Pith. sign in

REVIEW 5 major objections 5 minor 3 cited by

Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition 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 Vad-R1 claims that making a multimodal language model reason through a perception-to-cognition chain before answering improves video anomaly detection, and demonstrates this with a new dataset and a self-verifying reinforcement learning…

desk verdict Solid new task and dataset for video anomaly reasoning, but the self-verification reward and evaluation protocol need scrutiny before the reported gains are taken at face value. read the letter →

arxiv 2505.19877 v1 pith:TRXQETWH submitted 2025-05-26 cs.CV

Chao Huang , Benfeng Wang , Jie Wen , Chengliang Liu , Wei Wang , Li Shen , Xiaochun Cao This is my paper
classification cs.CV
keywords videoanomalyreasoningmultimodallargelanguagemodelschain-of-thoughtreinforcementlearningweaklysuperviseddetectiontemporalgroundingverificationrewardGRPO
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

Vad-R1 proposes a new task, Video Anomaly Reasoning (VAR): before answering whether a video is anomalous, a multimodal large language model must reason explicitly, moving from global perception of the scene, to local perception of suspicious clips, to shallow and then deep cognition about why the event is abnormal, what norms it violates, and what may follow. The paper's central claim is that this structured reasoning is not decoration—it improves anomaly detection itself, and an end-to-end 7B model trained on it outperforms open-source and proprietary baselines on both the new Vad-Reasoning test set and the VANE benchmark. To support that claim, the authors build Vad-Reasoning, a dataset of 8,641 videos with chain-of-thought annotations for a subset and weak video-level labels for the rest, and introduce AVA-GRPO, a reinforcement learning algorithm that rewards correct anomaly predictions through a self-verification step that trims the video and re-checks the model's verdict.

What carries the argument

The load-bearing object is the Perception-to-Cognition Chain-of-Thought (P2C-CoT), a four-step template (global perception, local perception, shallow cognition, deep cognition) followed by a structured answer tagging what, when, where, why, which, and how. It carries the argument in two ways: as the annotation schema for the Vad-Reasoning dataset, and as the output format that forces the model to think before answering. Around it, the training machinery is AVA-GRPO, an extension of the GRPO reinforcement learning objective in which an anomaly verification reward is computed by trimming the video according to the model's own prediction and re-asking the model; the reward is positive when removing a predicted abnormal span makes the video look normal, and negative when truncating a 'normal' prediction exposes an anomaly. This reward gives the RL stage a training signal from video-level weak labels alone.

What would settle it

Take videos with two well-separated anomalous events, train Vad-R1 with AVA-GRPO, and check whether removing one predicted span still yields a 'normal' verdict; if the positive reward fires while the remaining anomaly is plainly visible, the reward is measuring something other than correct detection.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the missing ingredient in MLLM-based video anomaly detection is reasoning, not model size or annotation volume. Vad-R1 is built on a 7B open-source video-language base model and trained in two stages: supervised fine-tuning on the P2C-CoT annotated subset, followed by reinforcement learning on weakly labeled videos. The RL stage uses an anomaly verification reward: for a video predicted abnormal, the predicted abnormal segment is removed and the model must then call the trimmed video normal; for a video predicted normal, dropping the beginning or ending segment must not flip the verdict to abnormal. The reported numbers support the claim: Vad-R1 reaches 0.875 accuracy, 0.862 F1, 0.713 mIoU and 0.770 R@0.3 on Vad-Reasoning, and exceeds all compared models on VANE category accuracy, including proprietary reasoning models. The paper also shows that prompting for the structured chain-of-thought helps base models more than random reasoning does, and that SFT alone or RL alone give less than the combination.

Load-bearing premise

The training signal assumes that removing the predicted abnormal segment from an abnormal video turns it into a normal video, which fails for multi-anomaly videos and imprecise time ranges.

Editorial extensions

If this is right

  • Prompting even a frozen general MLLM with the P2C-CoT template improves detection Recall and F1 over direct answering, so the template itself is a reusable inference-time intervention.
  • A 7B open-source model trained with SFT plus AVA-GRPO can surpass proprietary reasoning models on anomaly reasoning and detection benchmarks, suggesting reasoning skill, not parameter count, is the bottleneck.
  • The SFT-then-RL sequence is necessary: RL without SFT yields limited gains, so the CoT annotations create the capability that RL then amplifies.
  • The anomaly verification reward allows RL to exploit large pools of weakly labeled videos, reducing the need for detailed anomaly annotations at scale.

