AI Paper Atlas

Research map / Architectures and efficiency / Quantization

KBVQ-MoE: KLT-guided SVD with Bias-Corrected Vector Quantization for MoE Large Language Models

Zukang Xu, Zhixiong Zhao, Xing Hu, Zhixuan Chen, Dawei Yang

ICLR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 8 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
add IDRE (KLT-guided SVD shared component) (ablation)KBVQ-MoE · WikiText2, 3-bitperplexity: 18.72 -> 11.67Improved
add BCOS (channel-wise bias correction) (ablation)KBVQ-MoE · WikiText2, 3-bitperplexity: 18.72 -> 14.32Improved
add BCOS on top of IDRE (ablation)KBVQ-MoE · WikiText2, 3-bitperplexity: 11.67 -> 9.26Improved
add KLT guidance to SVD (ablation)KBVQ-MoE · WikiText2, 2-bitperplexity: 14.03 -> 11.87Improved
add SVD shared component (k/n=1/128) (ablation)KBVQ-MoE · WikiText2, 2-bitperplexity: 15.30 -> 11.87Improved
increase truncated rank from 1/128 to 1/32 (ablation)KBVQ-MoE · WikiText2, 2-bitperplexity: 11.87 -> 11.01No clear effect
add IDRE and BCOS to GPTVQ (ablation)GPTVQ · WikiText2, 2-bitperplexity: 12.88 -> 9.43Improved
add IDRE and BCOS to VPTQ (ablation)VPTQ · WikiText2, 2-bitperplexity: 10.17 -> 8.78Improved
KBVQ-MoE vs GPTQ at 3-bit on DeepseekV2-LiteKBVQ-MoE · Avg zero-shot accuracyaccuracy: 68.89 vs 68.73No clear effect
KBVQ-MoE vs VQ at 2-bit on Qwen3-30B-A3BKBVQ-MoE · Avg zero-shot accuracyaccuracy: 30.61 -> 63.37Improved
KBVQ-MoE vs GPTQ at 2-bit on Mixtral-8x7BKBVQ-MoE · Avg zero-shot accuracyaccuracy: 67.56 -> 75.69Improved
quantize decoder with KBVQ-MoE at 2-bitKBVQ-MoE · Decoder speedspeedup: 22.31 -> 35.24 tokens/sImproved

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