Agentic AI Improves Video Anomaly Detection

This research addresses the limitations of current Video Anomaly Detection (VAD) systems, which often struggle to provide both precise temporal localization ("when") and semantic understanding ("what") of anomalies. Inspired by human observation, the paper proposes a "Glance, Scrutinize, and Think" paradigm, moving VAD towards more sophisticated agentic reasoning. Initially, it introduces Glance then Scrutinize (GtS), a training-free framework that uses textual guidance for coarse-to-fine anomaly grounding, balancing accuracy and speed. Building on this, the paper further develops a tool-augmented agentic VAD method. This approach leverages a multimodal large language model (LLM) that learns to interact with tools, such as a video cropping tool, to inspect suspicious segments more densely. The agent can then iteratively self-correct mislocalized hypotheses through supervised fine-tuning and reinforcement learning. To support this, a new benchmark, VAGU-T, was created, featuring real-world videos with human-validated groundings, explanations, and tool-calling traces. Experiments show the agentic model delivers superior accuracy and faster inference.