Pith. sign in

REVIEW 4 major objections 7 minor 4 cited by

VAU-R1: Advancing Video Anomaly Understanding via Reinforcement Fine-Tuning

T0 review · 4 major / 7 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 Reinforcement fine-tuning with GRPO makes small video-language models reason about anomalies better than supervised fine-tuning, and VAU-Bench is built to measure it.

desk verdict The benchmark is the real contribution; the paper's central claim of consistent reasoning gains is contradicted by its own tables and rests on a circular LLM-as-judge evaluation. read the letter →

arxiv 2505.23504 v1 pith:ELEWJVUG submitted 2025-05-29 cs.CV

Liyun Zhu , Qixiang Chen , Xi Shen , Xiaodong Cun This is my paper
classification cs.CV
keywords videoanomalyunderstandingreinforcementfine-tuningGRPOchain-of-thoughtbenchmarktemporalgroundingmultimodallargelanguagemodelreasoningsurveillance
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 sets out to show that reinforcement fine-tuning (RFT) can turn a multimodal video-language model into an explainable video-anomaly reasoner, and that a purpose-built chain-of-thought benchmark can measure that ability. It introduces VAU-R1, which applies Group Relative Policy Optimization (GRPO) with format, accuracy, and temporal-IoU rewards to a small open MLLM, and VAU-Bench, a 4,602-video benchmark with QA pairs, temporal intervals, captions, and step-by-step rationales. On its own evaluations, the RFT model consistently beats its supervised-fine-tuned twin on question accuracy, temporal grounding, and five judged dimensions of reasoning coherence. If the result holds, data-efficient reinforcement fine-tuning would be a practical recipe for interpretable anomaly understanding in surveillance and safety applications.

What carries the argument

The load-bearing mechanism is Group Relative Policy Optimization (GRPO), a reinforcement-learning update that samples several candidate outputs per prompt, normalizes their rewards against the group mean and standard deviation, and optimizes the policy with a KL penalty to the reference model. The paper couples it with three reward signals: a binary format reward for structured <think>/<answer>/<glue> tags, a binary accuracy reward against ground truth, and a temporal IoU reward for the predicted anomaly interval. The other load-bearing object is VAU-Bench: 4,602 videos with 19 anomaly types, 3,700 temporal annotations, and LLM-written rationales that supply the reward targets and the evaluation ground truth.

What would settle it

A human audit of a random sample of VAU-Bench ground-truth rationales and temporal intervals against the raw videos would settle the core question: if a substantial fraction of rationales misidentify the anomaly or intervals miss the visible event, the reward targets and evaluation are unreliable. An independent check would be to re-score VAU-R1 outputs with a different, separately built judge plus human raters and see whether the RFT-over-SFT gains survive.

Watch

Extended reading notes

Core claim

The central discovery is that RFT over GRPO, with rewards tied to output format, answer correctness, and temporal intersection-over-union, improves a small video-language model's anomaly reasoning more than supervised fine-tuning (SFT) does, while also generalizing better to an unseen dataset. The paper decomposes video anomaly understanding into four tasks—multiple-choice QA, temporal grounding, anomaly reasoning, and anomaly classification—and reports that co-training with temporal grounding helps classification more than training classification alone. It also claims the first chain-of-thought benchmark for this setting, where each video carries descriptions, temporal annotations, QA pairs, and rationales, plus a GPT-based evaluation protocol that scores classification correctness, key-concept alignment, fluency, informativeness, and factual consistency.

Load-bearing premise

The load-bearing premise is that the LLM-written rationales, QA pairs, and temporal labels in VAU-Bench are accurate enough to serve as ground truth, and that the same model family's judgment scores truthfully reflect reasoning quality; if those annotations are biased, every improvement the paper measures could be an artifact.

Editorial extensions

If this is right

  • A 3B-parameter video-language model fine-tuned with RFT can outperform a 7B base model on temporal grounding in the paper's ECVA evaluation, so reinforcement fine-tuning can substitute for scale on some video anomaly tasks.
  • RFT generalizes better than SFT out-of-distribution: on UCF-Crime as an unseen test set, the RFT model keeps meaningful temporal grounding while SFT degrades.
  • Naive chain-of-thought prompting can hurt base models' QA accuracy; RFT is what turns reasoning into a gain rather than a source of hallucination.
  • Co-training temporal grounding with classification gives better anomaly classification than training classification alone, suggesting localization is a useful intermediate reward.
  • VAU-Bench ties perception, grounding, reasoning, and conclusion into one evaluation protocol, making reasoning quality measurable instead of assumed.