Reading between the lines

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

  • The trim-and-reask reward is a general recipe for weakly supervised temporal reasoning: any task where removing the predicted evidence should change the label, such as action localization or misinformation clip grounding, could reuse the same loop.
  • If the causal claim is right, token-overlap metrics like BLEU and ROUGE likely understate the real gain; a human study or adversarial questions about the 'why' and 'how' steps would test whether reasoning quality genuinely improves rather than merely detection scores.
  • A stress test for the reward is multi-anomaly videos: when one anomaly remains after trimming, the reward still fires, so the method may need a non-binary verification signal to scale to densely anomalous videos.
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 Anomaly Reasoning (VAR), a task that requires multimodal LLMs to produce structured chain-of-thought reasoning before answering anomaly-related questions. The authors design a Perception-to-Cognition Chain-of-Thought (P2C-CoT) with global/local perception and shallow/deep cognition stages, construct a new dataset Vad-Reasoning containing a CoT-annotated SFT subset and a larger weakly labeled RL subset, and propose AVA-GRPO, an extension of GRPO that adds an anomaly verification reward computed by re-feeding temporally trimmed videos to the policy. They train Qwen2.5-VL-7B in two stages and report superior results on the Vad-Reasoning test set and the VANE benchmark, including detection metrics (accuracy, F1, mIoU, R@K) and text-overlap metrics for reasoning quality.

Significance. The paper proposes a useful structured reasoning template for video anomalies, a sizable hybrid dataset, and a weak-label RL objective with a self-verification mechanism; if validated, these could be solid contributions to video anomaly detection and understanding. The detection evaluation is partially grounded in ground-truth labels and the VANE protocol excludes UCF-Crime to reduce training leakage, which is a strength. However, the reasoning-quality evaluation and the AVA-GRPO reward mechanism currently have circularity and robustness gaps, and the paper's 'reasoning' claims rely heavily on metrics that may measure imitation of the annotator model rather than independent reasoning quality. The dataset statistics and some comparison tables also contain inconsistencies that need correction before the central claims can be fully assessed.

major comments (5)
  1. [§3.2 and Appendix B.2] The dataset statistics are internally inconsistent: the main text says Vad-Reasoning-SFT contains 1,755 videos and the training set contains 8,203 videos (with 438 test videos), while Appendix B.2 reports 2,193 videos in Vad-Reasoning-SFT and a total of 8,641 videos, and Figure 2(c) labels the SFT subset as 1,755. Please correct the numbers and, more importantly, describe how the 438-video test set was constructed (random split, source-stratified, or otherwise) and whether any near-duplicate or source-overlap analysis was performed between the test set and the training subsets. Without this information, the main comparison in Table 2 is not fully reproducible.
  2. [§3.3, Algorithm 1] The anomaly verification reward is computed by sampling a new completion from the same policy πθ on the trimmed video and comparing its predicted category with the original prediction. This is self-referential: the policy can increase the reward by learning to label trimmed videos as normal regardless of whether the discarded segment was truly anomalous. The underlying assumption that removing the predicted abnormal segment leaves a normal video fails for videos with multiple anomalies or when the predicted temporal span is imprecise, and the paper does not analyze these failure modes. Moreover, Table 10 shows that adding the anomaly reward alone to GRPO yields ROUGE-L 0.496 versus 0.502 for 16 frames, so the claimed benefit of the verification mechanism is not demonstrated independently of the length reward. Please provide a frozen-verifier baseline or label-based verification, and analyze the multi-anomaly and imprecise-span cases.
  3. [§4.2 and Appendix D.1] The reasoning-quality metrics (BLEU, METEOR, ROUGE, and the LLM-guided scores) compare generated text against Qwen-generated references, and the SFT annotations were also generated by Qwen models. The reasoning gains may therefore reflect imitation of the annotation distribution rather than improved reasoning ability. The LLM-guided evaluation in Appendix D.1 does not specify which judge model was used, the prompt, the sample size, or any human correlation. Please add a human evaluation or an independent reasoning-quality protocol, and report the judge details.
  4. [§3.3, Eq. (3) and Algorithm 2] The paper states that because each completion is used for only one update, the clipped GRPO objective in Eq. (2) can be simplified to the unclipped form in Eq. (3), with πθno grad 'equivalent to πθ.' This is not justified: dropping the min and clip operations changes the optimization objective, and performing one update does not make the importance ratio equal to 1 unless the parameters are held fixed during the update. Please clarify the derivation or use the standard clipped objective.
  5. [§4.3, Table 4] The claim that SFT equips the model with basic reasoning capability is only partially supported. Compared with the base Qwen2.5-VL, +SFT improves BLEU-2 (0.113 to 0.219) and R@0.3 (0.610 to 0.677), but decreases ROUGE-1 (0.505 to 0.456) and ROUGE-L (0.477 to 0.429). Please explain this inconsistency or report additional metrics that align with the claimed benefit of the structured CoT annotations.
minor comments (5)
  1. [Table 2] The Open-R1-Video and VideoChat-R1 rows have identical accuracy, F1, mIoU, R@0.3, and R@0.5 values; this looks like a copy-paste error and should be corrected.
  2. [Appendix B.2] The text contains a typo: 'datset' should be 'dataset'.
  3. [Eq. (3)] The notation πθno grad is used without a formal definition; please define it or remove it.
  4. [Figure 10] The training-curve panels lack axis labels and units; please add them for interpretability.
  5. [Appendix E] The limitation section mentions only inference speed; given the concerns about the anomaly verification reward and the reasoning-evaluation protocol, the limitations discussion should also address those issues.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity: the anomaly-verification reward uses the trained policy as its own oracle and the length reward is fitted to the SFT-annotation length distribution, partially inflating reasoning-text metrics; the detection claims remain independently grounded.

