Verifier-Calibrated On-Policy Distillation: A Practical Algorithm for Teaching Models Without Making Them Forget
A distribution-first training proposal that uses student rollouts, verifier rewards, and calibrated teacher guidance instead of blindly imitating style tokens.
The note argues that SFT, RL, and on-policy distillation should be understood as different ways of moving the model distribution, then proposes verifier-calibrated on-policy distillation as a practical hybrid.
