HiProbe-VAD: Video Anomaly Detection via Hidden States Probing in Tuning-Free Multimodal LLMs

2025-07-23

AI Summary

Video anomaly detection (VAD) faces critical bottlenecks including high computational overhead and heavy reliance on large-scale annotated data. To address these, we propose HiProbe-VAD—a training-free, fine-tuning-free VAD framework that leverages frozen multimodal large language models (MLLMs). We make the first discovery that intermediate hidden-layer states of MLLMs exhibit high sensitivity and linear separability with respect to anomalous events. Building on this insight, we design a Dynamic Layer Significance Probing (DLSP) mechanism to adaptively identify the most discriminative hidden layer. HiProbe-VAD further integrates a lightweight anomaly scorer and a temporal localization module, operating directly on frozen MLLM features. The method achieves strong generalization and inherent interpretability without any parameter updates. Extensive experiments demonstrate that HiProbe-VAD significantly outperforms existing training-free approaches and surpasses most supervised and weakly supervised methods on UCF-Crime and XD-Violence, establishing a new paradigm for efficient, low-barrier VAD.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Data Mining & Knowledge Management: Anomaly/Outlier Detection

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Video Anomaly Detection (VAD) aims to identify and locate deviations from normal patterns in video sequences. Traditional methods often struggle with substantial computational demands and a reliance on extensive labeled datasets, thereby restricting their practical applicability. To address these constraints, we propose HiProbe-VAD, a novel framework that leverages pre-trained Multimodal Large Language Models (MLLMs) for VAD without requiring fine-tuning. In this paper, we discover that the intermediate hidden states of MLLMs contain information-rich representations, exhibiting higher sensitivity and linear separability for anomalies compared to the output layer. To capitalize on this, we propose a Dynamic Layer Saliency Probing (DLSP) mechanism that intelligently identifies and extracts the most informative hidden states from the optimal intermediate layer during the MLLMs reasoning. Then a lightweight anomaly scorer and temporal localization module efficiently detects anomalies using these extracted hidden states and finally generate explanations. Experiments on the UCF-Crime and XD-Violence datasets demonstrate that HiProbe-VAD outperforms existing training-free and most traditional approaches. Furthermore, our framework exhibits remarkable cross-model generalization capabilities in different MLLMs without any tuning, unlocking the potential of pre-trained MLLMs for video anomaly detection and paving the way for more practical and scalable solutions.

Problem

• Detects video anomalies without fine-tuning MLLMs
• Reduces computational demands and labeled data reliance
• Improves anomaly detection via hidden states analysis

Innovation

• Leverages pre-trained MLLMs without fine-tuning
• Dynamic Layer Saliency Probing extracts hidden states
• Lightweight scorer detects anomalies efficiently