  1. self definitional [Section 3.3, Algorithm 1 and 'Anomaly Verification Reward' paragraph]
    "Generate a new completion ˜o ∼ πθ(· |p, ˜v) ... if p == Abnormal and ˜p == Normal then Assign positive reward ri ← 0.5 ... If the trimmed video is subsequently predicted as normal, it suggests that the discarded segment is indeed abnormal and the model’s initial prediction was correct."

    The 'verification' of the original prediction is a second sample from the same policy πθ being optimized (Algorithm 1 lines 10-12). By construction, the reward is R_ano = 0.5 if πθ(v)=Abnormal and πθ(trim(v))=Normal, a self-consistency predicate, not a ground-truth check that the discarded segment was anomalous. The policy can increase reward by learning the mapping untrimmed->Abnormal and trimmed->Normal regardless of video content, so the reward does not externally verify anomaly correctness. No frozen verifier or ground-truth temporal labels are used, and Table 10 shows GRPO+ano_reward alone underperforms GRPO on ROUGE-L (0.496 vs 0.502), so the claimed contribution of the self-verification mechanism to the headline gains is not independently established.

  2. fitted input called prediction [Appendix C.2, 'Training Process of AVA-GRPO' and Table 10]
    "We first separately calculate the length of the reasoning text for abnormal videos and normal videos in Vad-Reasoning-SFT. During RL training, if the length of output satisfies the corresponding range, a length reward will be assigned. For normal videos, the length range of reasoning process is set as [140, 261], while it is set as [233, 456] for abnormal videos."

    The length-reward range is fitted to the SFT annotation distribution, and the test references used for BLEU/ROUGE were produced by the same Qwen-VL-Max/Qwen-Max annotation pipeline. Rewarding outputs whose length lies inside the fitted SFT range mechanically shifts generated text toward the reference length distribution, so part of the reported reasoning-quality gain (e.g., ROUGE-L rising from 0.502 to 0.530 when the length reward is added in Table 10) is a fitted-input effect rather than an independent reasoning improvement. This does not affect the detection metrics, which use ground-truth labels, but it partially inflates the 'anomaly reasoning' comparison.

full rationale

The central anomaly-detection claims are independently grounded: accuracy, F1, mIoU and R@K are computed against ground-truth video-level and temporal labels, and VANE is an external benchmark. The proposed P2C-CoT and dataset are also legitimate contributions whose test set is separate from training. However, two partial circularities weaken the reasoning-quality story. First, Algorithm 1's anomaly-verification reward draws the verification prediction from the same policy being trained, so the 'correctness' signal is self-consistency by construction rather than external verification; the inference that a trimmed video predicted normal implies the discarded segment was anomalous is an untested assumption that can be gamed, and the ablation evidence for the anomaly reward alone is not positive. Second, the length-reward range is fitted to the SFT annotation length distribution, while the BLEU/ROUGE references come from the same annotation pipeline; adding the length reward inflates ROUGE-L in a way that is partly an artifact of matching the reference length distribution. These issues are real but partial: they do not reduce the detection numbers, the VANE results, or the value of the released dataset, and no load-bearing self-citation chain is present. Hence a moderate score of 4.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method rests on proprietary-model-generated annotations, a heuristic verification reward, and standard RL machinery. The length-reward ranges are fitted to the SFT annotation distribution.

free parameters (6)
  • Anomaly verification reward (abnormal) = 0.5
    Hand-picked reward value for correct abnormal predictions in AVA-GRPO; directly shapes the RL objective.
  • Anomaly verification reward (normal) = -0.2
    Hand-picked penalty for normal predictions that flip to abnormal after trimming; designed to discourage temporal hacking.
  • Length reward = 0.2
    Hand-picked reward for output length within the SFT-derived ranges.
  • Length ranges for reasoning text = [140,261] normal, [233,456] abnormal
    Computed from Vad-Reasoning-SFT annotation lengths; these ranges are fitted to the training annotation distribution and used as a reward criterion.
  • KL penalty beta = 0.04
    Hand-picked regularization strength in Eq. 3.
  • Number of completions G = 4
    Group size for GRPO advantage normalization; chosen by hand.
assumptions (4)
  • domain assumption Qwen-VL-Max and Qwen-Max generate accurate, unbiased dense frame descriptions and CoT annotations.
    The Vad-Reasoning-SFT dataset is built entirely from proprietary-model outputs without human verification; the quality of these annotations is the training signal for reasoning.
  • ad hoc to paper Removing the predicted abnormal segment from an abnormal video yields a normal video if and only if the prediction is correct.
    Algorithm 1 assigns positive reward when the trimmed video is re-classified as normal; this assumes single-anomaly videos and accurate temporal localization.
  • ad hoc to paper A normal video that becomes abnormal after discarding its beginning or ending segment is evidence of temporal hacking.
    Used to assign the -0.2 penalty; this heuristic assumes the dropped segment contained the anomaly that was missed.
  • standard math The GRPO objective and KL penalty from DeepSeekMath apply to this MLLM setting.
    Eq. 2 and Eq. 3 adopt the standard GRPO loss; no new derivation is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought." pith.science (2026). https://pith.science/paper/TRXQETWH