Reading between the lines

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

  • A testable extension: if the same format-plus-accuracy-plus-IoU reward design transfers to other video reasoning tasks, such as action anticipation or accident explanation, RFT could become a general post-training recipe rather than an anomaly-specific one.
  • Because the current grounding reward covers time only, adding spatial localization or object-level rewards could strengthen causal reasoning, a direction the paper itself flags as future work.
  • The benchmark's training targets and its judge come from the same LLM family, so an independent human audit of a small slice of VAU-Bench would clarify whether the measured reasoning gains reflect true understanding or annotation bias.
  • A deployment consequence not tested in the paper: if RFT's gains survive video compression and low frame rates, the pipeline could be retrained cheaply for new anomaly categories in changing environments.
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

4 major / 7 minor

Summary. The paper proposes VAU-R1, a GRPO-based reinforcement fine-tuning (RFT) framework for video anomaly understanding, together with VAU-Bench, a Chain-of-Thought benchmark composed of multiple-choice QA, temporal annotations, rationales, and captions. The method decomposes video anomaly understanding into QA, temporal grounding, anomaly reasoning, and classification, and uses task-specific rewards. The authors claim that RFT consistently improves over supervised fine-tuning and base models on QA accuracy, temporal grounding, and reasoning coherence, and they introduce VAU-Eval, a DeepSeek-V3-based metric for reasoning quality. The main evidence is in Tables 1-2 and 5-7, plus qualitative case studies.

Significance. If the claims held, the paper would contribute a data-efficient post-training recipe and a reusable benchmark for interpretable video anomaly understanding. The code release, the large annotation effort, and the cross-dataset evaluation are valuable. However, the central claims as stated are not supported by the reported numbers: several tables show RFT underperforming SFT or the base model, and the reasoning-quality metric is scored by the same model family that generated the ground-truth rationales. With human validation of the benchmark and a more careful analysis of when RFT helps, the contributions could be solid; in its current form the evidence is mixed.

