Research map / Architectures and efficiency / Quantization
GSRQ: Gain-Shape Residual Quantization for Sub-1-bit KV Cache
Soosung Kim, Minjae Park, Eui-Young Chung, Jaeyong Chung
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 |
|---|---|---|---|
| replace standard K-means with GSKM (unweighted) (ablation) | GSRQ residual quantization · WikiText-2 (0.375-bit) | perplexity: 19.50 -> 12.77 | Improved |
| add raw gradient-norm weighting to GSKM (ablation) | GSKM · WikiText-2 (0.375-bit) | perplexity: 12.77 -> 13.57 | Hurt |
| add log-smoothed gradient weighting to GSKM (ablation) | GSKM · WikiText-2 (0.375-bit) | perplexity: 12.77 -> 12.49 | Improved |
| log-weighting instead of raw weighting vs. no weighting on keys (ablation) | GSKM · WikiText-2 (0.375-bit), key perplexity | perplexity: 6.87 -> 6.82 | Improved |
| GSRQ vs VQLLM at 1-bit | LLaMA-3-8B-Instruct KV cache · LongBench average | accuracy: 11.34 -> 33.54 | Improved |
| GSRQ vs VQLLM at 1-bit | LLaMA-3-8B-Instruct KV cache · GSM8K | exact match accuracy: 2.27 -> 26.16 | Improved |
| GSRQ vs VQLLM at 2-bit | Mistral-7B-Instruct-v0.1 KV cache · GSM8K | exact match accuracy: 27.90 -> 28.13 | No clear effect |
| GSRQ vs VQLLM at 1-bit | LLaMA-3-8B-Instruct KV cache · RULER | accuracy: 20.0 -> 72.0 | Improved |
| GSRQ vs VQLLM at 1-bit | LLaMA-3-8B KV cache · C4 perplexity | perplexity: 74.47 -> 8.92 | Improved |
| GSRQ vs VQLLM at 2-bit on Qwen3-8B | Qwen3-8B KV cache · Few-shot commonsense/math average | accuracy: 36.44 -> 40.75 | Improved |
| GSRQ vs VQLLM at 1-bit on Qwen3-8B | Qwen3-8B KV cache · TriviaQA (LongBench) | accuracy: 11.43 -> 8.17 | Hurt |
| GSRQ vs VQLLM at 2-bit on TruthfulQA | LLaMA-3-8B KV cache · TruthfulQA | accuracy: 51.05 -> 50.12 | Hurt |
Closest papers by meaning
- NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache (NeurIPS 2025 · similarity 0.91)
- KV Cache is 1 Bit Per Channel: Efficient Large Language Model Inference with Coupled Quantization (NeurIPS 2024 · similarity 0.90)
- KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization (NeurIPS 2024 · similarity 0.90)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.90)
- CommVQ: Commutative Vector Quantization for KV Cache Compression (ICML 2025 · similarity 0.89)
- Cache Me If You Must: Adaptive Key-Value Quantization for Large Language Models (ICML 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.
