ATI-OPD prevents training instability in long-sequence on-policy distillation by replacing high-disagreement student prefixes with teacher segments. This approach allows the student to resume reasoning from a corrected state rather than reinforcing errors through long-context feedback loops.
HOW THIS AFFECTS YOU
●
builderThis provides a more robust way to perform on-policy distillation for sequence-heavy applications.
●
researcherYou can use this method to stabilize the training of student models on long-form generative tasks.