Reasoning-Aware Compression Protects Critical Circuits in Efficient LLM Deployment

Summary

The paper presents a reasoning-aware compression framework for reducing the deployment cost of large reasoning models without treating every model component identically. It evaluates quantization settings on GSM8K, FOLIO, MATH-500, ProofWriter, and MuSiQue, combining held-out calibration data with hardware-level GPU energy measurements. The authors profile INT4 sensitivity across 196–224 layer-and-projection pairs and selectively restore the most vulnerable circuits from INT4 to FP16. Uniform INT4 compression can unexpectedly increase total energy because it lengthens reasoning chains; on GSM8K, a 25% power reduction became a net energy increase. Vulnerability also depends on the task and architecture: attention projections were more important for mathematical reasoning, while logical-inference sensitivity varied across architectures. Selective compression produced accuracy-energy tradeoffs unavailable to uniform methods. In the reported example, the Top-10% configuration for R1-Qwen-7B improved ProofWriter performance by 12 percentage points over FP16 while using 9.7% less energy, with the result validated on held-out data across all five benchmarks.