AI Paper Atlas

Research map / Architectures and efficiency / Quantization

KV-Cache Transform Coding for Compact Storage in LLM Inference

Konrad Staniszewski, Adrian Łańcucki

ICLR 2026 · Architectures and efficiency › Quantization · approach: KV cache quantization · 13 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
compress KV cache with KVTC at 16x (about 20x after DEFLATE)Llama 3.1 8B, MN-Minitron 8B, Mistral NeMo 12B · GSM8K, MMLU, Qasper, LITM, RULER-VTaccuracy: GSM8K 56.8 -> 56.9 (Llama 3.1 8B)No clear effect
raise KVTC compression to 64x (ablation)KVTC on Llama 3.1 8B · LITMaccuracy: 99.3 -> 90.2Hurt
raise KVTC compression to 64x (ablation)KVTC on MN-Minitron 8B · LITMaccuracy: 99.3 -> 59.5Hurt
compress the 4 attention sink tokens (s=0 instead of s=4) (ablation)KVTC 64x on Llama 3.1 8B · benchmark tasksaccuracyHurt
compress the most recent 128 tokens (w=0 instead of w=128) (ablation)KVTC 64x on Llama 3.1 8B · benchmark tasksaccuracyHurt
use KVTC instead of KIVI 2-bitLlama 3.1 8B · LITMaccuracy: 88.8 -> 99.3Improved
use KVTC instead of token eviction (TOVA, H2O)Llama 3.1 8B · MMLUaccuracy: 44.8 -> 60.1Improved
use KVTC instead of xKVLlama 3.1 8B · QasperF1: 35.6 -> 40.7Improved
add DEFLATE entropy coding after quantization (ablation)KVTC · compression ratiocompression ratio: approximately 1.23x on top of quantizationImproved
KVTC at 16x on reasoning modelQwen 2.5 R1 7B · LiveCodeBenchaccuracy: 36.7 -> 31.6Hurt
KVTC at 8x on reasoning modelQwen 2.5 R1 1.5B and 7B · LiveCodeBenchaccuracy: 16.4 -> 16.1 (1.5B); 36.7 -> 36.5 (7B)No clear effect
KVTC on Llama 3.3 70B with four-GPU pipelineLlama 3.3 70B Instruct · MATH-500accuracy: 75.6 -> 72.6 at 20xHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy