The authors first generated reverse chain-of-thought responses, in which a teacher sees both a problem and its known answer.
Translating answer-guided reasoning cuts the cost of model distillation
CoTranslator converts reasoning produced with access to answers into training traces resembling verified, answer-free reasoning.
Chinese Tech
Zinan Tang · Zhuoshi Pan · Kairui Chen · Ze Yan · Shaomian Zheng · Yukun Zhang · +2 more
Tsinghua University · Ant Group · Xiaomi Corporation
Research Digest··3 min read
Tang and colleagues treat the gap between reverse and forward chain-of-thought as a translation problem.
Why this paper
From Ant Group and 2 others · Part of Reasoning Distillation Alignment, now 35 papers
In one line
CoTranslator learns to translate reverse CoT to forward CoT, improving student accuracy while using less compute.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§