Lei Feng 冯磊

I study agents that plan over long interactions, choose tools and reusable skills, and learn from real-world trajectories. My interests include decision-oriented world models, conversational interaction with controllable user simulators, and memory systems that retain experience and retrieve relevant evidence for future tasks.

I develop reinforcement learning methods for agents facing sparse rewards and long decision sequences. My work focuses on assigning credit to useful actions, accounting for historical context and execution dependencies, stabilizing policy optimization, and learning from the joint decisions of collaborating agents.

I study the security, reliability, robustness, and privacy of AI agents and language and vision-language models. My work examines backdoors in multimodal alignment and compositional understanding, and develops defenses that preserve useful capabilities. I also investigate adversarial robustness and privacy risks.