Reasoning-Aware Compression Restores FP16 on 10% of Circuits for Net Energy Gains
Selective protection of vulnerable reasoning circuits in LRMs delivers measurable energy reductions without uniform accuracy loss. The approach exposes how standard INT4 quantization can increase total energy via extended chains. Results position circuit-aware compression as a practical lever for sustainable LRM deployment.
The framework benchmarks INT4 quantization across 196-224 layer-projection pairs using perturbation sweeps on held-out calibration data from GSM8K, FOLIO, MATH-500, ProofWriter, and MuSiQue. Hardware GPU energy traces show uniform INT4 extends reasoning chains on GSM8K, converting a nominal 25% power cut into net energy increase. Vulnerability maps differ by task: attention projections dominate mathematical reasoning sensitivity while feed-forward layers vary more on logical inference tasks.
Task-dependent patterns align with prior circuit-level studies on attention heads in arithmetic models and documented INT4 accuracy cliffs in chain-of-thought sequences. Selective restoration reaches Pareto fronts unavailable to uniform methods by protecting only the highest-perturbation modules, confirming that uniform quantization erodes the very efficiency it targets when reasoning depth increases.
Operational deployment requires per-architecture calibration runs before production quantization. Future scaling to larger LRMs will hinge on automated identification of reasoning circuits rather than blanket precision policies, directly addressing measured energy overheads in sustained inference workloads.
Validation on held-out splits across all five benchmarks indicates the method generalizes beyond the calibration distribution without retraining.
Twagirayezu: Selective restoration pipelines will cut average energy per correct answer by 12% on MATH-500 within 18 months of open-source release.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2609.05512)
- [2]Supporting Source(https://arxiv.org/abs/2306.00978)
- [3]Supporting Source(https://arxiv.org/abs/2402.16775)