AI Paper Atlas

Research map / Architectures and efficiency / Quantization

CodeQuant: Unified Clustering and Quantization for Enhanced Outlier Smoothing in Low-Precision Mixture-of-Experts

Xiangyang Yin, Xingyu Liu, Tianhua Xia, BO BAO, Vithursan Thangarasa, Valavan Manohararajah, Eric Sather, Sai Qian Zhang

ICLR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 1 citations (OpenAlex)

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
replace random rotation with AOS fine-tuned rotation (ablation)CodeQuant · DeepSeek-V2-Lite A4W4 Embedding-wiseaverage accuracy: 0.652 -> 0.667; perplexity down 0.23 on WikiText2 and 0.31 on C4Improved
add KL divergence penalty to centroid fine-tuning (ablation)CodeQuant ACCF · DeepSeek-V2-Lite A4W4 Block-wiseaverage accuracy: 0.658 -> 0.667Improved
add KL divergence penalty to centroid fine-tuning (ablation)CodeQuant ACCF · Phi-mini-MoE-Instruct A4W4 Block-wiseaverage accuracy: 0.694 -> 0.700Improved
add KL penalty to reduce router change (ablation)CodeQuant · DeepSeek-V2-Lite router top-K change ratechange rate (%): 24.33 -> 22.82Improved
add KL penalty to reduce router change (ablation)CodeQuant · Qwen3-30B-A3B router top-K change ratechange rate (%): 60.21 -> 59.58Improved
add permutation (POG) before block-wise clustering (ablation)CodeQuant · DeepSeek-V2-Lite A4W4 Block-wise g=1024average accuracy: 0.663 -> 0.668Improved
add permutation (POG) before block-wise clustering (ablation)CodeQuant · Phi-mini-MoE-Instruct A4W4 Block-wise g=1024average accuracy: 0.710 -> 0.714Improved
use CodeQuant clustering instead of SqueezeLLM at 2-bit weights (ablation)CodeQuant · DeepSeek-V2-Lite A4W2average accuracy: 0.496 -> 0.568Improved
reduce weight bits from 4 to 2 (A4W4 to A4W2) (ablation)CodeQuant · DeepSeek-V2-Liteaverage accuracy: 0.667 -> 0.568Hurt
use CodeQuant instead of QuaRotquantized Qwen3-30B-A3B · A4W4 Embedding-wiseaverage accuracy: 0.581 -> 0.694Improved
use CodeQuant instead of QuaRotquantized Mixtral-8x7B · A4W4 Embedding-wiseWikiText2 perplexity: 16.79 -> 4.65Improved
use CodeQuant instead of QuaRotquantized Qwen3-30B-A3B · GSM8K 8-shot A4W4accuracy: 0.508 -> 0.867Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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