@misc{pith2026250519877,
  author       = {Pith},
  title        = {Pith review of: Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TRXQETWH}},
  note         = {Machine review of arXiv:2505.19877}
}
read the original abstract

Recent advancements in reasoning capability of Multimodal Large Language Models (MLLMs) demonstrate its effectiveness in tackling complex visual tasks. However, existing MLLM-based Video Anomaly Detection (VAD) methods remain limited to shallow anomaly descriptions without deep reasoning. In this paper, we propose a new task named Video Anomaly Reasoning (VAR), which aims to enable deep analysis and understanding of anomalies in the video by requiring MLLMs to think explicitly before answering. To this end, we propose Vad-R1, an end-to-end MLLM-based framework for VAR. Specifically, we design a Perception-to-Cognition Chain-of-Thought (P2C-CoT) that simulates the human process of recognizing anomalies, guiding the MLLM to reason anomaly step-by-step. Based on the structured P2C-CoT, we construct Vad-Reasoning, a dedicated dataset for VAR. Furthermore, we propose an improved reinforcement learning algorithm AVA-GRPO, which explicitly incentivizes the anomaly reasoning capability of MLLMs through a self-verification mechanism with limited annotations. Experimental results demonstrate that Vad-R1 achieves superior performance, outperforming both open-source and proprietary models on VAD and VAR tasks. Codes and datasets will be released at https://github.com/wbfwonderful/Vad-R1.

Figures

Figures reproduced from arXiv: 2505.19877 by the authors.

Figure 1
Figure 1. Overview of Vad-R1. Vad-R1 is an end-to-end framework for video anomaly reasoning. A structured Perception-to-Cognition Chain-of-Thought is proposed to guide Vad-R1 in step-by-step reasoning. Based on the structured CoT, a new dataset for video anomaly reasoning is constructed, including fine-grained anomaly categories. A two-stage training pipeline is adopted to progressively enhance the reasoning capability of Vad… view at source ↗
Figure 2
Figure 2. Overview of the proposed Perception-to-Cognition CoT and Vad-Reasoning dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the two-stage training pipeline for Vad-R1. Stage 1 enables the model to acquire basic reasoning capability with CoT annotated video. Stage 2 further enhances the model’s reasoning capability through reinforcement learning. influence (How). Notably, for normal videos, the corresponding P2C-CoT is simplified into two steps. Please refer to Appendix B for more details. 3.2 Dataset: Vad-Reasoning Video … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Qualitative performance on VANE benchmark. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Illustration of multi-stage annotation process of Vad-Reasoning-SFT dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Statistical analyses of the proposed Vad-Reasoning dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: An abnormal example of Vad-Reasoning. Given this setup, the anomaly could involve the cat reacting suddenly, such as scratching or biting the boy when he leans closer. Such a reaction aligns with the \"Animals Hurt Human\" category, as the cat may feel threatened by th…
Figure 9
Figure 9. Figure 9: Prompt template for performing video anomaly reasoning. [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]
Figure 10
Figure 10. Figure 10: RL training curves of Vad-R1. unstable improvement. These results demonstrate that the combination of length and anomaly rewards is essential for improving the overall reasoning and detection performance. D.4 Training Curves [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: Qualitative result for an abnormal video. [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: Qualitative result for a normal video. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. MAVEN: A Multi-stage Agentic Annotation Pipeline for Video Reasoning Tasks

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    MAVEN pipeline generates multi-scale spatio-temporal event descriptions from videos using agentic adaptation and refinement, then produces training data that lets a fine-tuned 8B model outperform Gemini baselines on p...

  2. O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning

    cs.CV 2026-07 conditional novelty 6.0 of 10

    An object-centric, training-free agentic pipeline that tracks object state changes and reasons over them with a vision-language model achieves strong video-level AUROC on Phys-AD, LiquidAD, and IPAD, while producing i...

