Abstract
Video Anomaly Detection (VAD) aims to temporally localize abnormal events in videos. Existing approaches, whether explainable or not, rely heavily on training, fine-tuning, dataset-specific hyperparameters, and curated data, which limits their generalization to real-world scenarios. Although recent zero-shot methods mitigate this dependency by leveraging vision-language models (VLMs) for training-free inference, they typically depend on shallow captioning or similarity matching, lacking deep reasoning and temporal awareness. To overcome these challenges, we propose CoZSR-VAD, a zero-shot, reasoning-driven, two-stage framework built entirely upon a Large Vision-Language Model (LVLM). In the raw stage, a Chain-of-Anomaly Detection Thought Prompting (CoADTP) strategy integrates an anomaly text database with temporal cues, enabling contextual reasoning and generating coarse anomaly scores. In the refinement stage, a cross-modal re-ranking mechanism fuses visual and linguistic embeddings to inject semantic and temporal coherence, resulting in refined and temporally consistent predictions. Comprehensive experiments on two large-scale benchmarks demonstrate that CoZSR-VAD achieves performance comparable to baselines without any training or fine-tuning, highlighting its strong reasoning capability, generalization, and practical applicability in real-world anomaly detection scenarios. Code will be made publicly available at: https://github.com/MohdUbaidwani/CoZSR-VAD-Contextual-Zero-Shot-Reasoning-VAD.