Tsinghua's CoTranslator distills reverse CoT to boost 4B model math scores
Tsinghua · hf · 2026-10-16
Tsinghua, Ant Group and Xiaomi propose CoTranslator, which translates cheap reverse chain-of-thought traces (conditioned on known answers) to match verified forward reasoning distributions for reasoning distillation, avoiding post-hoc rationalization while cutting teacher calls. On AIME25/26, HMMT25/26, Omni and IMO, distilled Qwen3-4B-Base clearly beats SSR/SUP baselines (e.g., 41.06 vs 31.53 on Omni).
More from Research
- Free Textbook: Introduction to Machine Learning Covers the Math from Optimization to Generative Models — TheTuringPost · 2026-10-16
- Free ML math textbook: Laurent Younes' Introduction to Machine Learning covers Adam to generative models — TheTuringPost · 2026-10-16
- RFdiffusion3 published in Science, alongside Chaperone agent harness for bio-molecular design — anshulkundaje · 2026-10-16
- Microsoft paper: history-aware model routing lifts multi-turn agent task pass rate by 3.3 points — dair_ai · 2026-10-16
- Researcher: everyone should build a good eval at least once — bad evals are everywhere — gabriberton · 2026-10-16
- The square lattice connective constant remains unsolved — and might be ripe for AI — tak3sh8 · 2026-10-16