  3. DAMS:Dual-Branch Adaptive Multiscale Spatiotemporal Framework for Video Anomaly Detection

    cs.CV 2025-07 conditional novelty 4.0 of 10

    DAMS, a dual-branch architecture fusing adaptive temporal pyramids, CBAM attention, and CLIP pseudo-labels, reports 94.67 AUC on UCF-Crime and 84.00 AP on XD-Violence for weakly supervised video anomaly detection.

Reference graph

Works this paper leans on

94 extracted references · 41 canonical work pages · cited by 3 Pith papers

  1. [1]

    Ubnormal: New benchmark for supervised open-set video anomaly detection

    Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20143–20153, 2022. 3, 5, 18

    2022

  2. [2]

    Claude 3.5 haiku, 2024

    Anthropic. Claude 3.5 haiku, 2024. URL https://www.anthropic.com/claude/haiku. 8, 23

    2024

  3. [3]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments

    Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65–72,

  4. [4]

    Video generation models as world simulators

    Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1:8, 2024. 20

    2024

  5. [5]

    A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation

    Congqi Cao, Yue Lu, Peng Wang, and Yanning Zhang. A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20392–20401, June

  6. [6]

    Videollm-online: Online video large language model for streaming video

    Joya Chen, Zhaoyang Lv, Shiwei Wu, Kevin Qinghong Lin, Chenan Song, Difei Gao, Jia-Wei Liu, Ziteng Gao, Dongxing Mao, and Mike Zheng Shou. Videollm-online: Online video large language model for streaming video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18407–18418, 2024. 3

    2024

  7. [7]

    Prompt-enhanced multiple instance learning for weakly supervised video anomaly detection

    Junxi Chen, Liang Li, Li Su, Zheng-Jun Zha, and Qingming Huang. Prompt-enhanced multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18319–18329, 2024. 1, 3

    2024

  8. [8]

    Tevad: Improved video anomaly detection with captions

    Weiling Chen, Keng Teck Ma, Zi Jian Yew, Minhoe Hur, and David Aik-Aun Khoo. Tevad: Improved video anomaly detection with captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549–5559, 2023. 3

    2023

Show all 94 references
  1. [9]

    DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning

    DeepSeek-AI. DeepSeek-R1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3, 8

    2025 arXiv

  2. [10]

    Streaming video question-answering with in-context video kv-cache retrieval

    Shangzhe Di, Zhelun Yu, Guanghao Zhang, Haoyuan Li, Tao Zhong, Hao Cheng, Bolin Li, Wanggui He, Fangxun Shu, and Hao Jiang. Streaming video question-answering with in-context video kv-cache retrieval. arXiv preprint arXiv:2503.00540, 2025. 3

    2025 arXiv

  3. [11]

    Slowfastvad: Video anomaly detection via integrating simple detector and rag-enhanced vision-language model

    Zongcan Ding, Haodong Zhang, Peng Wu, Guansong Pang, Zhiwei Yang, Peng Wang, and Yanning Zhang. Slowfastvad: Video anomaly detection via integrating simple detector and rag-enhanced vision-language model. arXiv preprint arXiv:2504.10320, 2025. 2, 3

    2025 arXiv

  4. [12]

    Exploring what why and how: A multifaceted benchmark for causation understanding of video anomaly

    Hang Du, Guoshun Nan, Jiawen Qian, Wangchenhui Wu, Wendi Deng, Hanqing Mu, Zhenyan Chen, Pengxuan Mao, Xiaofeng Tao, and Jun Liu. Exploring what why and how: A multifaceted benchmark for causation understanding of video anomaly. arXiv preprint arXiv:2412.07183,

    arXiv

  5. [13]

    Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly

    Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on...

    2024

  6. [14]

    Video-r1: Reinforcing video reasoning in mllms

    Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms. arXiv preprint arXiv:2503.21776, 2025. 3, 7, 8, 23 10

    2025 arXiv

  7. [15]

    Vane-bench: Video anomaly evaluation benchmark for conversational lmms

    Hanan Gani, Rohit Bharadwaj, Muzammal Naseer, Fahad Shahbaz Khan, and Salman Khan. Vane-bench: Video anomaly evaluation benchmark for conversational lmms. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 3123–3140, 2025. 7, 20

    2025

  8. [16]

    Open-sora: Democratizing efficient video production for all

    HPCAI Tech. Open-sora: Democratizing efficient video production for all. https://github. com/hpcaitech/Open-Sora, 2024. 20

    2024

  9. [17]

    Abnormal event detection using deep contrastive learning for intelligent video surveillance system

    Chao Huang, Zhihao Wu, Jie Wen, Yong Xu, Qiuping Jiang, and Yaowei Wang. Abnormal event detection using deep contrastive learning for intelligent video surveillance system. IEEE Transactions on Industrial Informatics, 18(8):5171–5179, 2021. 1, 3

    2021

  10. [18]

    Weakly supervised video anomaly detection via self-guided temporal discriminative transformer

