1
FormInv: A Measurement Protocol for Semantic Invariance in Mathematical Reasoning Benchmarks
一次改写质量审计揭露了数学推理基准中语义漏洞，排名翻盘只需移除3.1%错误项，单一模型评估的盲区由此暴露。
arXiv:2605.29001v1 Announce Type: cross Abstract: A paraphrase-quality audit of MathCheck (ICLR 2025) detected 4 semantically incorrect paraphrases in…