中文CoReVAD：一种用于免训练视频异常检测的上下文推理框架

现有视频异常检测方法依赖特定训练，域依赖强且成本高，且仅输出异常分数，缺乏可解释性。新方法利用视觉语言模型同时实现异常检测与可解释推理，减少域依赖，降低训练成本，提供人类可理解的异常原因。

arXiv:2605.23116v1 Announce Type: new Abstract: Existing Video Anomaly Detection (VAD) methods typically rely on task-specific training, leading to strong domain dependency and high training costs. Moreover, most existing methods output only scalar anomaly scores, providing limited insight into why specific events are considered abnormal. Recent advances in Vision-Language Models (VLMs) have enabled both anomaly detection and human-interpretable reasoning. However, many VLM-based approaches still require additional training steps (e.g., instruction tuning or verbalized learning) or external La