EvoGuard: A scalable agent reinforcement learning framework for practical and continuously evolving AI-generated image detection

EvoGuard proposes a scalable agent reinforcement learning framework that turns AI-generated image detection into reasoning-based evidence synthesis over the outputs of multiple off-the-shelf detectors. Its capability-aware selection and dynamic orchestration mechanisms cross-validate conflicting or low-confidence signals over multiple rounds, and it is trained with GRPO using only low-cost binary labels; experiments report that it outperforms single detectors and static ensembles, achieving SOTA accuracy. New detectors can be plugged in without training, to cope with continuously evolving AI-generated images.