Hang ChenHáng Chén

My research encompasses a broad spectrum of language models, centered on how these models construct and utilize causal mechanisms. In my earlier work, I investigated methods to endow LLM representations with causal discrimination and explored the phenomenon of causal emergence within these complex architectures. Currently, my focus has shifted toward Mechanistic Interpretability (MI). I am particularly interested in the intersection of MI and parameter updating (such as SFT). My goal is to leverage mechanistic insights—identifying specific functional circuits—to guide more precise, surgical, and interpretable modifications to model behavior. By bridging these two fields, I aim to transform LLMs from “black boxes” into transparent systems that can be reliably controlled and updated for trusted applications. 😃