SAKI: Maximal-Coupling-Routed Teacher Supervision for On-Policy Distillation
On-policy distillation (OPD) reduces train-test state mismatch by training a student on its own generated trajectories, but weak students may visit teacher-misaligned prefixes where supervision is less representative. We introduce SAKI (Supervision Allocation with KL-constrained Interpolation), which combines a KL-constrained teacher-guided rollout with maximal coupling and reu
Reading in Plain English mode — toggle below to switch
On-policy distillation (OPD) reduces train-test state mismatch by training a student on its own generated trajectories, but weak students may visit teacher-misaligned prefixes where supervision is less representative. We introduce SAKI (Supervision Allocation with KL-constrained Interpolation), which combines a KL-constrained teacher-guided rollout with maximal coupling and reu
This research introduces a novel distillation technique that addresses a key challenge in on-policy distillation, improving the efficiency and effectiveness of training student models. By optimizing supervision allocation, it offers a method to create more performant and potentially smaller models, which is crucial for deployment in resource-constrained environments.