pith. sign in

arxiv: 2605.15325 · v1 · pith:XMXDSXEDnew · submitted 2026-05-14 · 💻 cs.CV

COPRA: Conditional Parameter Adaptation with Reinforcement Learning for Video Anomaly Detection

Darryl Cherian Jacob , Xinyu Liu , Kai Wang , Pan He This is my paper

Pith reviewed 2026-05-19 16:09 UTC · model grok-4.3

classification 💻 cs.CV
keywords video anomaly detectionvision-language modelsreinforcement learningparameter adaptationconditional adaptationcross-domain generalizationvideo understanding
0
0 comments X

The pith

COPRA uses reinforcement learning to generate input-specific parameter updates that dynamically adapt a frozen vision-language model to each video segment for anomaly detection.

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

Existing VLM approaches to video anomaly detection rely on static adaptations after training and process sparse frames during training but dense segments at inference, creating mismatches that hurt generalization to new environments or anomaly types. COPRA instead trains a reinforcement learning policy to output parameter updates conditioned on the current video input, applying these updates to a frozen base model in both phases. This closes the distribution and configuration gaps while keeping the core model unchanged. The method improves results on standard benchmarks and transfers to unrelated video tasks such as multiple-choice question answering and dense captioning. A reader would care because it offers a way to make large pretrained models responsive to shifting video contexts without repeated full fine-tuning.

Core claim

COPRA is a conditional parameter adaptation framework that employs reinforcement learning to produce input-specific parameter updates, thereby dynamically adapting a frozen vision-language model to each video segment during both training and inference and resolving mismatches in data distribution and model configuration that limit prior VLM-based video anomaly detection methods.

What carries the argument

A reinforcement learning policy that outputs input-conditioned parameter updates applied to a frozen VLM for each video segment.

If this is right

  • COPRA outperforms static adaptation baselines on in-domain video anomaly detection benchmarks.
  • It improves generalization to cross-domain settings involving unseen environments or anomaly types.
  • The same adaptation mechanism transfers to unseen tasks such as multiple-choice video question answering and dense captioning.
  • It functions as a weight-space generation approach supporting scalable and context-aware video understanding.

Where Pith is reading between the lines

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

  • Keeping the base VLM frozen while generating small per-segment updates could lower the cost of deploying such models across many different video domains.
  • The approach may suit real-time video monitoring where input statistics change over time without requiring model reloading.
  • Conditional parameter generation of this form could be tested on other foundation models facing similar train-inference distribution shifts in vision or multimodal settings.

Load-bearing premise

Reinforcement learning can stably generate useful input-conditioned parameter updates for a frozen VLM without instability or extensive per-domain hyperparameter tuning when new video distributions appear.

What would settle it

On a held-out video dataset with clear distribution shift, COPRA matching or underperforming a static post-training adaptation baseline in detection accuracy would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.15325 by Darryl Cherian Jacob, Kai Wang, Pan He, Xinyu Liu.

Figure 1
Figure 1. Figure 1: Illustration of the generalization gap in VAD caused by training–inference mismatch and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of COPRA, an instance-conditioned parameter generation framework for VAD. The key advantage is a learned functional memory that maps each instance to tailored parameters, enabling improved robustness and cross-domain generalization compared to static adaptation methods. Overview. Our COPRA framework adapts VLM parameters on a per-instance basis to overcome the limitations of static, shared adaptat… view at source ↗
Figure 3
Figure 3. Figure 3: AUC (%) performance for cross-dataset generalization across train–test dataset pairs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cross-domain evaluation of natural language explanation quality on HIVAU-70K [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative analysis of InternVL2-8B + COPRA representations and responses. Left: Under [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative temporal anomaly scoring example on XD-Violence. The grey bars show [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of COPRA’s segment-level reasoning process on an arrest video from UCF [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
read the original abstract

Vision-language models (VLMs) have shown strong performance in video anomaly detection (VAD) while providing interpretable predictions. However, existing VLM-based VAD methods suffer from a fundamental mismatch between training and inference in both data distribution and model configuration. First, most approaches rely on static post-training adaptation, limiting generalization under distribution shifts such as unseen environments or anomaly types. Second, they train VLMs on sparse frames from long videos, but perform inference on densely sampled short segments, creating inconsistencies between training and testing. To address these limitations, we propose COPRA, a conditional parameter adaptation framework for VLM-based VAD. Instead of fixed prompts or shared parameter updates, COPRA generates input-specific parameter updates to dynamically adapt a frozen VLM for each video segment during both training and inference. Experiments show strong performance on standard VAD benchmarks, consistently outperforming static baselines in both in-domain and cross-domain settings. Moreover, COPRA generalizes beyond VAD to unseen tasks such as multiple-choice Video Question Answering and Dense Captioning. These results highlight COPRA as an effective weight-space generation framework for scalable, adaptive, and context-aware video understanding. The code will be released at https://github.com/THE-MALT-LAB/COPRA

