Reproduce “Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haolei Xu, Yuchen Yan, Yongliang Shen, Wenqi Zhang, Guiyang Hou, Shengpei Jiang, Kaitao Song, Weiming Lu and 2 more
Researchers from Zhejiang University, SF Technology, and Microsoft Research Asia identified and addressed "Thought Leaps" in Chain-of-Thought (CoT) datasets, where human-generated reasoning inadvertently omits intermediate steps. Their CoT-Bridge model automatically detects these gaps and generates missing steps, leading to consistent performance improvements of up to +15.63% on challenging math benchmarks and enhanced generalization for large language models.
Key insights
- Thought Leaps in CoT datasets act as a substantial barrier to LLM learning and generalization, even more so than factual inaccuracies, significantly reducing model performance.
- Accurate identification of the exact positions of reasoning gaps is crucial; merely adding random steps can degrade rather than improve reasoning performance.
- Specialized training with a dedicated dataset (ScaleQM+) for the CoT Thought Leap Bridge Task yields superior and more consistent performance compared to zero-shot bridging by general-purpose LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.