    Chao Huang, Chengliang Liu, Jie Wen, Lian Wu, Yong Xu, Qiuping Jiang, and Yaowei Wang. Weakly supervised video anomaly detection via self-guided temporal discriminative transformer. IEEE Transactions on Cybernetics, 54(5):3197–3210, 2022. 1, 3

    2022

  11. [19]

    Self-supervised attentive generative adversarial networks for video anomaly detection

    Chao Huang, Jie Wen, Yong Xu, Qiuping Jiang, Jian Yang, Yaowei Wang, and David Zhang. Self-supervised attentive generative adversarial networks for video anomaly detection. IEEE transactions on neural networks and learning systems, 34(11):9389–9403, 2022. 1, 3

    2022

  12. [20]

    Long short-term dynamic prototype alignment learning for video anomaly detection

    Chao Huang, Jie Wen, Chengliang Liu, and Yabo Liu. Long short-term dynamic prototype alignment learning for video anomaly detection. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 866–874, 2024. 1, 3

    2024

  13. [21]

    Multi- modal evidential learning for open-world weakly-supervised video anomaly detection

    Chao Huang, Weiliang Huang, Qiuping Jiang, Wei Wang, Jie Wen, and Bob Zhang. Multi- modal evidential learning for open-world weakly-supervised video anomaly detection. IEEE Transactions on Multimedia, 2025. 1, 3

    2025

  14. [22]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749, 2025. 3

    2025 arXiv

  15. [23]

    Chat-univi: Unified visual representation empowers large language models with image and video understanding

    Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700–1...

    2024

  16. [24]

    Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection

    Hyekang Kevin Joo, Khoa V o, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230–3234. IEEE, 2023. 1, 3

    2023

  17. [25]

    Videochat: Chat-centric video understanding

    KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355,

    arXiv

  18. [26]

    Anomaly detection and localization in crowded scenes

    Weixin Li, Vijay Mahadevan, and Nuno Vasconcelos. Anomaly detection and localization in crowded scenes. IEEE transactions on pattern analysis and machine intelligence, 36(1):18–32,

  19. [27]

    Videochat-flash: Hierarchical compression for long-context video modeling

    Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling. arXiv preprint arXiv:2501.00574, 2024. 8, 23

    2024 arXiv

  20. [28]

    Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment 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. 3, 7, 8, 23

    2025 arXiv

  21. [29]

    Llama-vid: An image is worth 2 tokens in large language models

    Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 3, 9

    2024

  22. [30]

    Video-llava: Learning united visual representation by alignment before projection

    Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122,

    arXiv

  23. [31]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. InText summarization branches out, pages 74–81, 2004. 7

    2004

  24. [32]

    Future frame prediction for anomaly detection–a new baseline

    Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection–a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536–6545, 2018. 1, 3, 5, 18

    2018

  25. [33]

    Videomind: A chain-of-lora agent for long video reasoning

    Ye Liu, Kevin Qinghong Lin, Chang Wen Chen, and Mike Zheng Shou. Videomind: A chain-of-lora agent for long video reasoning. arXiv preprint arXiv:2503.13444, 2025. 3

    2025

  26. [34]

    A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction

    Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588–13...

    2021

  27. [35]

    Abnormal event detection at 150 fps in matlab

    Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. In Proceedings of the IEEE international conference on computer vision, pages 2720–2727, 2013. 18, 20

    2013

  28. [36]

    Video anomaly detection and explanation via large language models

    Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 1, 3

    2024 arXiv

  29. [37]

    Localizing anomalies from weakly-labeled videos

    Hui Lv, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Localizing anomalies from weakly-labeled videos. IEEE transactions on image processing, 30:4505–4515, 2021. 1, 3, 5, 18

    2021

  30. [38]

    Sherlock: Towards multi-scene video abnormal event extraction and localization via a global-local spatial-sensitive llm

    Junxiao Ma, Jingjing Wang, Jiamin Luo, Peiying Yu, and Guodong Zhou. Sherlock: Towards multi-scene video abnormal event extraction and localization via a global-local spatial-sensitive llm. In Proceedings of the ACM on Web Conference 2025, pages 4004–4013, 2025. 2, 3

    2025

  31. [39]

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

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3, 7, 9

    2023 arXiv

  32. [40]

    Gpt-4o system card

    OpenAI. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7, 8, 23

    2024 arXiv

  33. [41]

    Openai o1 system card

    OpenAI. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 1

    2024 arXiv

  34. [42]

    Openai o3 and o4-mini system card, 2025

    OpenAI. Openai o3 and o4-mini system card, 2025. URL https://openai.com/index/ o3-o4-mini-system-card/ . 8

    2025

  35. [43]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 7

    2002

  36. [44]

    Timechat: A time-sensitive multimodal large language model for long video understanding

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313–14323, 2024. 9

    2024

  37. [45]

    Self-distilled masked auto-encoders are efficient video anomaly detectors

    Nicolae-C Ristea, Florinel-Alin Croitoru, Radu Tudor Ionescu, Marius Popescu, Fahad Shahbaz Khan, Mubarak Shah, et al. Self-distilled masked auto-encoders are efficient video anomaly detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2024

  38. [46]

    Gen-2: The next step forward for generative ai

    Runway Research. Gen-2: The next step forward for generative ai. https://research. runwayml.com/gen2, 2024. 20

    2024

  39. [47]

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

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 3, 6

    2024 arXiv

  40. [48]

    Moviechat: From dense token to sparse memory for long video understanding

    Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patt...

