high-entropy-minority-tokens-rl

Optimize only high-entropy tokens during RL training to achieve better reasoning performance with 80% fewer gradient updates.

Independent GitHub snapshot · Automated classification · Original documentation governs