Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning

AI Digest

研究发现，数学推理中答案不变性与表示不变性不同，模型准确率与表示区分度正相关，为理解数学推理机制提供新视角。

The study reveals that answer invariance and representation invariance differ in mathematical reasoning. Model accuracy correlates with representation distinctiveness, offering new insights into reasoning mechanisms.

Key points

• 重新排列规则不影响答案，但模型内部表示可能变化 Reordering rules doesn't change answers but may alter internal representations
• 准确率与表示区分度呈正相关，SNR达0.86 Accuracy correlates with permutation SNR across models
• 答案不变性与表示不变性是不同概念 Answer invariance differs from representation invariance
• 结果为数学推理的表示研究提供新方向 Findings offer new perspectives on reasoning representations

Takeaway: 数学推理中，答案不变性与表示不变性可共存，模型表现与表示区分度密切相关。 / Answer invariance and representation invariance can coexist in mathematical reasoning, with model performance linked to representation distinctiveness.

Why it matters 揭示数学推理中答案与表示的差异，为模型设计提供新思路。