AI Paper Atlas

Research map / Architectures and efficiency / Quantization

No outlier channels but with outlier blocks

Shanwen Mao, Hao Zhang, Jiasheng Li, Haoyu Qiao, Chenxin Cai, Tingting Wu, Jie Liu

ICLR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs

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 OCP (ablation)AQLM (3-bit) · WikiText2, Llama3-8Bperplexity: 7.05 -> 6.93Improved
add OCP (ablation)VPTQ (3-bit) · WikiText2, Llama3-8Bperplexity: 7.01 -> 6.86Improved
add OCP (ablation)GPTVQ (3-bit) · WikiText2, Llama3-8Bperplexity: 6.03 -> 6.12Hurt
add OCP (ablation)GPTVQ (2-bit) · WikiText2, Llama3-8Bperplexity: 10.34 -> 9.14Improved
add OCP (ablation)VPTQ (2-bit) · WikiText2, Llama3-8Bperplexity: 9.19 -> 8.92Improved
add OCP (ablation)AQLM (2-bit) · MMLU average, Llama3-8Baccuracy: 44.67 -> 45.63Improved
add OCP (ablation)GPTVQ (3-bit) · MMLU average, Llama3-8Baccuracy: 50.71 -> 48.81Hurt
use linear-level compensation instead of random (ablation)NuBitQ · layer 31 of LLaMA3-8Brelative PPL reduction: 1.00× -> 5.27×Improved
use model-level compensation instead of random (ablation)NuBitQ · layer 31 of LLaMA3-8Brelative PPL reduction: 1.00× -> 2.12×Improved
use NuBitQ with OCP instead of NuBitQ alone at 2-bit scaleNuBitQ · WikiText2, Llama3-8Bperplexity: 5.79 -> 5.66Mixed
increase vector length d from 8 to 64 (ablation)NuBitQ · Ada Lovelace computation timetime (ms): 63.7 -> 98.8Hurt
shrink vector length to d=4 versus FP16 (ablation)NuBitQ · Ada Lovelace computation timetime (ms): 134.3 -> 63.6Improved

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