My research focuses on building multimodal models that can evaluate, learn from feedback, and continually improve themselves.

• Critic and reward modeling: I build and benchmark multimodal judges for reliable evaluation and alignment across diverse criteria and physical reasoning tasks (LLaVA-Critic, Multi-Crit, PhyCritic).
• Multimodal post-training: I study data selection and reinforcement learning for stronger multimodal reasoning and generation, with an emphasis on critic supervision and visual grounding (Self-Filter, LLaVA-Critic-R1, EVAN).
• Recursive self-improvement: I explore feedback loops that let models iteratively refine their outputs and discover better inference strategies, toward systems that can improve their own capabilities (RubSE, AutoTTS).