major comments (4)
  1. [Section 4.1, Table 2] The statement that RFT 'consistently outperforms' SFT on temporal grounding is directly contradicted by Table 2. On ECVA, SFT reaches 45.30 mIoU versus RFT 35.09 without thinking and 45.96 versus 33.25 with thinking; on MSAD with thinking, SFT is 35.17 versus RFT 30.70. The same table also shows that adding thinking can hurt RFT relative to itself (e.g., MSAD RFT 35.77 without thinking vs 30.70 with thinking). The abstract's claim of 'significantly improves temporal grounding' is therefore not supported. Please report per-setting results without aggregating into a single narrative, add confidence intervals or significance tests, and qualify the claim to the settings where RFT actually helps.
  2. [Section 3.4, Appendix B, Appendix A] VAU-Eval is circular by construction. Section 3.4 states that DeepSeek-V3 generated the QA pairs and reasoning chains for UCF-Crime and ECVA, and Appendix B states that DeepSeek-V3 is the judge for the five VAU-Eval dimensions. The same model family thus defines the target rationales and then grades candidate rationales. The only reference to human checking is a one-line 'manual review' in Appendix A with no protocol, sample size, or agreement statistics. Consequently, the reported reasoning-quality gains may reflect the policy learning to match DeepSeek-V3's stylistic preferences rather than improved anomaly reasoning. An independent human evaluation of a sample, with inter-annotator agreement and a comparison of judge scores versus human scores, is necessary before VAU-Eval can support the paper's reasoning claims.
  3. [Table 1, Appendix C, Table 5] The reasoning-quality comparison is internally inconsistent. In Table 1, SFT collapses VAU-Eval scores (Qwen2.5-VL-3B on MSAD: SFT Total 15.96 vs base 32.47 and RFT 33.38; on UCF-Crime: SFT 10.89 vs base 25.10 and RFT 25.49), and RFT appears to restore roughly base-level performance. Yet Appendix C, Table 5 shows that on ECVA the RFT model's Total (18.45) is below the base model (20.89) and its CLS score (1.45) is below the SFT score (1.50). The main text's claim of 'consistent gains across five V AU-Eval dimensions' is false as stated, and Appendix C even concedes that 'V AU-Eval reasoning scores does not always improve.' The paper needs to reconcile these statements and explain why SFT degrades reasoning so drastically; otherwise 'RFT outperforms SFT' is not a meaningful comparison.
  4. [Table 3, Section 4.1] The ablation claims are not supported by Table 3. The table shows that RFT with CLS alone (60.30 binary accuracy without thinking) is below the untrained baseline (62.77), and adding QA and CLS to TAG (RFT w/ QA-TAG-CLS, 64.70) decreases binary accuracy relative to RFT w/ TAG alone (67.81). The sentence 'jointly optimizing tasks via reinforcement learning yields complementary gains in both accuracy and reasoning' is contradicted by these numbers. Please report the full task-combination matrix, including error bars, and adjust the interpretation to match the actual ordering of configurations.
minor comments (7)
  1. [Figure 1 caption] The caption repeats the 'consistently outperforms' claim that Table 2 disproves; please align the caption with the actual per-setting results.
  2. [Section 4.1] The sentence 'RFT leads to consistent gains across five V AU-Eval dimensions—classification,' ends abruptly and should be completed or removed.
  3. [Throughout] The paper uses inconsistent spacing in 'V AU' and 'V AU-Eval' throughout; please standardize the notation (e.g., 'VAU').
  4. [Table 2] The note about rows highlighted in light yellow is not visible in the text version; please mark the UCF-Crime rows explicitly with a symbol or boldface.
  5. [Appendix C] The sentence 'V AU-Eval reasoning scores does not always improve' is grammatically awkward and, more importantly, contradicts the main text; please move it to a prominent position and discuss when RFT helps or hurts.
  6. [All result tables] No confidence intervals or significance tests are reported; given the mixed results, these are needed to support the comparative claims.
  7. [References] Reference [9] is cited with two different titles in the related work ('Exploring what why and how' and 'Uncovering what why and how'); please verify the correct title.

Circularity Check

1 steps flagged · score 6.0 of 10

VAU-Eval reasoning scores are circular: DeepSeek-V3 generates the ground-truth rationales and also serves as the judge, with no reported human validation.

  1. self definitional [Section 3.4 (Dataset Construction and Annotation; Reasoning Evaluation Metric: VAU-Eval) and Appendix B (Evaluation Details for Anomaly Reasoning)]
    "For UCF-Crime and ECVA, we use DeepSeek-V3 [18] to generate video-level summaries, QA pairs, and reasoning chains. ... To better assess anomaly reasoning, we propose VAU-Eval, a GPT-based metric that compares model-generated descriptions and analyses with ground truth annotations. ... we evaluate each response along five dimensions using DeepSeek-V3 [18] as the judge. ... We then use DeepSeek-V3 [18] to assess the similarity between the predicted answers and the ground truth across five aspects."

    The ground-truth reasoning chains are written by DeepSeek-V3, and the same model scores every model output against those chains in VAU-Eval. There is no independent human scoring: the claimed manual review is described only as 'we perform a manual review to ensure the accuracy and consistency of all generated annotations' with no protocol, sample size, or agreement statistics. Therefore improvements in CLS/KM/FLU/INF/FAC reflect, at least in part, increased stylistic agreement with the generator-judge rather than objectively better anomaly reasoning. The central claim that VAU-R1 improves reasoning coherence is thus partly self-referential by construction.

full rationale

The paper's flagship reasoning claim rests on VAU-Eval scores, which are generated by the same model family that produced the benchmark's ground-truth rationales. Because DeepSeek-V3 both defines the reference reasoning chains and judges candidate outputs against them, the reported reasoning-quality gains are partially circular: a model that imitates DeepSeek-V3's writing style will appear to reason better even without any improvement in underlying anomaly understanding. The objective QA accuracy and temporal grounding numbers are more self-contained, and the Appendix C admission that 'their VAU-Eval reasoning scores does not always improve' (with Qwen2.5-VL-3B+RFT total 18.45 below base 20.89 on ECVA) further weakens the abstract's 'consistent gains' claim, though that is a correctness risk rather than an additional circular step. No load-bearing self-citation chain is present. Overall, the reasoning-evaluation portion is circular by construction, while the non-reasoning results retain independent content, yielding a score of 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central results depend on two unvalidated provenance assumptions: LLM-generated annotations are treated as ground truth, and the same LLM is treated as a valid judge. Reward weights are unreported. No new physical entities or fitted constants are introduced.

