同量不同答：语言模型中的数值表示不变性

AI 摘要

该研究提出一个精确有理数轨道框架，用3,600道精确有理数题和8,600条提示，测试五种开源权重模型在五种保持数值恒等的改写（小数/分数、科学计数法、百分数、数字词、单位换算）下是否给出相同答案。结果显示，规范准确率高达0.969–0.996，但轨道正确率降至0.848–0.981，轨道不变性为0.851–0.981；大部分严格解析器崩溃源于乘法形式科学计数法不在解析语法内，而非推理失败。Mistral Small 4在单位换算输入上仅得0.699，并产生265个与标签相差恰好10的幂次的错误。另一项9,000次调用实验表明，表示共识并未优于复述共识，反而产生更多误报。

正文节选

Same Quantity, Different Answer: Numerical Representation Invariance in Language Models Abstract Numerically equivalent word problems should yield the same canonical answer whether a quantity is written as a decimal, fraction, percentage, number word, scientific notation, or an exactly converted unit. We generate 3,600 exact-rational problems and 8,600 prompts spanning five identity-preserving transformation families, and evaluate five open-weight systems. After a fixed syntax audit that normali

发布时间：2026-09-23 12:00

抓取时间：2026-09-23 12:04

来源机构：arXiv