rlad-knowledge-distillation

RLAD performs selective distillation during RL, using trust-region ratio objectives to balance exploration, exploitation, and teacher guidance for efficient reasoning model training.

Independent GitHub snapshot · Automated classification · Original documentation governs