TOPSD limits teacher supervision to verified prefixes in distillation
Original title: Knowing When to Teach: Learning from Teachable Prefixes in Multilingual Reasoning Distillation
It beats conventional on-policy self-distillation on math reasoning in ten languages, especially lower-resource ones.
Researchers describe Teacher Drift, where a self-distillation teacher follows a student's wrong intermediate steps instead of correcting them, and say this starts earlier in lower-resource languages. Teachable On-Policy Self-Distillation estimates where drift begins from teacher-student likelihood, then distills only prefixes from which the teacher still recovers the correct answer. Experiments cover multilingual mathematical reasoning across ten languages and three model scales, with gains over conventional OPSD particularly in low- and medium-resource languages.
- Teacher Drift tends to occur earlier in lower-resource languages.
- A prefix is chosen using teacher-student likelihood.
- Distillation runs only if the teacher recovers the correct answer.
- Tests span ten languages and three model scales.
From the original article
1 min read at huggingface.co
On-Policy Self-Distillation (OPSD) improves reasoning by training a student on its own reasoning, using the same model conditioned on reference solutions as a teacher. Applying OPSD to multilingual reasoning requires the teacher to provide useful feedback across languages, even when the student makes mistakes in intermediate steps.
By Jaewon Cho, Minseo Yoon, Jinyoung Park · 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.