free parameters (3)
  • Reward weights lambda_k = not reported
    Equation (1) defines the total reward as a weighted sum of task-specific rewards, but the lambda values are never specified. Changing these weights could materially change the behavior of the RFT training.
  • GRPO candidate count M = 4
    Appendix B sets M=4 candidate outputs per prompt. No sensitivity analysis is provided, and this hyperparameter affects the variance of the normalized rewards.
  • KL penalty beta = 0.04
    Appendix B sets the KL regularization coefficient to 0.04. No ablation is reported for this value, which controls how far the fine-tuned policy can drift from the reference model.
assumptions (5)
  • domain assumption DeepSeek-V3-generated rationales and QA labels are accurate ground truth for video anomaly reasoning.
    Section 3.4 states that DeepSeek-V3 generates video summaries, QA pairs, and reasoning chains. A brief manual review is claimed but not detailed. All training rewards and evaluation scores inherit any errors in these labels.
  • domain assumption DeepSeek-V3 as judge produces valid reasoning-quality scores for VAU-Eval.
    Appendix B describes VAU-Eval as a multi-turn evaluation where DeepSeek-V3 scores classification, key matching, fluency, informativeness, and factual consistency. No correlation with human ratings is reported.
  • domain assumption Anomalies can be perceived from the adopted frame sampling rates.
    Section 4 says Qwen models sample at 1 FPS and InternVL models sample 16 frames per video. If key anomaly moments fall between sampled frames, both QA accuracy and temporal grounding will be limited.
  • domain assumption Existing dataset labels and anomaly type annotations in MSAD, UCF-Crime, and ECVA are correct.
    VAU-Bench is built from these three datasets. Any errors in their original labels propagate into the new benchmark and into the models trained on it.
  • standard math GRPO update equations are mathematically valid.
    Equations (1)-(3) are taken from the prior GRPO work [31] and are not re-derived in this paper.
invented entities (1)
  • VAU-Eval
    purpose: A GPT-based scoring protocol that evaluates anomaly reasoning across five dimensions using DeepSeek-V3 as judge.
    The paper reports no correlation between VAU-Eval scores and human judgments. The judge is the same model family that generated the benchmark annotations, so the metric has no independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VAU-R1: Advancing Video Anomaly Understanding via Reinforcement Fine-Tuning." pith.science (2026). https://pith.science/paper/ELEWJVUG

@misc{pith2026250523504,
  author       = {Pith},
  title        = {Pith review of: VAU-R1: Advancing Video Anomaly Understanding via Reinforcement Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ELEWJVUG}},
  note         = {Machine review of arXiv:2505.23504}
}
read the original abstract

Video Anomaly Understanding (VAU) is essential for applications such as smart cities, security surveillance, and disaster alert systems, yet remains challenging due to its demand for fine-grained spatio-temporal perception and robust reasoning under ambiguity. Despite advances in anomaly detection, existing methods often lack interpretability and struggle to capture the causal and contextual aspects of abnormal events. This limitation is further compounded by the absence of comprehensive benchmarks for evaluating reasoning ability in anomaly scenarios. To address both challenges, we introduce VAU-R1, a data-efficient framework built upon Multimodal Large Language Models (MLLMs), which enhances anomaly reasoning through Reinforcement Fine-Tuning (RFT). Besides, we propose VAU-Bench, the first Chain-of-Thought benchmark tailored for video anomaly reasoning, featuring multiple-choice QA, detailed rationales, temporal annotations, and descriptive captions. Empirical results show that VAU-R1 significantly improves question answering accuracy, temporal grounding, and reasoning coherence across diverse contexts. Together, our method and benchmark establish a strong foundation for interpretable and reasoning-aware video anomaly understanding. Our code is available at https://github.com/GVCLab/VAU-R1.

