Latent Spherical Flow Policy for Reinforcement Learning with Combinatorial Actions

This paper introduces a novel reinforcement learning framework designed for environments with combinatorial action spaces. The authors propose a latent spherical flow policy that leverages diffusion-based generative modeling techniques to efficiently represent and optimize complex action combinations. The approach improves exploration and policy learning in high-dimensional combinatorial settings, demonstrating strong empirical performance across benchmark reinforcement learning tasks. The work contributes new methods for scalable decision-making in domains where agents must select structured sets of actions simultaneously.