Online video anomaly detection (VAD) necessitates real‐time inference without access to future temporal context. Existing training‐free methods, while promising, often suffer from correlation bias, misclassifying infrequent but benign actions (e.g., running for exercise) as anomalies due to a lack of causal reasoning. In this paper, we introduce multi‐agent counterfactual dialectics (MACD), a novel framework that reframes VAD from statistical outlier detection to causal falsification. MACD operationalizes this via a structured adversarial debate among three specialized agents: a Prosecutor for high‐recall risk screening, a Skeptic that performs instruction‐conditioned causal interventions to generate benign counterfactual explanations, and an Arbitrator that adjudicates the final verdict based on the “evidence gap.” To ground this reasoning in visual reality, we construct a visual‐semantic world model using dynamic scene graphs (DSG) and set‐of‐mark (SoM) prompting, reinforced by a visual re‐verification loop (VRL) and a contextual memory bank to mitigate hallucinations. Our “screen‐debate‐decide” protocol ensures computational efficiency by reserving expensive dialectics only for ambiguous frames. Extensive experiments on UCF‐Crime and XD‐Violence demonstrate that MACD achieves state‐of‐the‐art zero‐shot performance.
Ouyang et al. (Thu,) studied this question.