Figures

Figures reproduced from arXiv: 2505.23504 by the authors.

Figure 1
Figure 1. Effectiveness of Reinforcement Fine-Tuning. We compare QA accuracy and temporal anomaly grounding performance across different models. VAU-R1, trained via Reinforcement Fine-Tuning (RFT), consistently outperforms its Supervised Fine-Tuning (SFT) counterpart. This demonstrates that RFT enhances both reasoning and temporal localization capabilities in VAU tasks. 1 Introduction Anomalies are events or behaviors that de… view at source ↗
Figure 2
Figure 2. Overview of VAU-R1. VAU-R1 leverages Reinforcement Fine-Tuning to enhance the reasoning ability of MLLMs for video anomaly understanding. Specifically, we adopt Group Relative Policy Optimization (GRPO) to optimize the model with task-specific rewards, such as answer format, accuracy, and temporal Intersection-over-Union (IoU). We decompose the VAU task into four complementary tasks to facilitate comprehensive reaso… view at source ↗
Figure 3
Figure 3. Statistics of our VAU-Bench. (a) Distribution of main anomaly types. (b) Distribution of video durations (top) and the proportion of anomalous segments within each video (bottom). (c) The evaluation criteria for four VAU tasks. This simple accuracy reward encourages the model to choose the right answer during training. Temporal IoU Reward. To encourage precise temporal grounding, we introduce a temporal Intersection… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative case of the QA (top) and TAG (bottom) task. All ground-truths and correct answers are highlighted in orange. Both SFT and RFT perform inference using the same CoT prompt. RFT’s explicit chain-of-thought yields precise, interpretable QA choice and anomaly in…
Figure 5
Figure 5. Figure 5: More dataset statistics of our VAU-Bench. (a) Distribution of training, validation, and test splits across the four tasks included in VAU-Bench. (b) Word cloud visualization of frequent terms appearing in the multiple-choice questions and choices. Dataset Annotation. V…
Figure 6
Figure 6. Figure 6: Qualitative case of the Anomaly Reasoning task. All correct description and analysis are highlighted in orange. The evaluation results are presented on the right of the answer respectively. Both SFT and VAU-R1 perform inference using the same CoT prompt. VAU-R1’s outpu…
Figure 7
Figure 7. Figure 7: Example of VAU-Bench. An explosion case in an outdoor backyard, highlighting complex anomaly detection and dynamic scene understanding, labeled with a question-answer pair, key visual evidence, anomaly type, and a multi-part reasoning chain that includes location, caus…
Figure 8
Figure 8. Figure 8: Example of VAU-Bench. A stealing incident, demonstrating capabilities in human activity recognition and intent analysis. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Example of VAU-Bench. A normal scene, used to evaluate model robustness against false positives and to enhance dataset diversity. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. 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...

  2. AgenticVAU: Multi-Agent Explore-Verify Reasoning for Video Anomaly Understanding

    cs.CV 2026-08 conditional novelty 5.0 of 10

    A training-free multi-agent explore-verify system with four specialized agents and a shared evidence registry outperforms zero-shot and RL-finetuned baselines on video anomaly understanding benchmarks.

  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.

  4. Reinforcement Learning Meets Large Language Models: A Survey of Advancements and Applications Across the LLM Lifecycle

    cs.CL 2025-09 conditional novelty 3.0 of 10

    A survey that maps reinforcement learning methods, datasets, benchmarks, and open-source tools across the full training lifecycle of large language models, focusing on verifiable-reward reasoning.

Reference graph

Works this paper leans on

65 extracted references · 44 canonical work pages · cited by 4 Pith papers

  1. [1]

    Acsintoae, A

    A. Acsintoae, A. Florescu, M.-I. Georgescu, T. Mare, P. Sumedrea, R. T. Ionescu, F. S. Khan, and M. 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

    work page 2022

  2. [2]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin. Qwen2.5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  3. [3]

    J. Bi, S. Liang, X. Zhou, P. Liu, J. Guo, Y . Tang, L. Song, C. Huang, G. Sun, J. He, et al. Why reasoning matters? a survey of advancements in multimodal reasoning (v1).arXiv preprint arXiv:2504.03151, 2025

    arXiv 2025

  4. [4]

