Video anomaly understanding stops relying on pre-sampled frames: SEEK-VAU lets an agent search and verify, lifting evidence F1 from 28.1 to 43.7
Synopsis
Video anomaly understanding has long used fixed observation: a model receives a bundle of sampled frames or clustered events and then judges the anomaly, with no way to go back for missing evidence once reasoning starts. SEEK-VAU recasts this as budget-bounded agentic search and verification, letting the policy gather evidence across the precursor, trigger and aftermath stages of an event chain and check sufficiency before closing the case. On its SEEK-Bench, built on a Qwen3-VL-8B backbone, it reaches 55.6 Temporal mIoU, 45.5 QA Accuracy and 43.7 Evidence F1, against 44.2, 37.7 and 28.1 for the fixed-observation Qwen3-VL baseline.
Interpretation
Video anomaly understanding moves from a one-shot judgment to multi-round evidence gathering: the model first scans the whole video coarsely, then issues retrieval requests for suspect intervals, commits the returned frames as supporting evidence, and checks whether that evidence is enough before concluding. Earlier pipelines separated evidence acquisition from reasoning, so the visual material was chosen before inference and stayed fixed; if reasoning exposed a missing stage, the model could not go looking for it. On the SEEK-Bench validation split with the same Qwen3-VL-8B backbone, fixed observation with basic supervision scored 44.2 Temporal mIoU, 37.7 QA Accuracy and 28.1 Evidence F1, while active search without event-chain supervision either raised these to 50.1, 40.1 and 34.3.
Four executable actions separate acquisition from closure: scan builds a whole-video temporal context, seek retrieves evidence for a concrete visual cue and a second-level interval, verify judges the evidence sufficient, insufficient or redundant, and finalize closes the case only after verification passes. The retrieval tool takes only the input video and the interval the policy issues, returning four uniformly spaced frames under a fixed rule, so what is learned is where to look rather than how to sample. An ablation shows that closed-loop search without verification reaches 29.1 Evidence F1, and adding the verification gate raises it to 41.6 while Temporal mIoU rises by 3.7 and QA Accuracy by 9.2.
Evidence-faithful counterfactual verification writes into the reward whether the selected evidence actually supports the conclusion: the evidence must suffice on its own, admit a compact sufficient subset, and lose confidence when trigger-stage evidence is removed. Rewarding answer correctness alone cannot tell whether evidence supports the conclusion, whether redundant windows hide weak support, or whether trigger evidence is decision-critical. Under the same closed-loop protocol, adding this reward raises Temporal mIoU, QA Accuracy and Evidence F1 by a further 1.2, 2.8 and 2.1, while average frame use falls from 17.1 to 15.6.
Perspective
The demonstrated scope is interactive search-and-verify evaluation on the SEEK-Bench validation split, plus zero-shot anomaly-localization transfer of frozen checkpoints to the UCF-Crime and XD-Violence test splits; the latter measures localization only, not external evidence faithfulness, because those datasets lack aligned event-chain, evidence-window and semantic-QA annotations.
The verifier and the policy share the same Qwen3-VL-8B weights, so the residual risk is self-confirmation, and the authors frame it as an empirically audited self-consistency mechanism rather than a formal guarantee; the benchmark is derived from MSAD and ECVA and inherits category coverage, annotation noise and dataset bias; evaluation is limited to sparse-anomaly videos where the anomaly occupies a short fraction, and very long videos with an extremely low anomaly fraction need dedicated evaluation.