I'm a postdoctoral researcher at UCSD and the CMU/MBZUAI CLeaR group, working with Kun Zhang and Biwei Huang. Previously I worked closely with Sara Magliacane at the University of Amsterdam. I completed my PhD at City University of Hong Kong.
Feel free to contact me for research collaborations or other engagements.
Research Interests
My long-term research goal is to build agents that not only imagine the world, but also understand it, act in it, discover goals within it, and continually refine their internal models in an open-ended, self-improving loop. To achieve this, I study world models, RL, and robotics. Specifically, I focus on:
- Interactive and Adaptive World Models: How can we learn world models from large-scale interactive data that are sufficient for control and planning (TC-WM), while remaining robust under broad distribution shifts across diverse and non-stationary environments and tasks (AdaRL, Factored-RL, Ada-Diffuser), across embodiments (SCAR), and in continuously evolving domains (CSR)?
- Compositional and Generalizable Policy Learning: How can agents compositionally reuse learned structures to generalize across new tasks and environments through world models (FIOC-WM, DAFT-WM)? Furthermore, how can agents autonomously discover and acquire meaningful skills through self-play in rich interactive environments by leveraging interactive memory and unsupervised RL (IWR, HInt)?
- Agentic Exploration and Self-Improvement: How can agents actively explore in a purposeful and open-ended way to discover goals and achieve them (ECL, CIP, DreamSAC), and simultaneously improve their world models (WAV)?
Publications
* equal contribution, † equal advising
Preprint
World Action Verifier: Self-Improving World Models via Asymmetric Forward-Inverse Consistency →
ICLR World Model Workshop & RSI Workshop 2026
★ Outstanding paper at ICLR World Model Workshop, 2026
★ ICLR 2026 RSI Workshop Spotlight
2026
Learning Task-Sufficient World Models by Synergizing Agentic Exploration and Structured Modeling →
ICML 2026
Also at World Modelling Workshop, MILA, 2026
2025
2023
Learning Dynamic Attribute-factored World Models for Efficient Multi-object Reinforcement Learning →
NeurIPS 2023
Community Channel-Net: Efficient Channel-wise Interactions via Community Graph Topology →
Pattern Recognition, 2023