    C. Cao, Y . Lu, P. Wang, and Y . Zhang. A new comprehensive benchmark for semi-supervised video anomaly detection and anticipation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20392–20401, 2023

    work page 2023

  5. [5]

    Y . Chen, Z. Liu, B. Zhang, W. Fok, X. Qi, and Y .-C. Wu. Mgfn: Magnitude-contrastive glance- and-focus network for weakly-supervised video anomaly detection. InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 387–395, 2023

    work page 2023

  6. [6]

    D. Ding, L. Wang, L. Zhu, T. Gedeon, and P. Koniusz. Lego: Learnable expansion of graph operators for multi-modal feature fusion.arXiv preprint arXiv:2410.01506, 2024

    arXiv 2024

  7. [7]

    Doshi and Y

    K. Doshi and Y . Yilmaz. Towards interpretable video anomaly detection. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2655–2664, 2023

    work page 2023

  8. [8]

    H. Du, G. Nan, J. Qian, W. Wu, W. Deng, H. Mu, Z. Chen, P. Mao, X. Tao, and J. Liu. Exploring what why and how: A multifaceted benchmark for causation understanding of video anomaly. arXiv preprint arXiv:2412.07183, 2024

    arXiv 2024

Show all 65 references
  1. [9]

    H. Du, S. Zhang, B. Xie, G. Nan, J. Zhang, J. Xu, H. Liu, S. Leng, J. Liu, H. Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p...

    2024

  2. [10]

    K. Feng, K. Gong, B. Li, Z. Guo, Y . Wang, T. Peng, B. Wang, and X. Yue. Video-r1: Reinforcing video reasoning in mllms.arXiv preprint arXiv:2503.21776, 2025

    2025 arXiv

  3. [11]

    D. Gong, L. Liu, V . Le, B. Saha, M. R. Mansour, S. Venkatesh, and A. v. d. Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 17...

    2019

  4. [12]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  5. [13]

    K. Hara, H. Kataoka, and Y . Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018

    2018

  6. [14]

    Huang, B

    W. Huang, B. Jia, Z. Zhai, S. Cao, Z. Ye, F. Zhao, Z. Xu, Y . Hu, and S. Lin. Vision-r1: Incentiviz- ing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749, 2025

    2025 arXiv

  7. [15]

    Jaech, A

    A. Jaech, A. Kalai, A. Lerer, A. Richardson, A. El-Kishky, A. Low, A. Helyar, A. Madry, A. Beutel, A. Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024. 10

    2024 arXiv

  8. [16]

    J. Leng, Z. Wu, M. Tan, Y . Liu, J. Gan, H. Chen, and X. Gao. Beyond euclidean: Dual-space representation learning for weakly supervised video violence detection. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  9. [17]

    X. Li, Z. Yan, D. Meng, L. Dong, X. Zeng, Y . He, Y . Wang, Y . Qiao, Y . Wang, and L. Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning.arXiv preprint arXiv:2504.06958, 2025

    2025 arXiv

  10. [18]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  11. [19]

    K. Liu, W. Liu, C. Gan, M. Tan, and H. Ma. T-c3d: Temporal convolutional 3d network for real-time action recognition. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  12. [20]

    Liu and H

    K. Liu and H. Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490–1499, 2019

    2019

  13. [21]

    W. Liu, W. Luo, D. Lian, and S. Gao. Future frame prediction for anomaly detection–a new baseline. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6536–6545, 2018

    2018

  14. [22]

    Y . Liu, D. Yang, Y . Wang, J. Liu, J. Liu, A. Boukerche, P. Sun, and L. Song. Generalized video anomaly event detection: Systematic taxonomy and comparison of deep models.ACM Computing Surveys, 56(7):1–38, 2024

    2024

  15. [23]

    Z. Liu, Z. Sun, Y . Zang, X. Dong, Y . Cao, H. Duan, D. Lin, and J. Wang. Visual-rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025

    2025 arXiv

  16. [24]

    C. Lu, J. Shi, and J. Jia. Abnormal event detection at 150 fps in matlab. InProceedings of the IEEE international conference on computer vision, pages 2720–2727, 2013

    2013

  17. [25]

    Y . Lu, F. Yu, M. K. K. Reddy, and Y . Wang. Few-shot scene-adaptive anomaly detection. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 125–141. Springer, 2020

