CRIME: Investigating Compositional Risks in Video Moderation via Relational Evidence

Compositional risks arise when the relations among video elements establish a risk that the individual elements alone do not. Despite advances in multimodal large language models, detecting such risks remains difficult: relevant cues may be dispersed amid abundant video content, and recognizing them individually does not establish how they are related. The central challenge is to identify which connections matter to the risk being assessed, whether those connections hold, and whether they support a risk judgment. To tackle this challenge, we introduce CRIME (Compositional Risk Investigation for Video Moderation through Relational Evidence), an investigative framework for discovering and verifying relational evidence in videos. CRIME maintains a revisable Evidence Graph of video-grounded observations and candidate relations. Guided by evolving risk hypotheses, it searches for missing or uncertain evidence, checks consequential connections against the source video, and revises the graph and hypotheses accordingly. Multiple agents take distinct but interacting roles in proposing interpretations, seeking evidence, and scrutinizing candidate relations. Through the shared graph, one agent's provisional findings can be examined and corrected by others, allowing those corrections to shape subsequent investigation and the final judgment. Experiments across controlled synthetic, real-world, and cross-policy compositional-risk settings demonstrate the effectiveness and adaptability of CRIME. These results highlight the value of combining explicit relational representation with hypothesis-driven evidence discovery for video moderation.