Reproduce “Reasoning Scaffolding: Distilling the Flow of Thought from LLMs” on OpenResearch

About this paper

Xiangyu Wen, Junhua Huang, Zeju Li, Min Li, Jianyuan Zhong, Zhijian Xu, Mingxuan Yuan, Yongxiang Huang and 1 more

Researchers from CUHK and Huawei Noah's Ark Lab developed Reasoning Scaffolding, a framework that distills the "flow of thought" from large language models into smaller models by leveraging discrete semantic signals. This approach consistently improved Pass@1 accuracy by an average of 14% over base models and 8% over state-of-the-art Chain-of-Thought (CoT) fine-tuning methods across multiple reasoning benchmarks, significantly boosting logical robustness for small language models.

Key insights

• Distilling the "flow of thought" through explicit, discrete semantic signals (rather than just textual rationales) enables Small Language Models to internalize logical structure, moving beyond superficial mimicry.
• Structuring reasoning with predefined semantic signals improves logical robustness and enhances the interpretability of the model's step-by-step reasoning process.
• The mere act of guiding generation with structured signals provides a useful inductive bias, with correct semantic signals offering a substantial additional performance boost.

Summary by alphaXiv.