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The manuscript proposes COPRA, a conditional parameter adaptation framework that employs reinforcement learning to generate input-specific parameter updates for dynamically adapting a frozen vision-language model (VLM) on a per-video-segment basis. This addresses mismatches in data distribution and model configuration between training (sparse frames from long videos) and inference (dense short segments) in video anomaly detection (VAD). The method is claimed to outperform static baselines in both in-domain and cross-domain settings on standard VAD benchmarks while also generalizing to unseen tasks such as multiple-choice Video Question Answering and Dense Captioning. Code release is promised.

Significance. If the results hold under scrutiny, the work offers a potentially impactful approach to making VLMs more adaptive and generalizable for video tasks without full retraining, by operating in weight space via RL-generated deltas. Strengths include the explicit handling of train-inference mismatch, the cross-domain and cross-task generalization claims, and the commitment to code release which aids reproducibility. This could influence future designs for context-aware video understanding systems.

major comments (3)
  1. [Section 3.2] Section 3.2 (Policy Network): The architecture and conditioning mechanism of the policy network that produces input-specific parameter deltas are described at a high level but lack sufficient implementation details (e.g., input features from the video segment, output parameterization of the updates, and any regularization to prevent large deviations from the frozen VLM weights). This is load-bearing for the central claim of stable, useful conditional adaptation.
  2. [Section 4.2] Section 4.2 and Table 3: The reward formulation for the RL objective is not specified in enough detail to evaluate alignment with VAD performance metrics or robustness to distribution shifts; without this, it is difficult to determine whether reported gains rely on dataset-specific tuning or generalize as claimed.
  3. [Section 4.3] Section 4.3 (Ablations): No ablation studies isolate the contribution of the RL component versus simpler conditioning mechanisms (e.g., direct feature concatenation or hypernetwork baselines), which is necessary to substantiate that reinforcement learning is the key enabler rather than the overall adaptation idea.
minor comments (3)
  1. [Abstract] Abstract: The claim of 'strong performance' would be strengthened by briefly noting the specific metrics (e.g., AUC or F1) and number of benchmarks used.
  2. [Figure 2] Figure 2: The framework diagram would benefit from explicit arrows or labels indicating the flow of the RL-generated parameter updates during inference.
  3. [Related Work] Related Work: A short paragraph contrasting COPRA with prior parameter-efficient adaptation methods (e.g., LoRA or prompt tuning in VLMs) would clarify novelty.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and will revise the paper to provide the requested clarifications and additional experiments. These changes will improve the technical exposition and strengthen the presentation of our contributions.

read point-by-point responses

  1. Referee: [Section 3.2] Section 3.2 (Policy Network): The architecture and conditioning mechanism of the policy network that produces input-specific parameter deltas are described at a high level but lack sufficient implementation details (e.g., input features from the video segment, output parameterization of the updates, and any regularization to prevent large deviations from the frozen VLM weights). This is load-bearing for the central claim of stable, useful conditional adaptation.

    Authors: We agree that Section 3.2 would benefit from greater specificity to support reproducibility and the central claims. In the revised manuscript we will expand the section to describe: (1) the precise input features to the policy network, which are pooled visual embeddings from the frozen VLM encoder applied to the current video segment; (2) the output parameterization, in which the policy produces low-rank deltas (LoRA-style) for selected attention and feed-forward layers; and (3) the regularization, consisting of an L2 penalty on delta magnitude together with a KL term that keeps the adapted distribution close to the original frozen weights. We will also add pseudocode and a schematic diagram of the policy network. revision: yes

  2. Referee: [Section 4.2] Section 4.2 and Table 3: The reward formulation for the RL objective is not specified in enough detail to evaluate alignment with VAD performance metrics or robustness to distribution shifts; without this, it is difficult to determine whether reported gains rely on dataset-specific tuning or generalize as claimed.

