About

I am on the industry job market. Please get in touch about full-time opportunities.

I am a PhD candidate at the University of Cambridge’s Language Technology Lab, supervised by Nigel Collier. I expect to complete my PhD in January 2027.

I recently interned at Apple, working on LLM post-training for complex instruction following with Maartje ter Hoeve, Rin Metcalf, and Andrew Silva. I have also been a visiting researcher at Bocconi University, working with Dirk Hovy and Paul Röttger.

I co-authored the Current Capabilities chapter of the International AI Safety Report 2026, covering frontier LLMs and agentic systems. The report was chaired by Yoshua Bengio.

Before Cambridge, I received my MSc from ETH Zürich, with a thesis at EPFL, and my BSc from the University of Texas at Dallas. I’m from Nanjing. Outside research, I enjoy hiking and cooking, and being near mountains, lakes, or the ocean.

Research

In my PhD, I study how to build AI systems that understand people, behave as intended, and work productively within human society. My research connects three questions: how AI models people, how we shape its behavior, and what happens when it interacts with us.

Modeling people. General-purpose AI needs to work with people whose preferences, beliefs, and goals differ. I study how language models can capture human variation, simulate individuals and populations, and adapt to particular users. This work supports both the study of human behavior and the development of personalized agents.

Shaping AI behavior. Understanding people informs what AI should do, but making models behave accordingly remains a separate challenge. I study behavioral steering and how post-training affects alignment and calibration. My position paper argues that failures from hallucination to homogenization share a common problem: model behavior is miscalibrated relative to the uncertainty and variation it should reflect.

AI in society. I examine how these behaviors play out in interaction, from when agents should ask for clarification to social biases, persuasion, and creativity in AI teams. These settings reveal both opportunities for collaboration and limitations that evaluations of models in isolation can miss. I also contribute to broader assessments of AI capabilities and risks through the International AI Safety Report.

My full publication list is on Google Scholar.

Service

I review for NeurIPS, ICML, ICLR, ACL ARR, CHI, ICWSM, LREC-COLING, and JMIR. I received an Outstanding Reviewer award at EMNLP 2024 and Special Recognition for Outstanding Review at CHI 2024.

  • Co-organizer, Cambridge Language Technology Lab Seminar.
  • Expert section writer, International AI Safety Report 2026 (August–December 2025).
  • Panelist, Widening Natural Language Processing Workshop at EMNLP 2023.

Invited talks

  • From Individuals to Groups to Interactions: What Is Missing in LLM Simulations?
    Stanford University, 2026.
  • LLM Social Simulation Is a Promising Research Method.
    Cambridge Judge Business School, 2025.