Code does not improve LLM math reasoning; structured traces do

A new research paper explores the impact of code on mathematical reasoning in large language models. The study found that while code improves programming abilities, it does not generally enhance mathematical reasoning and can even compete with knowledge-intensive tasks. The researchers discovered that structured reasoning traces, like math-text mixtures, are more effective for improving reasoning than executable code alone. They suggest that increasing the density of structured math-domain samples offers a targeted approach to boost mathematical reasoning without sacrificing programming performance. AI

影响 Clarifies which data characteristics improve LLM reasoning, suggesting more precise data-centric optimization strategies.

排序理由 Academic paper detailing findings on LLM training data. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

• code
• language model
• mathematical reasoning

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →