📄 中文摘要
大型语言模型（LLMs）在数学问题上表现出色的表象掩盖了其实际的能力不足。虽然它们能自信地讨论复利公式和TDEE的计算步骤，但在面对相同问题时却可能给出不同的答案。这并不是幻觉问题，而是架构问题，试图通过提示来修复这一问题并不是有效的解决方案。以Claude为例，计算30年期抵押贷款时，给出的两次答案分别为$2,528和$2,533，而正确答案应为$2,528.27。这种概率不一致性显示了LLMs在处理数学问题时的局限性，强调了引入结构化数学工具的必要性。
为何大型语言模型需要结构化数学工具（而非提示工程）
出处: Why LLMs Need Structured Math Tools (Not Prompt Engineering)
发布: 2026年3月31日
📄 English Summary
Why LLMs Need Structured Math Tools (Not Prompt Engineering)
Large Language Models (LLMs) often appear competent in mathematics, confidently discussing compound interest formulas and the calculation of TDEE. However, they can provide different answers to the same question, indicating a deeper issue. This inconsistency is not merely a hallucination problem but an architectural one, suggesting that prompt engineering is not the right solution. For instance, when asked to calculate a 30-year mortgage on $400,000 at 6.5% APR, Claude provided two different monthly payment estimates: $2,528 and $2,533, while the correct answer is $2,528.27. This probabilistic inconsistency highlights the limitations of LLMs in mathematical tasks and underscores the need for structured mathematical tools.