AnomalyMiner: Closed-Loop Evidence Mining for Aerial Anomaly Understanding

Aerial anomaly understanding requires models to identify, localize, and explain abnormal events in complex aerial videos. Existing methods based on fixed perception, visual enhancement, or global fusion often assume that providing more visual information yields more useful anomaly evidence. In practice, additional information may emphasize background structures, irrelevant thermal responses, or misleading regions rather than support the current anomaly judgment. This information–evidence gap is especially pronounced in low-light RGB–infrared scenes: RGB may miss subtle anomalous cues, whereas infrared may introduce salient but anomaly-irrelevant responses. The key challenge is therefore not simply to enhance, fuse, or acquire more visual inputs, but to continually admit, validate, and repair anomaly-relevant evidence throughout reasoning. Accordingly, we propose AnomalyMiner, a closed-loop evidence-mining framework for aerial anomaly understanding, which organizes perception and reasoning into a stateful evidence trajectory: perception supports the current anomaly hypothesis, while reasoning specifies what evidence should be sought next. Evidence-Gain Interleave first controls the forward flow of evidence by admitting localized infrared only when it yields counterfactual gain for downstream anomaly reasoning, rather than indiscriminately expanding multimodal context. However, admitted evidence can still update the reasoning state in a way that redirects subsequent perception toward misleading cues. Evidence-State Backtrack therefore monitors whether the current state regresses from the preceding state and repairs a misdirected search from that reliable preceding state. To make these admission and repair decisions directly learnable, we introduce Evidential Preference Policy Optimization, which routes process feedback to the corresponding key evidence-bearing tokens and action tokens for acquisition, interaction, and repair. AnomalyMiner achieves 74.75% AP on the complete A2Seek benchmark and 45.18% on its nighttime subset, while showing consistent cross-dataset generalization on M3OT and VTUAV.