Yan Zhou (Terry)

Hello! I am a Master’s student in Data Science at Harvard University. I am currently part of the CRISP group, working under the supervision of Prof. Demba Ba on my master’s thesis.

I am primarily interested in mechanistic interpretability. In particular, my research focuses on understanding models’ internal computations through analyzing weight-space and activation-space structures.

I have also been working on knowledge distillation and model arithmetic with Prof. David Alvarez-Melis in the ML Foundations group, and on AI safety and interpretability in the AI4LIFE group with Prof. Hima Lakkaraju. Previously, I completed my Bachelor’s degree in Politics and Data Science at the London School of Economics.

I am always happy to discuss research ideas! Please feel free to reach out to me at terryzhou [at] fas [dot] harvard [dot] edu.

News

Aug 20, 2026	Our work Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs is accepted to EMNLP 2026 Findings!
Jun 27, 2026	I’ll be presenting our work User Persona Subspaces Modulate Refusal Behavior in Language Models at the Mech Interp Workshop at ICML 2026!

Selected Publications

EMNLP — Findings

Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs

Yan Zhou, Sara Kangaslahti, Jonathan Geuter, Nihal V. Nayak, Marco Fumero, Francesco Locatello, and David Alvarez-Melis

In Findings of the Association for Computational Linguistics: EMNLP, 2026

URL PDF Bib

@inproceedings{zhou2026adapt,
  title = {Thinking at the Right Size: Amortized Distillation Across Post-Trained LLMs},
  author = {Zhou, Yan and Kangaslahti, Sara and Geuter, Jonathan and Nayak, Nihal V. and Fumero, Marco and Locatello, Francesco and Alvarez-Melis, David},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP},
  year = {2026},
}

ICML — Workshop

User Persona Subspaces Modulate Refusal Behavior in Language Models

Yan Zhou, Shichang Zhang, Zidi Xiong, and Himabindu Lakkaraju

ICML Workshop on Mechanistic Interpretability, 2026

URL PDF Bib

@article{zhou2026user,
  title = {User Persona Subspaces Modulate Refusal Behavior in Language Models},
  author = {Zhou, Yan and Zhang, Shichang and Xiong, Zidi and Lakkaraju, Himabindu},
  journal = {ICML Workshop on Mechanistic Interpretability},
  year = {2026},
}