Bridged Self-Distillation unifies on-policy learning from any teacher
Original title: Bridged Self-Distillation for On-Policy Learning
A self-teacher scores the student’s own rollouts after an external teacher supplies only a reasoning trace.
Researchers propose Bridged Self-Distillation (BSD) for transferring reasoning from a stronger teacher to a smaller student. An external teacher supplies a reasoning trace, while a copy of the student conditioned on that trace scores rollouts in the student’s token space. A verifier supplies the outcome signal, and unsuccessful rollouts are adjusted with that token-level guidance. The same procedure covers shared-tokenizer, cross-tokenizer, and closed API teachers, and the authors say it beats scenario-specific baselines.
- The external teacher is queried only to produce text, including offline traces.
- BSD applies to shared-tokenizer, cross-tokenizer, and closed API teachers.
- Code is published at github.com/TUMLegalTech/Bridged-Self-Distillation.
From the original article
1 min read at huggingface.co
On-policy distillation can transfer the reasoning capabilities of a strong teacher to a smaller student. However, distillation from a stronger teacher model faces two mismatches: its solutions may be difficult for a smaller student to imitate, and its token-level probabilities may be incompatible with, or even unavailable to, the student. Existing methods handle these settings through different forms of supervision.
By Runsheng Chen, Yihong Liu, Raoyuan Zhao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.