[2603.25385] GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs

Computer Science > Machine Learning arXiv:2603.25385 (cs) [Submitted on 26 Mar 2026] Title:GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs Authors:Selim An, Il hong Suh, Yeseong Kim View a PDF of the paper titled GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs, by Selim An and 2 other authors View PDF HTML (experimental) Abstract:Quantization techniques such as BitsAndBytes, AWQ, and GPTQ are widely used as a standard method in deploying large language models but often degrades accuracy when using low-bit representations, e.g., 4 bits. Low-rank correction methods (e.g., LQER, QERA, ASER) has been proposed to mitigate this issue, however, they restore all layers and insert error-correction modules into every decoder block, which increases latency and memory overhead. To address this limitation, we propose GlowQ, a group-shared low-rank approximation for quantized LLMs that caches a single shared right factor per input-sharing group and restores only the groups or layers that yield the highest accuracy benefit. GlowQ computes the high-precision projection once per input-sharing group and reuses it across its modules, reducing parameter and memory overhead, and retaining the expressivity of layer-specific corrections. We also propose a selective variant, GlowQ-S, that applies the cached shared module only where it provides the largest benefit. Compared with strong baselines, our approach reduces TTFB by (5.6%) and increases throughput by (9.6%) on ...