NVIDIA Study Challenges On-Policy Distillation Superiority

An NVIDIA study across 17 teacher-student model pairs (1.5B–235B) found that Semi-OPD, distilling on the initial student's offline rollouts, beats on-policy distillation in 14 cases, challenging the assumption that on-policy sampling is always better.

2026-10-12 ~ 2026-10-13 · 3 related posts

• NVIDIA: Offline Student Rollouts Often Beat On-Policy Distillation in 14 of 17 Pairs — nvidia · 2026-10-12
• NVIDIA paper: offline rollouts beat on-policy distillation in 14 of 17 teacher-student pairs — siyan_zhao · 2026-10-13

1 near-duplicate retellings: siyan_zhao