    2024

  41. [49]

    Real-world anomaly detection in surveillance videos

    Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479–6488, 2018. 1, 3, 5, 17, 18, 20

    2018

  42. [50]

    Hawk: Learning to understand open-world video anomalies

    Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and Yingcong Chen. Hawk: Learning to understand open-world video anomalies. Advances in Neural Information Processing Systems, 37:139751–139785, 2024. 2, 3, 5, 7, 8, 9, 18, 21, 23

    2024

  43. [51]

    Gemini 2.5 flash preview model card, 2025

    Gemini Team. Gemini 2.5 flash preview model card, 2025. URL https://storage. googleapis.com/model-cards/documents/gemini-2.5-flash-preview.pdf . 7, 8

    2025

  44. [52]

    Gemini 2.5 pro preview model card, 2025

    Gemini Team. Gemini 2.5 pro preview model card, 2025. URL https://storage. googleapis.com/model-cards/documents/gemini-2.5-pro-preview.pdf . 7, 8

    2025

  45. [53]

    Kimi k1.5: Scaling reinforcement learning with llms

    Kimi Team. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025. 1

    2025 arXiv

  46. [54]

    QwQ: Reflect deeply on the boundaries of the unknown, 2024

    Qwen Team. QwQ: Reflect deeply on the boundaries of the unknown, 2024. URL https: //qwenlm.github.io/blog/qwq-32b-preview/ . 1

    2024

  47. [55]

    Qwen2.5 technical report

    Qwen Team. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5

    2024 arXiv

  48. [56]

    QVQ-Max: Think with evidence, 2025

    Qwen Team. QVQ-Max: Think with evidence, 2025. URL https://qwenlm.github.io/ blog/qvq-max-preview/. 1, 7, 8, 23

    2025

  49. [57]

    Qwen2.5-vl technical report

    Qwen Team. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 5, 6, 7, 8, 23

    2025 arXiv

  50. [58]

    Qwen3: Think deeper, act faster, 2025

    Qwen Team. Qwen3: Think deeper, act faster, 2025. URL https://qwenlm.github.io/ blog/qwen3/. 7

    2025

  51. [59]

    Llamav-o1: Rethinking step-by-step visual reasoning in llms

    Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamav-o1: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 3

    2025 arXiv

  52. [60]

    Federated weakly supervised video anomaly detection with multimodal prompt

    Benfeng Wang, Chao Huang, Jie Wen, Wei Wang, Yabo Liu, and Yong Xu. Federated weakly supervised video anomaly detection with multimodal prompt. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 21017–21025, 2025. 1, 3

    2025

  53. [61]

    Modelscope text-to-video technical report, 2023

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report, 2023. 20

    2023

  54. [62]

    Videolcm: Video latent consistency model, 2023

    Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model, 2023. 20

    2023

  55. [63]

    Open-r1-video, 2025

    Xiaodong Wang and Peixi Peng. Open-r1-video, 2025. URL https://github.com/ Wang-Xiaodong1899/Open-R1-Video . 8, 23

    2025

  56. [64]

    Timezero: Temporal video grounding with reasoning-guided lvlm

    Ye Wang, Boshen Xu, Zihao Yue, Zihan Xiao, Ziheng Wang, Liang Zhang, Dingyi Yang, Wenxuan Wang, and Qin Jin. Timezero: Temporal video grounding with reasoning-guided lvlm. arXiv preprint arXiv:2503.13377, 2025. 3, 7

    2025 arXiv

  57. [65]

    Internvideo2

    Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 8, 23

    2025 arXiv

  58. [66]

    Not only look, but also listen: Learning multimodal violence detection under weak supervision

    Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedi...

    2020

  59. [67]

    Open-vocabulary video anomaly detection

    Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297–18307, 2024. 1, 3 13

    2024

  60. [68]

    Vadclip: Adapting vision-language models for weakly supervised video anomaly detection

    Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6074–6082, 2...

