👋 Hi there! I'm Pin-Jie, and I'm a PhD student at the LLMs Lab at Virginia Tech. Personal Weibsite / X / Google Scholar
My research focuses on building capable LLMs and agent systems with less data and minimal supervision, through synthetic data, fine-tuning transfer, and PEFT. I explore methods that improve the efficiency of model updating (e.g., SFT, DPO, RLVR) while effectively blending diverse skills, including safety, tool use, coding, multilingual, and instruction-following. My overarching goal is to build modular, reusable systems in which the intelligent system continually learns and evolves over time. Currently, it covers the areas:
-
Efficient and continual model development. I explore strategies to make LLM updates faster and more reusable. In particular, I study how alignment capabilities can be transferred through model weights, enabling reuse across distinct model versions and eliciting the CoT behavior in latent space (e.g., fine-tuning transfer with diff vectors; Master Key Hypothesis).
-
Continually evolving agents. Building LLM agents that evolve skills and effectively extract useful signals from raw trajectories for long-horizon tasks.
-
Model harnessing. Designing systems that effectively orchestrate LLM agents with tools, memory, and feedback to improve system reasoning.