    Authors: We thank the referee for noting this gap. The reward is constructed to align directly with VAD evaluation metrics while promoting robustness: it comprises a primary term that rewards accurate anomaly scoring (frame-level AUC) and a consistency term across consecutive segments, plus a small penalty on policy output magnitude to discourage overfitting to training distributions. In the revision we will present the complete mathematical definition of the reward, list all weighting coefficients, and add a short analysis of reward-component sensitivity together with results under alternative formulations to illustrate generalization behavior. revision: yes

  3. Referee: [Section 4.3] Section 4.3 (Ablations): No ablation studies isolate the contribution of the RL component versus simpler conditioning mechanisms (e.g., direct feature concatenation or hypernetwork baselines), which is necessary to substantiate that reinforcement learning is the key enabler rather than the overall adaptation idea.

    Authors: We accept that the existing ablations in Section 4.3 compare COPRA primarily against static and non-conditional baselines and do not yet isolate the RL policy against simpler conditioning alternatives. In the revised version we will add new ablation experiments that implement (i) direct feature concatenation into the VLM and (ii) a hypernetwork baseline that generates the same parameter updates without RL. Performance on the standard VAD benchmarks will be reported for these variants, allowing readers to assess the specific benefit of the reinforcement-learning-driven policy. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method without self-referential derivations

full rationale

The paper proposes an empirical RL-based framework for input-conditioned parameter adaptation of frozen VLMs, evaluated via experiments on VAD benchmarks and generalization tasks. No equations, derivations, or fitted parameters are presented that reduce any claimed prediction or result to the inputs by construction. The method description relies on standard RL concepts applied to parameter generation, with performance claims grounded in reported benchmark comparisons rather than self-citation chains or definitional equivalences. This is a standard empirical contribution with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract does not introduce explicit free parameters, axioms, or new entities; the method implicitly relies on standard RL training assumptions and VLM fine-tuning practices from prior literature.

pith-pipeline@v0.9.0 · 5761 in / 1077 out tokens · 43650 ms · 2026-05-19T16:09:32.571812+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

64 extracted references · 64 canonical work pages · 2 internal anchors

  1. [1]

    WACV , pages=

    Unlocking vision-language models for video anomaly detection via fine-grained prompting , author=. WACV , pages=

    work page

  2. [2]

    CVPR , pages=

    Real-world anomaly detection in surveillance videos , author=. CVPR , pages=

    work page

  3. [3]

    Workshop on Neural Network Weights as a New Data Modality , year=

    Uncovering Latent Chain of Thought Vectors in Large Language Models , author=. Workshop on Neural Network Weights as a New Data Modality , year=

    work page

  4. [4]

    CVPR , pages=

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks , author=. CVPR , pages=

    work page

  5. [5]

    2026 , eprint=

    A Survey of Weight Space Learning: Understanding, Representation, and Generation , author=. 2026 , eprint=

    work page 2026

  6. [6]

    Huang, Chao and Shi, Yushu and Wen, Jie and Wang, Wei and Xu, Yong and Cao, Xiaochun , booktitle =. Ex-. 2025 , volume =

    work page 2025

  7. [7]

    Zhang, Huaxin and Xu, Xiaohao and Wang, Xiang and Zuo, Jialong and Han, Chuchu and Huang, Xiaonan and Gao, Changxin and Wang, Yuehuan and Sang, Nong , journal=

    work page

  8. [8]

    Zhu, Liyun and Chen, Qixiang and Shen, Xi and Cun, Xiaodong , journal=

    work page

  9. [9]

    ICML , pages =

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time , author =. ICML , pages =. 2022 , volume =

    work page 2022

  10. [10]

    ICLR , year=

    Editing models with task arithmetic , author=. ICLR , year=

    work page

  11. [11]

    2024 , eprint=

    Neural Network Diffusion , author=. 2024 , eprint=

    work page 2024

  12. [12]

    NeurIPS , year=

    Self-Supervised Representation Learning on Neural Network Weights for Model Characteristic Prediction , author=. NeurIPS , year=

    work page

  13. [13]

    2022 , eprint=

    Learning to Learn with Generative Models of Neural Network Checkpoints , author=. 2022 , eprint=

    work page 2022

  14. [14]

    ICCV , pages=

    What does a platypus look like? generating customized prompts for zero-shot image classification , author=. ICCV , pages=

    work page

  15. [15]

    2024 , eprint=

    Video Anomaly Detection and Explanation via Large Language Models , author=. 2024 , eprint=

    work page 2024

  16. [16]

