Selective-Advantage Entropy-Adaptive Horizon GRPO: Asymmetric Token-Level Discounting for Efficient Reinforcement Learning of Language Models

In three linesSA-AH-GRPO, an extension of GRPO, applies asymmetric entropy-based token discounting for LLM reinforcement learning. On GSM8K, Qwen 2.5-3B achieves Pass@1=0.858 with 3.6× variance reduction vs standard GRPO while preserving gradients on correct trajectories.

Summary generated by Claude — human-verified