Calibrating Teacher--Student Discrepancy for On-Policy Distillation
Cal-OPD tries to strip teacher-side noise out of the signal a student model learns.
The paper says standard on-policy distillation can treat a teacher model’s own deviations as if they were the real capability gap between teacher and student. That problem grows under privileged OPD, where extra information can shift the teacher’s likelihoods more strongly. The authors propose estimating the teacher’s self-deviation region with positive and negative privileged interventions, then keeping only discrepancy outside that region. On math reasoning benchmarks, Cal-OPD reportedly beats standard OPD and variants while using about 52-65% of the original discrepancy signal. HF Daily Papers' note
The paper says standard on-policy distillation can treat a teacher model’s own deviations as if they were the real capability gap between teacher and student. That problem grows under privileged OPD, where extra information can shift the teacher’s likelihoods more strongly. The authors propose estimating the teacher’s self-deviation region with positive and negative privileged interventions, then keeping only discrepancy outside that region. On math reasoning benchmarks, Cal-OPD reportedly beats standard OPD and variants while using about 52-65% of the original discrepancy signal. HF Daily Papers' note
score 4