    2020

  18. [26]

    Lv and Q

    H. Lv and Q. Sun. Video anomaly detection and explanation via large language models.arXiv preprint arXiv:2401.05702, 2024

    2024 arXiv

  19. [27]

    G. Pang, C. Shen, L. Cao, and A. V . D. Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 54(2):1–38, 2021

    2021

  20. [28]

    Ramachandra and M

    B. Ramachandra and M. Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2569–2578, 2020

    2020

  21. [29]

    Rodrigues, N

    R. Rodrigues, N. Bhargava, R. Velmurugan, and S. Chaudhuri. Multi-timescale trajectory pre- diction for abnormal human activity detection. InThe IEEE Winter Conference on Applications of Computer Vision (WACV), March 2020

    2020

  22. [30]

    Y . Shao, H. He, S. Li, S. Chen, X. Long, F. Zeng, Y . Fan, M. Zhang, Z. Yan, A. Ma, et al. Eventvad: Training-free event-aware video anomaly detection.arXiv preprint arXiv:2504.13092, 2025

    2025 arXiv

  23. [31]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  24. [32]

    Simonyan and A

    K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos.Advances in neural information processing systems, 27, 2014

    2014

  25. [33]

    Sultani, C

    W. Sultani, C. Chen, and M. Shah. Real-world anomaly detection in surveillance videos. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6479–6488, 2018. 11

    2018

  26. [34]

    H. Tan, Y . Ji, X. Hao, M. Lin, P. Wang, Z. Wang, and S. Zhang. Reason-rft: Reinforcement fine-tuning for visual reasoning.arXiv preprint arXiv:2503.20752, 2025

    2025

  27. [35]

    J. Tang, H. Lu, R. Wu, X. Xu, K. Ma, C. Fang, B. Guo, J. Lu, Q. Chen, and Y . Chen. Hawk: Learning to understand open-world video anomalies.Advances in Neural Information Process- ing Systems, 37:139751–139785, 2024

    2024

  28. [36]

    Y . Tian, G. Pang, Y . Chen, R. Singh, J. W. Verjans, and G. Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. InProceedings of the IEEE/CVF international conference on computer vision, pages 4975–4986, 2021

    2021

  29. [37]

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. InProceedings of the IEEE international conference on computer vision, pages 4489–4497, 2015

    2015

  30. [38]

    Vijay, W.-X

    M. Vijay, W.-X. LI, B. Viral, and V . Nuno. Anomaly detection in crowded scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1975–1981, 2010

    1975

  31. [39]

    L. Wang, W. Li, W. Li, and L. Van Gool. Appearance-and-relation networks for video classi- fication. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1430–1439, 2018

    2018

  32. [40]

    L. Wang, Y . Xiong, Z. Wang, Y . Qiao, D. Lin, X. Tang, and L. Van Gool. Temporal segment networks: Towards good practices for deep action recognition. InEuropean conference on computer vision, pages 20–36. Springer, 2016

    2016

  33. [41]

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Ge, Y . Fan, K. Dang, M. Du, X. Ren, R. Men, D. Liu, C. Zhou, J. Zhou, and J. Lin. Qwen2-vl: En- hancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.1...

    2024 arXiv

  34. [42]

    W. Wang, Z. Chen, W. Wang, Y . Cao, Y . Liu, Z. Gao, J. Zhu, X. Zhu, L. Lu, Y . Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization.arXiv preprint arXiv:2411.10442, 2024

    2024 arXiv

  35. [43]

    X. Wang, R. Girshick, A. Gupta, and K. He. Non-local neural networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7794–7803, 2018

    2018

  36. [44]

    P. Wu, C. Pan, Y . Yan, G. Pang, P. Wang, and Y . Zhang. Deep learning for video anomaly detection: A review.arXiv preprint arXiv:2409.05383, 2024

    2024 arXiv

  37. [45]

    P. Wu, X. Zhou, G. Pang, Y . Sun, J. Liu, P. Wang, and Y . Zhang. Open-vocabulary video anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297–18307, 2024

    2024

  38. [46]

