SFT can generalize too: locking updates to one on-policy direction halves training time and edges past GRPO
Constraining supervised fine-tuning updates to the direction identified by on-policy training lets SFT match GRPO's generalization while cutting training time by more than half.