欺骗即教学？通过对抗强化学习锻造感知鲁棒性

Multimodal Large Language Models (MLLMs) exhibit perceptual fragility when faced with visually complex scenes, primarily due to their reliance on finite training datasets that are costly to scale and limit model robustness. AOT-SFT, a large-scale adversarial dataset, is introduced to bootstrap MLLM robustness. Building on this, AOT (Adversarial Opponent Training) is proposed as a self-play framework that enhances MLLM robustness by generating its own training data. This method facilitates a co-evolution between an image-editing Attacker and a Defender MLLM, where the Attacker creates a diverse and dynamic curriculum of image manipulations, compelling the model to continuously adapt and improve.