    P. Wu, X. Zhou, G. Pang, Z. Yang, Q. Yan, P. Wang, and Y . Zhang. Weakly supervised video anomaly detection and localization with spatio-temporal prompts. InProceedings of the 32nd ACM International Conference on Multimedia, pages 9301–9310, 2024

    2024

  39. [47]

    P. Wu, X. Zhou, G. Pang, L. Zhou, Q. Yan, P. Wang, and Y . Zhang. Vadclip: Adapting vision- language models for weakly supervised video anomaly detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6074–6082, 2024

    2024

  40. [48]

    Y . Yang, K. Lee, B. Dariush, Y . Cao, and S.-Y . Lo. Follow the rules: reasoning for video anomaly detection with large language models. InEuropean Conference on Computer Vision, pages 304–322. Springer, 2024

    2024

  41. [49]

    M. Ye, W. Liu, and P. He. Vera: Explainable video anomaly detection via verbalized learning of vision-language models.arXiv preprint arXiv:2412.01095, 2024

    2024 arXiv

  42. [50]

    T. Yuan, X. Zhang, K. Liu, B. Liu, C. Chen, J. Jin, and Z. Jiao. Towards surveillance video- and-language understanding: New dataset baselines and challenges. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22052–22061, 2024. 12

    2024

  43. [51]

    Zanella, W

    L. Zanella, W. Menapace, M. Mancini, Y . Wang, and E. Ricci. Harnessing large language models for training-free video anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527–18536, 2024

    2024

  44. [52]

    Zhang, X

    H. Zhang, X. Xu, X. Wang, J. Zuo, C. Han, X. Huang, C. Gao, Y . Wang, and N. Sang. Holmes- vad: Towards unbiased and explainable video anomaly detection via multi-modal llm.arXiv preprint arXiv:2406.12235, 2024

    2024 arXiv

  45. [53]

    Zhang, X

    H. Zhang, X. Xu, X. Wang, J. Zuo, X. Huang, C. Gao, S. Zhang, L. Yu, and N. Sang. Holmes- vau: Towards long-term video anomaly understanding at any granularity.arXiv preprint arXiv:2412.06171, 2024

    2024 arXiv

  46. [54]

    aha moment

    H. Zhou, X. Li, R. Wang, M. Cheng, T. Zhou, and C.-J. Hsieh. R1-zero’s" aha moment" in visual reasoning on a 2b non-sft model.arXiv preprint arXiv:2503.05132, 2025

    2025 arXiv

  47. [55]

    H. Zhou, J. Yu, and W. 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

    2023

  48. [56]

    judgement

    L. Zhu, L. Wang, A. Raj, T. Gedeon, and C. Chen. Advancing video anomaly detection: A concise review and a new dataset. InThe Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 13 A Further Dataset Details train:2939 val:734 t...

    2024

  49. [57]

    Classification Correctness (10 pts)

  50. [58]

    Key Object and Action Matching (10 pts)

  51. [59]

    Fluency and Coherence (10 pts)

  52. [60]

    Informativeness and Domain Awareness (10 pts)

  53. [61]

    describing a political argument instead of an escalator malfunction

    Factual Consistency (10 pts) Evaluation Details for Anomaly Reasoning.To evaluate the alignment between model-generated outputs and our annotated ground truth in video anomaly understanding, we introduceV AU-Eval, a GPT-based evaluation protocol. The evaluation is structured a...

  54. [62]

    **Initial Setting**: The video begins with a group of people gathered in a backyard, sitting on chairs and lounging on a blanket, indicating a casual and relaxed atmosphere

  55. [63]

    This is evident from the bright flash, smoke, and sparks

    **Key Action**: A firework unexpectedly ignites, leading to a sudden explosion. This is evident from the bright flash, smoke, and sparks

  56. [64]

    This reaction is typical in response to a sudden and dangerous event

    **Reactions**: The people react with panic, quickly getting up and moving away from the explosion site. This reaction is typical in response to a sudden and dangerous event

  57. [65]

    ### Cause and Effect: - **Cause**: The firework malfunctioned or was mishandled, leading to the explosion

    **Environmental Factors**: The outdoor setting, with open space and proximity to a driveway, suggests that the firework was intended to be set off in a controlled manner, but the explosion was unanticipated and hazardous. ### Cause and Effect: - **Cause**: The firework malfunc...

Pith tools

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