Cal-OPD filters a teacher model's own errors out of distillation training

single source· 1 articles · confidence: medium · first seen 2026-09-17 20:00 UTC

What this means for you

Nothing to install: this is a preprint with no code or weights released. If you run distillation pipelines, the diagnostic is worth copying — measure the teacher's own deviation before treating the teacher-student gap as capability, and expect to train on a smaller fraction of it.

A preprint proposes a fix for a known flaw in on-policy distillation, a training method where a smaller student model is scored on its own outputs by a stronger teacher and learns the gap. That gap mixes the teacher's capability with the teacher's own errors, which standard training copies. Cal-OPD estimates the teacher's self-deviation using privileged interventions — extra context the teacher sees and the student will not — and keeps only the signal beyond it. On maths reasoning benchmarks it retains 52–65% of the original signal and beats standard on-policy distillation across model scales. No evaluation date, code or weights are given.

Key facts

  • ·The preprint is arXiv 2609.21619, posted 17 September 2026. source
  • ·The proposed method is called Calibrated On-Policy Distillation, or Cal-OPD. source
  • ·Cal-OPD keeps only about 52–65% of the original teacher-student discrepancy as the optimisation signal. source
  • ·It estimates the teacher's self-deviation region using positive and negative privileged interventions. source
  • ·Experiments were run on mathematical reasoning benchmarks, with the method reported to outperform standard on-policy distillation and its variants across model scales. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire