research
Research on LLM post-training, reasoning, agents, and efficient LLM systems.
LLM Post-Training
I work with Prof. Hao Peng on multi-task reinforcement learning and sequential on-policy distillation for LLM post-training. The project studies how task order and rollout-state distributions relate to transfer and interference across tasks, and how parameter-update geometry helps explain differences in long-horizon learning.
LLM Agents and Combinatorial Optimization
With Prof. Hua Xu, I studied LLM agents for large-scale mixed-integer linear programming and combinatorial optimization. The work explores code-generating heuristics, tool-using optimization pipelines, and structured interfaces between solver states and LLM token spaces.