困惑性悖论：为何代码在大型语言模型提示中压缩效果优于数学

The study reveals that code generation exhibits a strong tolerance for aggressive prompt compression, while chain-of-thought reasoning degrades gradually. Validation across six code benchmarks (HumanEval, MBPP, HumanEval+, MultiPL-E) and four reasoning benchmarks (GSM8K, MATH, ARC-Challenge, MMLU-STEM) confirms the generalizability of the compression threshold across languages and difficulties. Furthermore, the first per-token perplexity analysis uncovers a 'perplexity paradox': code syntax tokens are preserved under high perplexity, whereas reasoning processes are adversely affected. These findings provide new insights into the differences in prompt compression performance between code and mathematics, laying the groundwork for future adaptive algorithms.