To Deceive is to Teach? Forging Perceptual Robustness via Adversarial Reinforcement Learning

Researchers introduce AOT (Adversarial Opponent Training), a self-play framework that improves Multimodal Large Language Models' robustness by having an AI attacker generate adversarial image manipulations to train a defender model. The method addresses perceptual fragility in MLLMs when processing visually complex scenes, reducing hallucinations through dynamic adversarial training.

• →MLLMs show perceptual fragility with visually complex scenes due to finite training datasets.
• →AOT-SFT provides a large-scale adversarial dataset for bootstrapping MLLM robustness.
• →The AOT framework uses co-evolution between an image-editing Attacker and Defender MLLM.
• →Extensive experiments show AOT enhances perceptual robustness and reduces hallucinations.
• →The method establishes a scalable paradigm for training more reliable multimodal AI systems.