    2024

  61. [69]

    Streaming video understanding and multi-round interaction with memory-enhanced knowl- edge

    Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowl- edge. arXiv preprint arXiv:2501.13468, 2025. 3

    2025 arXiv

  62. [70]

    Llava-o1: Let vision language models reason step-by-step

    Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-o1: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 3

    2024 arXiv

  63. [71]

    Pllava: Parameter-free llava extension from images to videos for video dense captioning

    Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3

    2024 arXiv

  64. [72]

    Feature prediction diffusion model for video anomaly detection

    Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature prediction diffusion model for video anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5527–5537, 2023. 3

    2023

  65. [73]

    Follow the rules: reasoning for video anomaly detection with large language models

    Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. In European Conference on Computer Vision, pages 304–322. Springer, 2024. 2, 3

    2024

  66. [74]

    Svbench: A benchmark with temporal multi-turn dialogues for streaming video understanding

    Zhenyu Yang, Yuhang Hu, Zemin Du, Dizhan Xue, Shengsheng Qian, Jiahong Wu, Fan Yang, Weiming Dong, and Changsheng Xu. Svbench: A benchmark with temporal multi-turn dialogues for streaming video understanding. arXiv preprint arXiv:2502.10810, 2025. 3

    2025

  67. [75]

    Dota: Unsupervised detection of traffic anomaly in driving videos.IEEE transactions on pattern analysis and machine intelligence, 45(1):444–459, 2022

    Yu Yao, Xizi Wang, Mingze Xu, Zelin Pu, Yuchen Wang, Ella Atkins, and David J Crandall. Dota: Unsupervised detection of traffic anomaly in driving videos.IEEE transactions on pattern analysis and machine intelligence, 45(1):444–459, 2022. 1, 3

    2022

  68. [76]

    Vera: Explainable video anomaly detection via verbalized learning of vision-language models

    Muchao Ye, Weiyang Liu, and Pan He. Vera: Explainable video anomaly detection via verbalized learning of vision-language models. arXiv preprint arXiv:2412.01095, 2024. 1, 3

    2024 arXiv

  69. [77]

    Unhackable temporal rewarding for scalable video mllms

    En Yu, Kangheng Lin, Liang Zhao, Yana Wei, Zining Zhu, Haoran Wei, Jianjian Sun, Zheng Ge, Xiangyu Zhang, Jingyu Wang, et al. Unhackable temporal rewarding for scalable video mllms. arXiv preprint arXiv:2502.12081, 2025. 5, 6

    2025 arXiv

  70. [78]

    Towards surveillance video-and-language understanding: New dataset baselines and challenges

    Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset baselines and challenges. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 22052...

    2024

  71. [79]

    Generative cooperative learning for unsupervised video anomaly detection

    M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744–14754,

  72. [80]

    Har- nessing large language models for training-free video anomaly detection

    Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Har- nessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527–18536,

  73. [81]

    Vision-r1: Evolving human-free alignment in large vision-language models via vision- guided reinforcement learning

    Yufei Zhan, Yousong Zhu, Shurong Zheng, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Vision-r1: Evolving human-free alignment in large vision-language models via vision- guided reinforcement learning. arXiv preprint arXiv:2503.18013, 2025. 3

    2025 arXiv

  74. [82]

    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. 8, 23

    2025 arXiv

  75. [83]

    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. arXiv preprint arXiv:2306.02858, 2023. 3, 7, 9 14

    2023 arXiv

  76. [84]

    Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm

    Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 1, 3, 7, 8, 9, 18, 23

    2024 arXiv

  77. [85]

    Holmes-vau: Towards long-term video anomaly understanding at any granularity

    Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Xiaonan Huang, Changxin Gao, Shanjun Zhang, Li Yu, and Nong Sang. Holmes-vau: Towards long-term video anomaly understanding at any granularity. arXiv preprint arXiv:2412.06171, 2024. 2, 3, 7, 8, 9, 18, 23

    2024 arXiv

  78. [86]

    Long context transfer from language to vision

    Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 3

    2024 arXiv

  79. [87]

    Llama-adapter: Efficient fine-tuning of language models with zero-init attention

    Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 3, 7

    2023 arXiv

  80. [88]

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

    Xingjian Zhang, Siwei Wen, Wenjun Wu, and Lei Huang. Tinyllava-video-r1: Towards smaller lmms for video reasoning. arXiv preprint arXiv:2504.09641, 2025. 3, 7

    2025 arXiv

  81. [89]

    Video instruction tuning with synthetic data

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 8, 23

    2024 arXiv

  82. [90]

    Graph convo- lutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection

    Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convo- lutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237–...

    2019

  83. [91]

    Dual memory units with uncertainty regulation for weakly supervised video anomaly detection

    Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 3769–3777, 2023. 3

    2023

  84. [92]

    Animals Hurt Human\

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Yuchen Duan, Hao Tian, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479, 2025. 8, 23 15 A Summary...

    2025 arXiv

  85. [93]

    A structured reasoning process enclosed in <think></think> tags

  86. [94]

    Abnormal\

    A final explanation enclosed in <answer></answer> tags For abnormal videos, the reasoning should be based on a structured 4-step process: <think> must include the following four steps enclosed in corresponding tags: <step1>: Scene Description — Provide an objective overview of...

Pith tools

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