The authors introduce Semi-OPD (semi-on-policy distillation), which samples rollouts once from the initial student and keeps them fixed throughout training.
Offline student rollouts often beat on-policy distillation for language models
Across 17 teacher-student pairs, a simple semi-on-policy method outperforms on-policy in 14 cases with up to 13.6% accuracy gain and 11.4x speedup.
Independent
Siyan Zhao · Yonggan Fu · Jindong Jiang · Shih-Yang Liu · Song Bian · Byung-Kwan Lee · +5 more
Research Digest··3 min read
Zhao et al.
Why this paper
Independent
In one line
Semi-OPD, which uses initial offline student rollouts, often outperforms on-policy distillation across diverse teacher-student pairs.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§