ReDiST builds student-conditioned traces for reasoning distillation
Original title: Beyond the Answer: Student-Conditioned Trace Construction for Reasoning Distillation
Offline selection ranks traces by answer recovery and verified-answer probability before distilling only the student model.
ReDiST constructs evidence-rich reasoning traces conditioned on a target student, revising correct drafts against evidence and repairing unsuccessful drafts with the verified answer. A frozen student-matched reader ranks candidates using answer recovery and verified-answer probability. Selected traces are distilled with an auxiliary answer loss, and inference uses only the distilled student. The authors report gains on source and transfer question-answering benchmarks over answer-only, original-trace, and evaluated rewriting baselines.
- Construction and selection happen offline.
- Rankings can differ between source and transfer benchmarks.
- Supervision balances trace content with final-answer loss.
- Baselines include answer-only and original-trace supervision.
- Products
- ReDiST
From the original article
1 min read at arxiv.orgA reasoning trace can accompany a correct answer without providing effective supervision for a particular student. We study how to construct traces that improve learning beyond answer-only supervision by considering both their intermediate content and their accessibility to the target learner. We introduce ReDiST, a student-conditioned trace-construction and distillation framework for evidence-rich reasoning.
By Shaoang Li · arxiv.org
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.