Jiajun Fan

Jiajun Fan

Reinforcement Learning Researcher

CS Ph.D. of UIUC

Welcome

I’m Jiajun Fan, a Computer Science Ph.D. student at the University of Illinois Urbana-Champaign, working at the intersection of reinforcement learning theory and large-scale AI systems. My research focuses on developing self-evolving AI systems that can learn continuously from human/AI feedback while maintaining reliability and sample efficiency.

Currently seeking research internship opportunities for Summer 2025. My CV

Selected Publications

My research has been published at top venues including ICLR, ICML, and NeurIPS. Recent highlights:

  1. Online Reward-Weighted Fine-Tuning of Flow Matching with Wasserstein Regularization
    ICLR 2025
    For the first time, we proposed a framework for online post-training with RL for the flow matching model from both theoretical and practical perspectives (Collapse-Free Continuous Self-Learning, No Human-Collected Data Needed, Self-Evolution by Agent itself).

  2. Learnable Behavior Control: Breaking Atari Human World Records
    ICLR 2023 (Oral Presentation)
    We demonstrate for the first time the dramatic advantages of MoE in RL and break 24 human world records (100x cost reduction, Stable Self-Evolution of MoE in RL)

  3. Generalized Data Distribution Iteration
    ICML 2022
    For the first time, we theoretically demonstrate the importance of data optimization in RL. (Theoretical guarantee for continuous and stable self-evolution of MoE and single agent in RL)

Research Focus

My work addresses fundamental challenges in AI development through three key areas:

  • Self-Evolving AI Systems: Creating AI systems that can continuously improve through online learning while preventing collapse and maintaining diversity
  • Data-Efficient Learning: Developing algorithms that achieve superhuman performance with orders of magnitude less data
  • Theoretical Foundations: Building rigorous mathematical frameworks for understanding and improving AI learning processes

Research Highlights

  • Flow Matching with Wasserstein Regularization: Developed the first theoretically-grounded framework for continuous model evolution with provable diversity guarantees, reducing data requirements by orders of magnitude [ICLR 2025]
  • Learnable Behavior Control: Created a unified framework that broke 24 Atari world records while using 500x less data than previous methods [ICLR 2023 Oral, top 5/4176]
  • Generalized Data Distribution: Pioneered a new reinforcement learning paradigm that achieves state-of-the-art performance through optimized data distribution [ICML 2022]

Latest News

  • [2025-02] New paper on self-evolving Flow Matching Generative Models accepted to ICLR 2025
  • [2025-01] Selected as reviewer for ICML 2025, ICLR 2025, and NeurIPS 2024
  • [2024-08] Started research on collapse-free self-evolution at UIUC

Contact

I’m always interested in discussing research ideas and potential collaborations:

“The goal is not just to build better AI, but to understand intelligence itself”