for evaluating spatiotemporal understanding in video comprehension. Additionally, we provide a comprehensive benchmark for assessing these tasks, with each task featuring 750 images and tailored evaluation metrics. Our extensive experiments reveal that current VLMs face significant challenges in achieving satisfactory performance, high-lighting the difficulties of effective spatiotemporal understanding. Overall, the Video-CoT dataset and benchmark open new avenues for research in multimedia understanding and support future innovations in intelligent systems requiring advanced video analysis capabilities. By making these resources publicly available, we aim to encourage further exploration in this critical area. Project website:https://video-cot.github.io/ .

Figures

Figures reproduced from arXiv: 2506.08817 by the authors.

Figure 1
Figure 1. Overview of the Video-CoT dataset. The Video-CoT dataset encompasses three key dimensions: spatio-temporal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. The video length statistic. descriptions of video content, enabling more intuitive interaction with video data across various contexts. Spatio-Temporal Grounding aims to align textual queries with corresponding spatio-temporal regions in videos, integrating spatial and temporal reasoning to enhance multimodal understanding. This task includes two components: Spatial Video Grounding (SVG), which focuses on locating s… view at source ↗
Figure 4
Figure 4. Pipeline for Constructing the Video-CoT Dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 2
Figure 2. Figure 2: Additionally, the dataset demonstrates a balanced distribu [PITH_FULL_IMAGE:figures/full_fig_p004_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 18 canonical work pages

  1. [1]

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

    arXiv 2025

  2. [2]

    Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. 2021. Frozen in time: A joint video and image encoder for end-to-end retrieval. InProceedings of the IEEE/CVF international conference on computer vision. 1728–1738

    2021

  3. [3]

    David Chen and William B Dolan. 2011. Collecting highly parallel data for paraphrase evaluation. InProceedings of the 49th annual meeting of the association for computational linguistics: human language technologies. 190–200

    2011

  4. [4]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168(2021)

    arXiv 2021

  5. [5]

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

    arXiv 2025

  6. [6]

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: Program-aided language models. In International Conference on Machine Learning. 10764–10799

    work page 2023

  7. [7]

    something something

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. 2017. The" something something" video database for learning and evaluating visual common sense. InProceedings of the IEEE international conference on computer vision. 5842–5850

    2017

  8. [8]

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

    arXiv 2025

Show all 58 references
  1. [9]

    Xiaoshuai Hao and Wanqian Zhang. 2023. Uncertainty-aware alignment network for cross-domain video-text retrieval.Advances in Neural Information Processing Systems36 (2023), 38284–38296

    2023

  2. [10]

    Xiaoshuai Hao, Wanqian Zhang, Dayan Wu, Fei Zhu, and Bo Li. 2022. Listen and look: Multi-modal aggregation and co-attention network for video-audio retrieval. InIEEE International Conference on Multimedia and Expo. 1–6

    2022

  3. [11]

    Xiaoshuai Hao, Wanqian Zhang, Dayan Wu, Fei Zhu, and Bo Li. 2023. Dual alignment unsupervised domain adaptation for video-text retrieval. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 18962– 18972

    2023

  4. [12]

    Xiaoshuai Hao, Yucan Zhou, Dayan Wu, Wanqian Zhang, Bo Li, and Weiping Wang. 2021. Multi-Feature Graph Attention Network for Cross-Modal Video- Text Retrieval. InProceedings of the 2021 International Conference on Multimedia Retrieval. 135–143

    2021

  5. [13]

    Xiaoshuai Hao, Yucan Zhou, Dayan Wu, Wanqian Zhang, Bo Li, Weiping Wang, and Dan Meng. 2021. What matters: Attentive and relational feature aggregation network for video-text retrieval. InIEEE International Conference on Multimedia and Expo. 1–6

    2021

  6. [14]

    Xiaoshuai Hao, Yi Zhu, Srikar Appalaraju, Aston Zhang, Wanqian Zhang, Bo Li, and Mu Li. 2023. Mixgen: A new multi-modal data augmentation. InProceedings of the IEEE/CVF winter conference on applications of computer vision. 379–389

    2023

  7. [15]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes.arXiv preprint ar...