    Visual Instruction Tuning , url =

    Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae , booktitle =. Visual Instruction Tuning , url =

    work page

  17. [17]

    MLE-UVAD: Minimal Latent Entropy Autoencoder for Fully Unsupervised Video Anomaly Detection

    Yuang Geng and Junkai Zhou and Kang Yang and Pan He and Zhuoyang Zhou and Jose C. Principe and Joel Harley and Ivan Ruchkin , year=. 2603.23868 , archivePrefix=

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    ICLR , year=

    HyperNetworks , author=. ICLR , year=

    work page

  19. [19]

    Edward J Hu and yelong shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , booktitle=. Lo. 2022 , url=

    work page 2022

  20. [20]

    Federated Learning for Video Violence Detection: Complementary Roles of Lightweight CNNs and Vision-Language Models for Energy-Efficient Use , year=

    Thuau, S\'ebastien and Haidar, Siba and Chelouah, Rachid , booktitle=. Federated Learning for Video Violence Detection: Complementary Roles of Lightweight CNNs and Vision-Language Models for Energy-Efficient Use , year=

    work page

  21. [21]

    Journal of Imaging , VOLUME =

    Borodin, Kirill and Kondrashov, Kirill and Vasiliev, Nikita and Gladkova, Ksenia and Larina, Inna and Gorodnichev, Mikhail and Mkrtchian, Grach , TITLE =. Journal of Imaging , VOLUME =. 2025 , NUMBER =

    work page 2025

  22. [22]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages =

    Cai, Zhaolin and Li, Fan and Zheng, Ziwei and Qin, Yanjun , title =. Proceedings of the 33rd ACM International Conference on Multimedia , pages =. 2025 , isbn =. doi:10.1145/3746027.3755575 , abstract =

    work page doi:10.1145/3746027.3755575 2025

  23. [23]

    CVPR , month =

    Ye, Muchao and Liu, Weiyang and He, Pan , title =. CVPR , month =. 2025 , pages =

    work page 2025

  24. [24]

    CVPR , pages=

    Ubnormal: New benchmark for supervised open-set video anomaly detection , author=. CVPR , pages=

    work page

  25. [25]

    Zhang, Huaxin and Xu, Xiaohao and Wang, Xiang and Zuo, Jialong and Huang, Xiaonan and Gao, Changxin and Zhang, Shanjun and Yu, Li and Sang, Nong , booktitle=. Holmes-. 2025 , volume=

    work page 2025

  26. [26]

    ECCV , pages =

    Yang, Yuchen and Lee, Kwonjoon and Dariush, Behzad and Cao, Yinzhi and Lo, Shao-Yuan , title =. ECCV , pages =. 2024 , isbn =. doi:10.1007/978-3-031-73004-7_18 , abstract =

    work page doi:10.1007/978-3-031-73004-7_18 2024

  27. [27]

    CVPR , pages=

    Harnessing Large Language Models for Training-free Video Anomaly Detection , author=. CVPR , pages=

    work page

  28. [28]

    Qwen2.5-VL Technical Report

    Shuai Bai and Keqin Chen and Xuejing Liu and Jialin Wang and Wenbin Ge and Sibo Song and Kai Dang and Peng Wang and Shijie Wang and Jun Tang and Humen Zhong and Yuanzhi Zhu and Mingkun Yang and Zhaohai Li and Jianqiang Wan and Pengfei Wang and Wei Ding and Zheren Fu and Yiheng Xu and Jiabo Ye and Xi Zhang and Tianbao Xie and Zesen Cheng and Hang Zhang and...

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    2025 , eprint=

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling , author=. 2025 , eprint=

    work page 2025

  30. [30]

    CVPR , month =

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Lee, Yong Jae , title =. CVPR , month =. 2024 , pages =

    work page 2024

  31. [31]

    The CropAndWeed Dataset: A Multi-Modal Learning Approach for Efficient Crop and Weed Manipulation

    Thakare, Kamalakar Vijay and Raghuwanshi, Yash and Dogra, Debi Prosad and Choi, Heeseung and Kim, Ig-Jae , booktitle =. 2023 , volume =. doi:10.1109/WACV56688.2023.00550 , url =

    work page doi:10.1109/wacv56688.2023.00550 2023

  32. [32]

