Reasoning distillation depends on chain-of-thought (CoT) data that includes correct answers, yet producing such data efficiently remains a known bottleneck. The standard approach, forward CoT generation followed by rejection sampling, does yield verified examples but appears to be resource-intensive.
The new paper introduces CoTranslator, a method that instead translates reverse CoT into forward CoT. By reversing the reasoning process, the approach aims to generate verified CoT data more efficiently than sampling and filtering from scratch.
Because the abstract is truncated, details on the exact mechanism and experimental results are not available from this source. The paper is posted on arXiv under identifier 2610.20828.