    2023 arXiv

  8. [16]

    Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. 2024. Vtimellm: Empower llm to grasp video moments. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14271–14280

    2024

  9. [17]

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

    2024 arXiv

  10. [18]

    Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Suk- thankar, and Li Fei-Fei. 2014. Large-scale video classification with convolutional neural networks. InProceedings of the IEEE conference on Computer Vision and Pattern Recognition. 1725–1732

    2014

  11. [19]

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. 2017. The kinetics human action video dataset.arXiv preprint arXiv:1705.06950(2017)

    2017 arXiv

  12. [20]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners.Advances in neural information processing systems35 (2022), 22199–22213

    2022

  13. [21]

    Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles

  14. [22]

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

    2024 arXiv

  15. [23]

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

    2025 arXiv

  16. [24]

    Ruilin Luo, Zhuofan Zheng, Yifan Wang, Yiyao Yu, Xinzhe Ni, Zicheng Lin, Jin Zeng, and Yujiu Yang. 2025. URSA: Understanding and Verifying Chain-of- thought Reasoning in Multimodal Mathematics.arXiv preprint arXiv:2501.04686 (2025)

    2025

  17. [25]

    Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Mar- ianna Apidianaki, and Chris Callison-Burch. 2023. Faithful chain-of-thought reasoning. InThe 13th International Joint Conference on Natural Language Pro- cessing and the 3rd Conference of the Asia-Paci...

    2023

  18. [26]

    Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. 2019. The jester dataset: A large-scale video dataset of human gestures. InProceedings of the IEEE/CVF International Conference on Computer vision Workshops

    2019

  19. [27]

    Jinhui Pang, Jinglin He, Noureldin Mohamed Abdelaal Ahmed Mohamed, Changqing Lin, Zhihui Zhang, and Xiaoshuai Hao. 2025. A Hierarchical Re- inforcement Learning Framework for Multi-UAV Combat Using Leader-Follower Strategy.arXiv preprint arXiv:2501.13132(2025)

    2025 arXiv

  20. [28]

    Wanqian Zhang Jinchao Zhang Zexian Yang Xiaoshuai Hao Bo Li Sa Zhu, Huashan Chen. 2025. Uneven Event Modeling for Partially Relevant Video Retrieval. InIEEE International Conference on Multimedia and Expo. 1–6

    2025

  21. [29]

    Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. 2020. Finegym: A hierarchical video dataset for fine-grained action understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2616–2625

    2020

  22. [30]

    Fangxun Shu, Lei Zhang, Hao Jiang, and Cihang Xie. 2023. Audio-visual llm for video understanding.arXiv preprint arXiv:2312.06720(2023)

    2023 arXiv

  23. [31]

    Xiaoshuai Hao Wanqian Zhang Xiaohua Chen Wei Wang Shuai Cheng, Lin Wang

  24. [32]

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. 2012. UCF101: A dataset of 101 human actions classes from videos in the wild.arXiv preprint arXiv:1212.0402(2012)

    2012 arXiv

  25. [33]

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Com- monsenseqa: A question answering challenge targeting commonsense knowledge. arXiv preprint arXiv:1811.00937(2018)

    2018 arXiv

  26. [34]

    Yingbo Tang, Shuaike Zhang, Xiaoshuai Hao, Pengwei Wang, Jianlong Wu, Zhongyuan Wang, and Shanghang Zhang. 2025. Affordgrasp: In-context affor- dance reasoning for open-vocabulary task-oriented grasping in clutter.arXiv preprint arXiv:2503.00778(2025)

    2025 arXiv

  27. [35]

    Zongheng Tang, Yue Liao, Si Liu, Guanbin Li, Xiaojie Jin, Hongxu Jiang, Qian Yu, and Dong Xu. 2021. Human-centric spatio-temporal video grounding with visual transformers.IEEE Transactions on Circuits and Systems for Video Technology32, 12 (2021), 8238–8249

    2021

  28. [36]

