The authors first ran controlled experiments with Qwen3 students to isolate how the amount and placement of teacher generation affects both rollout quality and student learning.
Adaptive teacher intervention boosts distillation accuracy and stability
MAESTRO tunes when and how long a teacher generates in on-policy distillation, improving accuracy across eight math benchmarks.
Chinese Tech
Yuhao Wang · Ruiyang Ren · Yinan Zhang · Ruiqing Zhang · Jing Liu · Chunyan Miao
Nanyang Technological University · Baidu Inc.
Research Digest··2 min read
Wang et al.
Why this paper
From Baidu Inc. and Nanyang Technological University
In one line
Adaptive teacher intervention guided by policy disagreement improves accuracy and reduces training length in on-policy distillation.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 0.6B)
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§