Research map / Architectures and efficiency / Quantization
GlowQ: Group-Shared LOw-Rank Approximation for Quantized LLMs
Selim An, Il hong Suh, Yeseong Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add covariance-aligned whitening to stacked SVD (ablation) | GlowQ (group-shared) · C4, Qwen 3-8B | perplexity: 14.60 -> 13.40 | Improved |
| share right factor within input-sharing groups instead of layer-wise (ablation) | GlowQ (whitened) · C4, Qwen 3-8B | perplexity: 13.85 -> 13.40 | Improved |
| cache shared projection per group | GlowQ vs layer-wise · LLaMA 2 latency | TTFB / throughput: TTFB -5.57%; throughput +9.61% | Improved |
| restore only about half of units (selective restore) (ablation) | GlowQ-S · LLaMA 2 latency | TTFB / throughput: TTFB -23.39%; throughput +37.44% | Improved |
| restore only about half of units (selective restore) (ablation) | GlowQ-S vs GlowQ · 4 models, 7 tasks | avg accuracy / C4 ppl: -0.55% accuracy; +0.15 ppl C4 | Hurt |
| add GlowQ on top of GPTQ | GPTQ · WikiText-2 LLaMA 3.2-3B | perplexity: 9.32 -> 8.19 | Improved |
| add GlowQ on top of BnB | BnB · WikiText-2 LLaMA 3.2-3B | perplexity: 8.29 -> 8.10 | Improved |
| use shared B instead of layer-wise correction | GlowQ · Qwen1.5-MoE-A2.7B WikiText-2 | perplexity: 7.39 (layerwise) vs 7.41; memory about 63% lower | No clear effect |
| use GlowQ instead of L2QER at W4A4 | GlowQ · WikiText-2 LLaMA 3.2-3B | perplexity: 9.42 -> 9.21 | Improved |
| GlowQ vs baselines on perplexity across models | GlowQ · WikiText-2 W4A16 | perplexity: best or tied-best on 9 of 11 variants | Mixed |
Closest papers by meaning
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.90)
- LQER: Low-Rank Quantization Error Reconstruction for LLMs (ICML 2024 · similarity 0.90)
- SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models (ICML 2025 · similarity 0.89)
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.89)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.89)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
