Research map / Video / Video segmentation and tracking
HeadHunt-VAD: Hunting Robust Anomaly-Sensitive Heads in MLLM for Tuning-Free Video Anomaly Detection
Zhaolin Cai, Fan Li, Ziwei Zheng, Haixia Bi, Lijun He
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 |
|---|---|---|---|
| probe attention-head features instead of aggregated final-layer features (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 72.10 -> 87.03 | Improved |
| select heads with Robust Head Identification instead of random K heads (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 66.65 -> 87.03 | Improved |
| use RHI across five prompts instead of a single coarse prompt (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 81.86 -> 87.03 | Improved |
| use manually crafted oracle prompt instead of RHI (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 87.03 -> 87.11 | No clear effect |
| replace logistic regression scorer with 2-layer MLP (ablation) | HeadHunt-VAD · UCF-Crime / XD-Violence | AUC / AP: +0.22% AUC, +0.18% AP | No clear effect |
| replace logistic regression scorer with linear SVM (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 87.03 -> 84.95 | Hurt |
| add Gaussian smoothing in temporal locator (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 82.44 -> 87.03 | Improved |
| calibrate threshold on validation set instead of fixed values (ablation) | HeadHunt-VAD · UCF-Crime | AUC (%): 80.32 -> 87.03 (best fixed tau=0.50) | Improved |
| use head-level probing instead of text generation (LAVAD, LLaVA-1.5) | HeadHunt-VAD · UCF-Crime | AUC (%): 80.28 -> 87.03 (LAVAD) | Improved |
| single forward pass over top-K heads instead of autoregressive decoding | HeadHunt-VAD · inference efficiency | speed: Speed:+89.8% | Improved |
Closest papers by meaning
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.89)
- Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
- MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection (NeurIPS 2025 · similarity 0.89)
- Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · similarity 0.89)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.89)
- VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree (NeurIPS 2025 · 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.
