Hello! I am a Ph.D. student in Computer Science at the University of California, Santa Cruz. I am grateful to be advised by Professor Leilani Gilpin in the Artificial Intelligence Explainability and Accountability (AIEA) Lab.
My research explores the development of more efficient and sustainable AI agents. With work at the intersection of reinforcement learning and large language models, I employ novel compression techniques to to extend AI efficiency gains into the agentic domain for solving sequential decision-making problems, namely Markov decision processes.
Outside of research, I love exploring the outdoors, primarily through cycling and rock climbing. Please don't hesitate to reach out if you are interested in collaborating!
In preparation for ICML 2026. My current work, in which I leverage reward signals to guide VLM and LLM agent pruning in MDPs using a novel REINFORCE-inspired pruning objective, extending LLM pruning to the online agentic domain for the first time.
Featured in COLM 2025. My internship project at LG AI. I developed a novel constraint-based LLM-as-a-judge approach that allowed for highly performant knowledge distillation of LLM agents for automated web browsing.
This DistilBERT-based ColBERT candidate generation model achieves top-k recall performance within 10% of the Mind2Web candidate generation baseline, but with 1/20th of the inference time.
This work develops a hierarchical GPT-4 agent that solves Minigrid tasks zero-shot by verbalizing subgoals, then applies a procedural data augmentation strategy over these subgoals to cheaply generate large amounts of data for imitation learning. This approach enables efficient knowledge distillation from the GPT-4 agent to an RL network to effectively bootstrap policies in sparse reward Minigrid environments.
This work is the first to adapt the decision transformer architecture to the online reinforcement learning setting. Our hybrid approach, in which an exploration policy in parallel for stability early in training, then increasingly using the decision transformer for exploration as it stabilizes, surpasses state-of-the-art RL algorithms in Mujoco.
Featured in the NeurIPS 2021 Workshop on Deep RL, this work proposes a actor-critic architecture which separates the policy and value functions during training. Our approach generalizes better than state-of-the-art RL methods while using fewer parameters.
UC Santa Cruz (2024-Present)
Ph.D. Student in Computer Science
Dr. Leilani Gilpin | AIEA Lab
University of Michigan (2022-2024)
M.S in Computer Science
Dr. Joyce Chai | SLED Lab
Kansas State University (2018-2022)
B.S in Computer Science, B.S in Mathematics
Dr. Bill Hsu | KDD Lab
My current work, in which I leverage reward signals to guide LLM agent pruning in MDPs using a novel REINFORCE-inspired pruning objective, extending LLM pruning to the online agentic domain for the first time.
Resulting from my internship project at LG AI, in which I developed a novel constraint-based LLM-as-a-judge approach for LLM agent knowledge distillation.