CoTranslator: Reverse to Forward Chain-of-Thought Translator for Reasoning Distillation

Key points

• Reasoning distillation requires chain-of-thought (CoT) data with correct answers, but obtaining such data efficiently remains challenging.1
• Forward CoT generation (FCG) followed by rejection sampling (RS) produces verified, question-only responses, yet may require many teacher calls for difficult problems.1
• Reverse CoT generation (RCG) is cheaper because it conditions on the known answer, but access to this answer can lead to post-hoc rationalization and a mismatch with the student's answer-free inference setting.1
• These results show that CoTranslator offers an efficient and transferable approach to reasoning distillation.1

Sources (2)

• [1]CoTranslator: Reverse to Forward Chain-of-Thought Translator for Reasoning Distillation
  Hugging Face Daily Papers · Oct 14, 12:00 AM

  We propose CoTranslator, which learns to translate reverse CoT responses toward the distribution of verified forward responses.

  Reasoning distillation requires chain-of-thought (CoT) data with correct answers, but obtaining such data efficiently remains challenging.

• [2]CoTranslator: Reverse to Forward Chain-of-Thought Translator for Reasoning Distillation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 07:15 PM · same content

Extractive summary: sentences quoted from the sources.