DiPRL: Learning Discrete Programmatic Policies via Architecture Entropy Regularization

In three linesDiPRL introduces a programmatic reinforcement learning method that learns discrete, interpretable policies without post-hoc discretization. Using architecture entropy regularization, the approach converges toward discrete programs during training, avoiding performance collapse and eliminating the need for additional fine-tuning.

Summary generated by Claude — human-verified