What's this about?
SAKI lets a student generate on-policy trajectories while a coupling mechanism records where teacher and student agree. Direct teacher correction is concentrated on actual divergence points.
What exactly is on-policy distillation?
The student learns on situations created by its own behaviour, reducing the gap between training examples and states it will encounter at use time.
Why does this matter?
7 HF upvotes. The verifier reports 4.22× rollout throughput while preserving the intended sampling distribution.