Video anomaly understanding (VAU) goes beyond flagging unusual activity. The goal is to generate language-level interpretations of anomalous events in long videos, which requires visual evidence that supports an explanation rather than a mere detection signal.

This arXiv paper focuses on the temporal dimension of that evidence. The authors address how to select the most relevant moments from a long video when building an explanation, a step they argue is underdeveloped in existing VAU pipelines.

The abstract is truncated in the source, so specific methods and results are not available here. What is clear is the paper's central concern: choosing the right evidence over time is essential for making anomaly explanations meaningful.