gain-rl-angle-concentration

Improve RL training efficiency by 2.5× using angle concentration between token hidden states as a cost-effective data scheduling signal, selecting high-gradient samples dynamically.

Independent GitHub snapshot · Automated classification · Original documentation governs