Research map / Video / Video segmentation and tracking
Steering and Rectifying Latent representation manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection
Zhaolin Cai, Fan Li, Huiyu Duan, Lijun He, Guangtao Zhai
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global scrutiny gate (ablation) | SteerVAD · UCF-Crime | AUC: 85.94 -> 87.15 | Improved |
| use anisotropic steering instead of additive steering (ablation) | SteerVAD · UCF-Crime | AUC: 85.02 -> 87.15 | Improved |
| add local gating module instead of static scaling (ablation) | SteerVAD · UCF-Crime | AUC: 84.21 -> 87.15 | Improved |
| add hierarchical meta-controller on top of linear probing (ablation) | SteerVAD · UCF-Crime | AUC: 81.33 -> 87.15 | Improved |
| select heads with RSA instead of mid-to-last layers (ablation) | SteerVAD · UCF-Crime | AUC: 84.68 -> 87.15 | Improved |
| select heads with RSA instead of random heads (ablation) | SteerVAD · UCF-Crime | AUC: 69.57 -> 87.15 | Improved |
| select heads with RSA instead of first-to-mid layers (ablation) | SteerVAD · UCF-Crime | AUC: 63.16 -> 87.15 | Improved |
| increase calibration data from 1% to 100% (ablation) | SteerVAD · UCF-Crime | AUC: 87.15 -> 87.42 | No clear effect |
| change random seed of 1% calibration subset (ablation) | SteerVAD · UCF-Crime | AUC: 87.15 +- 0.04 | No clear effect |
| increase experts from 4 to 8 or 16 (ablation) | SteerVAD · UCF-Crime | AUC: 87.15 -> 87.09 -> 86.88 | Hurt |
| reduce experts from 4 to 2 (ablation) | SteerVAD · UCF-Crime | AUC: 87.15 -> 85.91 | Hurt |
| increase input frames from 4 to 16 (ablation) | SteerVAD · UCF-Crime | AUC: 82.54 -> 87.15 | Improved |
Closest papers by meaning
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.91)
- MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection (NeurIPS 2025 · similarity 0.90)
- Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · similarity 0.89)
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · similarity 0.89)
- VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree (NeurIPS 2025 · similarity 0.89)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
