CriticOPD Fixes Unreliable Teacher Supervision in Online Policy Distillation

Researchers from BYU and UNC Chapel Hill propose CriticOPD, addressing the overlooked recoverability flaw in online policy distillation where teachers often follow students' errors instead of correcting them. The critique-guided approach improves distillation accuracy by 4.57 points.

2026-10-15 ~ 2026-10-17 · 2 related posts

• CriticOPD shows teachers follow student errors and fixes on-policy distillation with critiques — Kaixiang Zhao · 2026-10-15
• CriticOPD: Teacher Critiques Fix On-Policy Distillation, +4.57 Points on Qwen3-1.7B — tw_killian · 2026-10-17