I am a first-year M.S. student in Computer Science at Brown University, working on Natural Language Processing (NLP) and Large Language Models (LLMs). My research interests lie in LLM post-training and interpretability, with a focus on understanding, adapting, and controlling the internal mechanisms that shape large language model behavior after pretraining.
I am particularly interested in how internal representations, contextual information, and intermediate reasoning processes give rise to model behavior, and how these insights can inform post-training and inference-time control. This includes activation steering, retrieval-augmented generation (RAG) and reinforcement learning methods for improving the generalization, reliability, and safety of LLMs.
I am currently a research intern at UCLA NLP in Summer 2026. I also work closely with Prof. Kuan-Hao Huang at Texas A&M University. Previously, I received my Bachelor's degree in Computer Engineering with First Class Honors from Nanyang Technological University, where I was fortunate to be advised by Prof. Wenya Wang. I am broadly open to collaborations on understanding and improving large language models, as well as other exciting problems in NLP and AI.
We propose In-Context Routing, an implicit in-context learning method that captures reusable ICL routing structure and modulates attention logits to simulate few-shot behavior without explicit demonstrations at inference time.
We study structure-aware exemplar selection for in-context learning in complex semantic parsing tasks, and introduce a structural alignment perspective for selecting more effective demonstrations.
We introduce TRACES, a proactive safety auditing framework that models the evolving trajectory state of multi-turn LLM agents to detect unsafe behavior before task completion.
We introduce Steering Vector Fields, a context-aware framework for inference-time control that adapts steering directions according to local representation geometry.
We introduce In-Context Prompt Tuning, an efficient framework for personalizing large vision-language models by transforming fine-grained visual semantics from reference images into adaptive continuous prompts.