Hermes: An Evidence-Driven Agentic Framework for Trustworthy and Explainable AI-Generated Video Detection

Recent advances in generative video models have blurred the boundary between real and synthetic content, raising urgent concerns about digital authenticity. Multimodal large language models (MLLMs) are promising for AI-generated video (AIGV) detection due to their broad perceptual and reasoning capabilities; however, existing MLLM-based detectors remain prone to hallucinated evidence and unstable reasoning, leading to false alarms and generic, unverifiable explanations. To address these issues, we propose Hermes, an evidence-driven agentic framework for trustworthy and explainable AIGV detection. Hermes comprises three key components: (1) Adaptive Instance-Conditioned Detection Strategy Planning, (2) Evidence-Centric Reasoning and Verification, and (3) Graph-Grounded Evidence Deliberation. Specifically, Hermes analyzes each video and uses instance-conditioned retrieval-augmented generation to retrieve relevant forensic knowledge and compose a tailored detection strategy. It then constructs a verifiable Evidence Reasoning Graph (ERG) to keep the reasoning grounded in concrete video evidence. Finally, multi-agent deliberation audits and refines the ERG to reconcile conflicting evidence and improve the reliability of the final judgment. Together with a library of forensic tools, these components enable evidence-grounded authenticity judgments and structured, verifiable explanations. Extensive experiments show that Hermes, without task-specific training, achieves state-of-the-art performance and produces auditable explanations.