Research Interests

My research studies how LLM-based agents can sustain progress in complex settings that require continual exploration rather than fast convergence. I design systems that generate hypotheses, evaluate their own failures, and update their strategies at test time using programmatic feedback loops and structured evaluation.

I am particularly interested in:

• Scientific discovery, where agents iteratively refine hypotheses, design experiments or queries, and adapt search strategies under feasibility and budget constraints.
• Red teaming of LLM systems, where agents must diversify attacks, reason about defenses, and adapt to changing policies and evaluators while operating under strict query budgets.
• Open-ended behavior, where maintaining novelty, diversity, and adaptability over long horizons is essential, and where premature convergence and reward exploitation are core failure modes.