The research introduces a reasoning-aware compression framework designed to mitigate the energy costs associated with deploying Large Reasoning Models. Current compression techniques often apply uniform quantization, potentially harming essential reasoning circuits. This framework benchmarks quantization conditions across five reasoning benchmarks: GSM8K, FOLIO, MATH-500, ProofWriter, and MuSiQue. Hardware-level GPU energy measurement is used during this process. A perturbation sweep is then performed on a held-out calibration split to profile per-module INT4 vulnerability across 196-224 layer and projection pairs. The most sensitive circuits are selectively restored to FP16.
Three key findings emerged from the study. First, INT4 quantization can increase energy by extending reasoning chains, resulting in a 25% power reduction becoming a net energy increase on GSM8K. Second, vulnerability is task-dependent, with attention projections being more critical for mathematical reasoning and sensitivity patterns differing by architecture in logical inference. Third, selective compression achieves Pareto-optimal points inaccessible to uniform methods, as demonstrated by an R1-Qwen-7B Top-10% gain on ProofWriter (+12 pp) at -9.7% energy, validated on held-out data across five reasoning benchmarks.
This approach offers a targeted strategy for compression, focusing on protecting the most vulnerable components of an LRM. The framework’s ability to identify task-specific sensitivities and achieve Pareto-optimal points represents a significant improvement over standard uniform quantization methods. This allows for more efficient deployment of LLMs with minimal performance degradation.
Source: https://arxiv.org/abs/2609.05512
