Discrete Flow Matching Policy Optimization
Apr 7, 2026·
,,,,,·
1 min read
M. Su
Po-Chung (Bryan) Hsieh
W. Wu
M. Lu
J. Chen
J. Y. C. Hu
H. Liu
DoMinO is a reinforcement learning framework for fine-tuning discrete flow-matching models. It formulates discrete flow-matching inference as an inner Markov decision process, which enables exact trajectory log-likelihood optimization. Fine-tuning uses REINFORCE- and PPO-based methods with cross-entropy and generalized KL regularization to mitigate reward hacking and preserve sequence naturalness.
On regulatory DNA and protein design, DoMinO outperforms prior state-of-the-art baselines by 10.6% on DNA predicted activity and 15.4% on protein pred-ddG.
My role: project co-lead. See the DoMinO project page.
Authors
M.S. Student in Electrical and Computer Engineering
I am an M.S. student in Electrical and Computer Engineering at UC San Diego, with a B.S. in Electrical Engineering from National Taiwan University. I build AI agents and reinforcement learning methods for scientific discovery, including multi-agent systems for biomedical feature discovery and policy optimization for discrete generative models, and I previously worked as an LLM engineer intern at the Inventec Artificial Intelligence Center.