Research

Interest

I study world models that faithfully simulate how environments evolve in response to actions. These models can serve as substitutes for their target environments, generating interaction data to support decision-making methods such as reinforcement learning and planning. I am also interested in broader computer vision such as generative modeling and unsupervised representational learning.

Goal

I want to build world models that:

  1. perfectly simulate external environments, such as the video game World of Warcraft;

  2. generalize across diverse tasks and environments, eliminating the need to train a separate world model for each task; and

  3. capture the underlying concepts and rules of the real world, such as the law of gravity.

Selected publications

  1. Concept-Guided Spatial Regularization for World Models in Atari Pong Yukuan Lu, Zaishuo Xia, Weyl Lu, and Yubei Chen.
    arXiv preprint, 2026.

    An original and self-led first research project. This work shows that:

    • As of 2026, successful world-model agents—policies trained with a world model—frequently rely on

      weak underlying world models, a phenomenon I term the agent–world-model gap.

    • Our method improves world models in Atari Pong, while broader generalization remains open.

Other resources

Talks

  1. A Brief Overview of Code World Models

    UC Davis SSL Lab · August 14, 2026

    Slides · Notes

0%