Reinforcement learning has become a central paradigm for extending large language model (LLM) capabilities, enabling agentic behavior, tool use, and policy-level control. Yet RL fine-tuning often triggers catastrophic forgetting, trading new gains in downstream performance for losses in core pretrained skills. We aim to develop an optimization-driven, capability-retaining RL method that raises downstream performance while explicitly preserving original knowledge.

My interests sit at the intersection of deep learning theory and application. SuperUROP lets me pursue that interest and helps me conduct research that can translate theoretical ideas into deployment-ready impact.