    Zaigham and Mahmood, Arif and Khan, M

    Zaheer, M. Zaigham and Mahmood, Arif and Khan, M. Haris and Segu, Mattia and Yu, Fisher and Lee, Seung-Ik , booktitle=. Generative Cooperative Learning for Unsupervised Video Anomaly Detection , year=

    work page

  33. [33]

    CVPR , pages=

    Learning Memory-guided Normality for Anomaly Detection , author=. CVPR , pages=

    work page

  34. [34]

    Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection , year=

    Gong, Dong and Liu, Lingqiao and Le, Vuong and Saha, Budhaditya and Mansour, Moussa Reda and Venkatesh, Svetha and Van Den Hengel, Anton , booktitle=. Memorizing Normality to Detect Anomaly: Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection , year=

    work page

  35. [35]

    URL https://ieeexplore.ieee.org/document/ 10658325/

    Wu, Peng and Zhou, Xuerong and Pang, Guansong and Sun, Yujia and Liu, Jing and Wang, Peng and Zhang, Yanning , booktitle =. 2024 , volume =. doi:10.1109/CVPR52733.2024.01732 , url =

    work page doi:10.1109/cvpr52733.2024.01732 2024

  36. [36]

    AAAI , articleno =

    Zhou, Hang and Yu, Junqing and Yang, Wei , title =. AAAI , articleno =. 2023 , isbn =. doi:10.1609/aaai.v37i3.25489 , abstract =

    work page doi:10.1609/aaai.v37i3.25489 2023

  37. [37]

    AAAI , articleno =

    Wu, Peng and Zhou, Xuerong and Pang, Guansong and Zhou, Lingru and Yan, Qingsen and Wang, Peng and Zhang, Yanning , title =. AAAI , articleno =. 2024 , isbn =. doi:10.1609/aaai.v38i6.28423 , abstract =

    work page doi:10.1609/aaai.v38i6.28423 2024

  38. [38]

    Text Prompt with Normality Guidance for Weakly Supervised Video Anomaly Detection , year=

    Yang, Zhiwei and Liu, Jing and Wu, Peng , booktitle=. Text Prompt with Normality Guidance for Weakly Supervised Video Anomaly Detection , year=

    work page

  39. [39]

    2023 , volume=

    Joo, Hyekang Kevin and Vo, Khoa and Yamazaki, Kashu and Le, Ngan , booktitle=. 2023 , volume=

    work page 2023

  40. [40]

    AAAI , articleno =

    Chen, Yingxian and Liu, Zhengzhe and Zhang, Baoheng and Fok, Wilton and Qi, Xiaojuan and Wu, Yik-Chung , title =. AAAI , articleno =. 2023 , isbn =. doi:10.1609/aaai.v37i1.25112 , abstract =

    work page doi:10.1609/aaai.v37i1.25112 2023

  41. [41]

    AAAI , year=

    Self-Training Multi-Sequence Learning with Transformer for Weakly Supervised Video Anomaly Detection , author=. AAAI , year=

    work page

  42. [42]

    Caron, H

    Tian, Yu and Pang, Guansong and Chen, Yuanhong and Singh, Rajvinder and Verjans, Johan W. and Carneiro, Gustavo , booktitle =. 2021 , volume =. doi:10.1109/ICCV48922.2021.00493 , url =

    work page doi:10.1109/iccv48922.2021.00493 2021

  43. [43]

    ECCV , pages=

    Scale-aware spatio-temporal relation learning for video anomaly detection , author=. ECCV , pages=. 2022 , organization=

    work page 2022

  44. [44]

    and Abdel-Aty, Mohamed , title =

    Kim, Younggun and Abdelrahman, Ahmed S. and Abdel-Aty, Mohamed , title =. ICCV , month =. 2025 , pages =

    work page 2025

  45. [45]

    , journal=

    Yao, Yu and Wang, Xizi and Xu, Mingze and Pu, Zelin and Wang, Yuchen and Atkins, Ella and Crandall, David J. , journal=. 2023 , volume=. doi:10.1109/TPAMI.2022.3150763 , url =

    work page doi:10.1109/tpami.2022.3150763 2023

  46. [46]

    Li and Y

    Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y.K. Li and Y. Wu and Daya Guo , title =. CoRR , volume =. 2024 , url =

    work page 2024

  47. [47]

    Shen, Haozhan and Liu, Peng and Li, Jingcheng and Fang, Chunxin and Ma, Yibo and Liao, Jiajia and Shen, Qiaoli and Zhang, Zilun and Zhao, Kangjia and Zhang, Qianqian and others , journal=

    work page

  48. [48]

