Improving Math Reasoning through Value-Guided Informative Search
Turning value-guided search improvements into informative learning signals for mathematical reasoning with verifiable rewards.
Publications
You can also find my work on Google Scholar.
2026
Iterative Gumbel Planning for Continuous Control
Aligning search and policy representations for continuous control, addressing variance inflation through iterative Gumbel planning with factorized categorical policies.
Scaling Tasks, Not Samples: Mastering Humanoid Control through Multi-Task Model-Based Reinforcement Learning
Scaling tasks rather than samples: shared world models enable sample-efficient multi-task learning for humanoid control.
2024
Real-time power scheduling through reinforcement learning from demonstrations
Combining expert demonstrations and reinforcement learning for real-time power system scheduling.
EfficientZero V2: Mastering Discrete and Continuous Control with Limited Data
A unified framework for sample-efficient reinforcement learning across discrete and continuous actions, visual inputs, and low-dimensional states.
2023
SpeedyZero: Mastering Atari Games with Limited Data and Time
Fast distributed model-based reinforcement learning, reaching human-level Atari performance in 35 minutes with 300k samples.
2021
Mastering Atari Games with Limited Data
EfficientZero learns a world model for planning and achieves super-human performance on Atari 100k with only two hours of game experience.