AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Principled SVD-based Delta Compression via Quantization Error Minimization

Boya Xiong, Shuo Wang, Weifeng Ge, Guanhua Chen, Yun Chen

ICML 2026 · Architectures and efficiency › Quantization · approach: KV cache quantization

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 changedWhereMeasuredEffect
add reconstruction target correction (RTC) (ablation)PRINMIX · RTC ablation (AVG)AVG accuracy: 60.2 -> 61.5Improved
add RTC (ablation)PRINMIX · AIME2024 (R1-Distill-Qwen-14B)accuracy: 27.5 -> 31.1Improved
add RTC (ablation)PRINMIX · Math500 (R1-Distill-Qwen-14B)accuracy: 78.2 -> 80.2Improved
add RTC (ablation)PRINMIX · LLAVA-1.5 GQA and SQAaccuracy: 62.8 vs 62.7 on GQA; 72.2 vs 72.1 on SQANo clear effect
vary fmax from 2 to 6 (ablation)PRINMIX · fmax sweep, R1-Distill-Qwen-14BAVG accuracy: 57.0, 55.0, 55.7, 56.4, 56.4 vs 50.5 for Delta-CoMeNo clear effect
replace Delta-CoMe's mixed precision with ILP quantization-error minimizationPRINMIX · 13-14B models AVGAVG accuracy: 63.3 -> 64.7Improved
replace Delta-CoMe's mixed precision with ILP quantization-error minimizationPRINMIX · 7B models AVGAVG accuracy: 71.9 -> 74.0Improved
replace Delta-CoMe with PRINMIXPRINMIX · AIME2024 (R1-Distill-Qwen-7B)accuracy: 30.0 -> 36.7Improved
replace Delta-CoMe with PRINMIXPRINMIX · MBPP/HumanEval 7B coderaccuracy: 82.7 vs 83.1 on MBPP; 85.0 vs 85.6 on HumanevalNo clear effect
replace Delta-CoMe with PRINMIXPRINMIX · MoE/enc-dec/Qwen3 mix AVGAVG accuracy: 42.9 -> 44.5Improved
compress delta with PRINMIX instead of LoRA at alpha=1/16PRINMIX · DeepSeek-LLM-7B code+math AVGAVG accuracy: 35.5 -> 40.8Improved
apply delta compression (PRINMIX) instead of PRINMIX-LoRA at alpha=1/64PRINMIX · DeepSeek-LLM-7B code+math AVGAVG accuracy: 35.2 -> 38.4Improved

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