Learning to Watch: Active Video Anomaly Understanding via Interleaved Policy Optimization

Video anomaly understanding (VAU) relies on sparse, context-dependent cues. However, existing passive paradigms suffer from observational aliasing, where static sampling fails to disambiguate semantically distinct events. To overcome this, we propose $Anom\text{-}\pi$, a closed-loop framework that reconceptualizes video understanding as an active sequential decision-making process within a dynamic environment. Inspired by human video-reviewing behavior, this framework unifies internal cognitive reasoning and strategic evidence acquisition into an interleaved policy, utilizing temporal atomic operators such as local backtracking, temporal expansion, and fine-grained sampling to endow the model with perceptual proactivity. To learn such complex interaction strategies under video-level weak supervision, we design Interactive Direct Preference Optimization (iDPO) to achieve trajectory-level policy alignment, guided by an Active Evidence Inquiry (AEI) utility that balances task success, informative evidence acquisition, and interaction cost. This approach enables the agent to learn to actively disambiguate hypotheses while suppressing redundant exploration. Extensive experiments demonstrate that our framework, with only 2B parameters, achieves highly competitive performance, significantly outperforming state-of-the-art large-scale VAU models in complex scenarios.

Long videos often contain only a few moments that explain what is really happening. In safety-related videos, for example, a person may seem to be standing normally until a brief frame reveals a fall, a fight, or another unusual event. Many AI systems make their decision after looking at a fixed set of snapshots. If those snapshots miss the important moment, the system has little chance to correct itself. Our work builds an AI video reviewer, Anom-pi, that can decide when it needs to look again. It can zoom in on selected moments, check what happened just before or after them, and inspect suspicious parts of the video more carefully. It also learns not to keep looking forever: once it has enough evidence, it stops and gives a clear answer. Across several public video collections, this active reviewing strategy helps the system make better decisions in difficult cases while avoiding unnecessary extra viewing.