OPMD: On-Policy Mechanism Distillation for Intermediate Computation
September 28, 2026
On-Policy Mechanism Distillation (OPMD) moves beyond output supervision by transferring teacher-derived intermediate computations to a student model. It constructs a student-compatible teacher to ensure that transferred representations are geometrically and computationally usable.
HOW THIS AFFECTS YOU
●
researcherThis offers a more effective way to distill model logic than simple representation matching or output-only training.