Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning

Wang, Shenzhi, Yu, Le, Gao, Chang · arXiv · 2025