Zhenting Qi

Zhenting Qi

Ph.D. Student in Computer Science

Harvard University

About

Welcome! I am a 2nd-year Computer Science Ph.D. student at Harvard University, working with Prof. Yilun Du, Prof. David Parkes, and Prof. Hima Lakkaraju. Before joining Harvard, I obtained my bachelor's degree from Zhejiang University. I have also spent time at FAIR at Meta, Google DeepMind, and the MIT-IBM Watson AI Lab.

Research

My research studies how to design the mechanisms under which AI systems self-evolve, from search and evolution within a single agent to coordination across a collective of agents.

  • Within an agent, these mechanisms take the form of search and evolution over the agent's own attempts, letting it reach solutions it would rarely produce by sampling alone (rStar, Satori, BES).
  • Across agents, they become rules of interaction: in Economy of Minds, agents bid for the right to act and pay those whose work they build on, and coordinated strategies emerge that no one programmed.

Yet improvement matters only if it transfers, and gains that rest on memorization or self-reinforcement often do not (Scylla, EvoLM, Generalization Gap). I therefore aim for systems that learn modularity, reusability, and composability, forming structures such as specialized agents and skills that can be reused and recombined, with robustness to local perturbations and generality to out-of-distribution scenarios.

For more information about my research, please see Google Scholar, Semantic Scholar, or DBLP.

News

  1. Our paper Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions has been accepted to NeurIPS 2026.

  2. Recognized as a Silver Reviewer at ICML 2026.

  3. Joining FAIR at Meta as a Research Scientist Intern this summer, working on scalable multi-agent coordination.

  4. Our paper On the Generalization Gap in Self-Evolving Language Model Reasoning, from my internship at Google DeepMind, has been accepted to ICML 2026.

  5. Received an Honorable Mention for the Jane Street Graduate Research Fellowship.

  6. Released Confucius Code Agent: Scalable Agent Scaffolding for Real-World Codebases, a software engineering agent for large-scale codebases.

  7. Our paper EvoLM: In Search of Lost Language Model Training Dynamics has been accepted to NeurIPS 2025 (oral).

  8. Our paper Satori: Reinforcement Learning with Chain-of-Action-Thought Enhances LLM Reasoning via Autoregressive Search has been accepted to ICML 2025.

  9. Will be joining Google DeepMind (Mountain View office) as a Student Researcher, working on language model post-training.

  10. I will continue my research journey at Harvard as a PhD student!

  11. Our papers Mutual Reasoning Makes Smaller LLMs Stronger Problem-Solvers, Quantifying Generalization Complexity for Large Language Models, Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems have been accepted to ICLR 2025.

Publications

* equal contribution

2026

2025