ShortOPD Recovers Pruned LLMs via On-Policy Distillation
July 13, 2026
ShortOPD recovers performance in structured-pruned LLMs by using short-to-long on-policy distillation. It utilizes the compressed model's own states with dense token-level supervision from a frozen teacher to prevent the collapse of free-form generation.
HOW THIS AFFECTS YOU
●
builderYou can use more aggressive pruning to reduce costs while maintaining the generative quality required for production.
●
researcherThis provides a more effective training recipe for recovering capabilities in compressed models.