    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

  29. [37]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. 2025. Kimi k1. 5: Scaling reinforcement learning with llms.arXiv preprint arXiv:2501.12599(2025)

    2025 arXiv

  30. [38]

    Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. 2023. Chatvideo: A tracklet-centric multimodal and versatile video understanding system.arXiv preprint arXiv:2304.14407(2023)

    2023 arXiv

  31. [39]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171(2022)

    2022 arXiv

  32. [40]

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al . 2025. InternVideo2. 5: Empowering Video MLLMs with Long and Rich Context Modeling.arXiv preprint arXiv:2501.12386(2025)

    2025 arXiv

  33. [41]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems35 (2022), 24824–24837

    2022

  34. [42]

    Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang

  35. [43]

    Yujie Wu, Huaihai Lyu, Yingbo Tang, Lingfeng Zhang, Zhihui Zhang, Wei Zhou, and Siqi Hao. 2025. Evaluating GPT-4o’s Embodied Intelligence: A Comprehensive Empirical Study.TechRxiv preprint techrxiv.174495686.69962588/v1(2025)

    2025

  36. [44]

    Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. Msr-vtt: A large video description dataset for bridging video and language. InProceedings of the IEEE conference on computer vision and pattern recognition. 5288–5296

    2016

  37. [45]

    Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. 2024. Visa: Reasoning video object segmentation via large language models. InEuropean Conference on Computer Vision. Springer, ACM MM, October 27–31, 2025, Dublin, Ire...

    2024

  38. [46]

    En Yu, Kangheng Lin, Liang Zhao, Yana Wei, Zining Zhu, Haoran Wei, Jianjian Sun, Zheng Ge, Xiangyu Zhang, Jingyu Wang, et al. 2025. Unhackable temporal rewarding for scalable video mllms.arXiv preprint arXiv:2502.12081(2025)

    2025 arXiv

  39. [47]

    Lingfeng Zhang, Xiaoshuai Hao, Qinwen Xu, Qiang Zhang, Xinyao Zhang, Peng- wei Wang, Jing Zhang, Zhongyuan Wang, Shanghang Zhang, and Renjing Xu

  40. [48]

    Lingfeng Zhang, Hao Wang, Erjia Xiao, Xinyao Zhang, Qiang Zhang, Zixuan Jiang, and Renjing Xu. 2024. Multi-Floor Zero-Shot Object Navigation Policy. arXiv preprint arXiv:2409.10906(2024)

    2024 arXiv

  41. [49]

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

    2025 arXiv

  42. [50]

    Y Zhang, B Li, H Liu, Y Lee, L Gui, D Fu, J Feng, Z Liu, and C Li. 2024. Llava-next: A strong zero-shot video understanding model. (2024)

    2024

  43. [51]

    Mapnav: A novel memory representation via annotated semantic maps for vlm-based vision-and-language navigation.arXiv preprint arXiv:2502.13451 (2025)

    2025 arXiv

  44. [52]

    Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. 2023. Multimodal chain-of-thought reasoning in language models.arXiv preprint arXiv:2302.00923(2023)

    2023 arXiv

  45. [53]

    Zhu Zhang, Zhou Zhao, Yang Zhao, Qi Wang, Huasheng Liu, and Lianli Gao

  46. [55]

    Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2022. Automatic chain of thought prompting in large language models.arXiv preprint arXiv:2210.03493 (2022)

    2022 arXiv

  47. [2017]

    InProceedings of the IEEE International Conference on Computer Vision

    Dense-captioning events in videos. InProceedings of the IEEE International Conference on Computer Vision. 706–715

  48. [2020]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Where does it exist: Spatio-temporal video grounding for multi-form sentences. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10668–10677

  49. [2024]

    In European Conference on Computer Vision

    Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision. 453–470

  50. [2025]

    InIEEE International Conference on Multimedia and Expo (ICME)

    Multi-Granularity Based Collaborative Learning for Semi-Supervised Hashing. InIEEE International Conference on Multimedia and Expo (ICME). 1–6

Pith tools

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