Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation
TGOPD checks whether the teacher is right for a prompt before letting its token-level supervision shape the student.
The paper argues that vanilla on-policy distillation can amplify a teacher’s confident mistakes because it applies dense supervision uniformly. TGOPD uses verifier-scored teacher probes to decide whether a prompt gets dense OPD or falls back to verifier-grounded GRPO. The authors report gains over vanilla OPD across 4B and 35B students in math, code, and instruction-following settings. They also report higher teacher-node GPU utilization in one measured 4B run, from 9.8% to 78.9%. HF Daily Papers' note
The paper argues that vanilla on-policy distillation can amplify a teacher’s confident mistakes because it applies dense supervision uniformly. TGOPD uses verifier-scored teacher probes to decide whether a prompt gets dense OPD or falls back to verifier-grounded GRPO. The authors report gains over vanilla OPD across 4B and 35B students in math, code, and instruction-following settings. They also report higher teacher-node GPU utilization in one measured 4B run, from 9.8% to 78.9%. HF Daily Papers' note
score 4