arxivcs.LG2026-07-16
A Continuous-Time Reinforcement Learning Framework for Fine-Tuning Discrete Diffusion Models
Zikun Zhang, Jiayuan Sheng, David D. Yao, Wenpin Tang
We formulate reinforcement learning (RL) in continuous time with discrete state spaces and possibly arbitrary action spaces via a stochastic control approach, where the state dynamics are modeled as a controlled continuous-time Markov chain (CTMC). We consider policy optimization…