The authors introduce DTDD, an adaptive on-policy distillation (OPD) framework.
Adaptive distillation triggers teacher takeover when student diverges
DTDD uses segment-level divergence to switch from on-policy distillation to teacher-generated recovery, improving stability and sample efficiency across reasoning and agentic tasks.
Big Tech
Qijia He · Zequn Li · Spencer Luo · Tianle Chen · Pradyumna Narayana · Zixian Ma · +3 more
Google · University of Washington · Boston University
Research Digest··3 min read
The authors propose Divergence-Triggered Dynamic Distillation (DTDD), which monitors student-teacher divergence during training and, when a segment diverges, rewinds to that segment and lets the teacher generate a recovery suffix that the student learns via behavior cloning.
Why this paper
From Google and 2 others
In one line
Divergence-triggered teacher takeover during on-policy distillation stabilizes long-horizon training and accelerates learning from base initializations.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§