ByteDance paper reveals generalization limits of on-policy distillation

A ByteDance-led paper with university collaborators finds that on-policy and multi-teacher distillation consolidate RL capabilities in-distribution, but student models lag 18-24% behind the best teacher on out-of-distribution generalization, and even behind the base model.

2026-11-05 ~ 2026-11-06 · 2 related posts

• ByteDance study finds on-policy distillation leaves models 2.45-16.25% below base on out-of-distribution tasks — ByteDance · 2026-11-05
• ByteDance paper: on-policy distillation students lag best teachers by 18-24% on OOD tasks — brandondamos · 2026-11-06