The Perplexity Paradox: Why Code Compresses Better Than Math in LLM Prompts

Researcher Warren Johnson's paper 'The Perplexity Paradox' introduces TAAC (Task-Aware Adaptive Compression), a new algorithm for compressing LLM prompts. It validates that code generation tolerates aggressive compression while math reasoning degrades, due to a paradox where critical numbers are pruned. TAAC achieves a 22% cost reduction while preserving 96% of output quality, outperforming fixed-ratio compression by 7% across multiple benchmarks like HumanEval and GSM8K.