    ECCV , year=

    Not only Look, but also Listen: Learning Multimodal Violence Detection under Weak Supervision , author=. ECCV , year=

    work page

  49. [49]

    SPICE : Semantic Propositional Image Caption Evaluation

    Anderson, Peter and Fernando, Basura and Johnson, Mark and Gould, Stephen. SPICE : Semantic Propositional Image Caption Evaluation. ECCV. 2016

    work page 2016

  50. [50]

    METEOR : An Automatic Metric for MT Evaluation with High Levels of Correlation with Human Judgments

    Lavie, Alon and Agarwal, Abhaya. METEOR : An Automatic Metric for MT Evaluation with High Levels of Correlation with Human Judgments. Proceedings of the Second Workshop on Statistical Machine Translation. 2007

    work page 2007

  51. [51]

    COMET : A Neural Framework for MT Evaluation

    Rei, Ricardo and Stewart, Craig and Farinha, Ana C and Lavie, Alon. COMET : A Neural Framework for MT Evaluation. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18653/v1/2020.emnlp-main.213

    work page doi:10.18653/v1/2020.emnlp-main.213 2020

  52. [52]

    ROUGE : A Package for Automatic Evaluation of Summaries

    Lin, Chin-Yew. ROUGE : A Package for Automatic Evaluation of Summaries. Text Summarization Branches Out. 2004

    work page 2004

  53. [53]

    Bleu: a Method for Automatic Evaluation of Machine Translation , booktitle =

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing. B leu: a Method for Automatic Evaluation of Machine Translation. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. 2002. doi:10.3115/1073083.1073135

    work page doi:10.3115/1073083.1073135 2002

  54. [54]

    Lawrence Zitnick and Devi Parikh , booktitle=

    Ramakrishna Vedantam and C. Lawrence Zitnick and Devi Parikh , booktitle=. 2014 , pages=

    work page 2014

  55. [55]

    ECCV , pages =

    Wu, Jhih-Ciang and Hsieh, He-Yen and Chen, Ding-Jie and Fuh, Chiou-Shann and Liu, Tyng-Luh , title =. ECCV , pages =. 2022 , isbn =. doi:10.1007/978-3-031-19778-9_42 , abstract =

    work page doi:10.1007/978-3-031-19778-9_42 2022

  56. [56]

    CVPR , year=

    Learning Temporal Regularity in Video Sequences , author=. CVPR , year=

    work page

  57. [57]

    ICCV , pages =

    Lu, Cewu and Shi, Jianping and Jia, Jiaya , title =. ICCV , pages =. 2013 , isbn =. doi:10.1109/ICCV.2013.338 , abstract =

    work page doi:10.1109/iccv.2013.338 2013

  58. [58]

    ICCV , pages=

    Gods: Generalized one-class discriminative subspaces for anomaly detection , author=. ICCV , pages=

    work page

  59. [59]

    2023 , issue_date =

    Thakare, Kamalakar Vijay and Dogra, Debi Prosad and Choi, Heeseung and Kim, Haksub and Kim, Ig-Jae , title =. 2023 , issue_date =. doi:10.1016/j.patcog.2023.109567 , journal =

    work page doi:10.1016/j.patcog.2023.109567 2023

  60. [60]

    2019 , organization=

    Fang, Jianwu and Yan, Dingxin and Qiao, Jiahuan and Xue, Jianru and Wang, He and Li, Sen , booktitle=. 2019 , organization=

    work page 2019

  61. [61]

    Silhouettes: A graphical aid to the in- terpretation and validation of cluster analysis,

    Peter J. Rousseeuw , keywords =. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis , journal =. 1987 , issn =. doi:https://doi.org/10.1016/0377-0427(87)90125-7 , url =

    work page doi:10.1016/0377-0427(87)90125-7 1987

  62. [62]

    Journal of Machine Learning Research , year =

    Laurens van der Maaten and Geoffrey Hinton , title =. Journal of Machine Learning Research , year =

    work page

  63. [63]

    Scaling Sentence Embeddings with Large Language Models

    Jiang, Ting and Huang, Shaohan and Luan, Zhongzhi and Wang, Deqing and Zhuang, Fuzhen. Scaling Sentence Embeddings with Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.181

    work page doi:10.18653/v1/2024.findings-emnlp.181 2024

  64. [64]

    2025 , eprint=

    InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models , author=. 2025 , eprint=

    work page 2025