Home
I'm a researcher working on machine learning and AI. Previously I was a research scientist at Google DeepMind, where my core work was the design of objective functions for improving the training and inference efficiency of large language models.
Research
World understanding with human common sense and intuition (2026—present)
How come humans learn a new concept with two examples, while an RL agent needs thousands of training steps?
This is my new research interest. How can we use human commonsense knowledge to accelerate AI training? How to design an AI agent that learns new concepts and acquire skills in the wild?
Model routing for efficient inference (2022—2025)
With many models in the serving pool, can each query be sent to the right one?
Use small models wherever they suffice and reserve large models for the few genuinely hard queries — cutting inference cost without sacrificing answers.
Teacher-based LLM training (2022—2025)
Can an already-trained model help train another one?
Training a small model only on the tokens a large teacher deems easy; and using a small teacher to help train a larger student.
Model evaluation & model comparison (2015—2022)
From samples alone, are two distributions the same?
Sample-efficient, kernel-based statistical tests for comparing two distributions, comparing a sample against a known model, and deciding which of several candidate models best fits the data.
Selected Recognition
Experience
Education
Patents
Academic Service
Co-organizer
Last updated: 2026-06-30