Running summarizing tools on a new article

Abstract: Quantization is widely used to enable memory-efficient deployment of large language models, yet its impact on instruction-following behavior remains poorly characterized. Prior work typically reports aggregate accuracy degradation, obscuring how compression affects individual instructions and decision boundaries. In this work, we conduct a fine-grained behavioral analysis of 4-bit quantization on instruction-tuned language models, comparing FP16 and 4-bit NF4 variants across 1,000 multi-constraint prompts from IFEval++. Contrary to the common assumption that aggressive quantization uniformly degrades performance, we find that instruction-following behavior is largely preserved. Across models, 72-76% of prompts exhibit no change in instruction-violation patterns under quantization, yielding an effective instruction-following retention of up to 82.5% as measured by a novel robustness coefficient. Behavioral changes are rare, structured, and bidirectional: while some prompts degrade, a comparable fraction improve, indicating controlled perturbation rather than random failure. To explain these effects, we introduce a boundary distance metric that measures proximity to instruction-following failure and an instruction-type fragility analysis that reveals selective sensitivity. We show that quantization primarily exposes intrinsic capability boundaries, with semantic and language-level constraints exhibiting higher fragility than structural or keyword-based constraints, and that these patterns vary across model architectures. Together, our results reframe quantization as not only an efficiency mechanism but also a diagnostic probe for revealing latent model weaknesses. This perspective proposes new uses of quantization for capability analysis and targeted robustness evaluation beyond deployment-focused compression.