3 ms·Beyond 80/20: High-Entropy Minority Tokens Drive Effective RL for LLM Reasoning3 points by mdp2021 6mo ago