AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

ImpQuant: Fine-Grained Importance-Aware Quantization for Large Vision-Language Models

Jundong Zhou, Tianao Cai, Yujie Huang, Xinbing Wang, Guang-Zhong Yang, Nanyang Ye

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ImpQuant, a post-training quantization approach for large vision-language models that accounts for heterogeneous information density across multimodal tokens rather than treating inputs as homogeneous sequences.

Problem

Post-training quantization of large vision-language models for low-bit deployment; existing PTQ pipelines borrowed from text-only LLMs treat multimodal inputs as homogeneous sequences, ignoring heterogeneous information density.

Approach

Fine-grained, importance-aware quantization that exploits the varying information density of different multimodal inputs instead of applying uniform LLM-style PTQ.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add token importance-aware weight quantization (W-imp) (ablation)ImpQuant · VQA average (W4A8)average accuracy (%): 71.38 -> 73.85Improved
add token importance-aware weight quantization (W-imp) (ablation)ImpQuant · OCRBench (W4A8)accuracy (%): 74.70 -> 82.90Improved
add token importance-aware weight quantization (W-imp) (ablation)ImpQuant · TextVQA (W4A8)accuracy (%): 77.89 -> 80.88Improved
add channel outlier-aware activation quantization (A-out) (ablation)ImpQuant · VQA average (W4A8)average accuracy (%): 73.33 -> 73.85Improved
add outlier-aware activation quantization (A-out) (ablation)ImpQuant · OCRBench (W4A8)accuracy (%): 82.30 -> 82.90Improved
10x weight on important visual tokens in calibration lossGPTQ-style W3A16 · VizWiz and TextVQAaccuracy (%): 67.09 -> 68.38 (VizWiz); 79.03 -> 80.64 (TextVQA)Improved
10x weight on important textual tokens in calibration lossGPTQ-style W3A16 · VizWiz and TextVQAaccuracy (%): 67.09 -> 67.75 (VizWiz); 79.03 -> 79.96 (TextVQA)Improved
ImpQuant instead of best baseline MBQQwen2.5-VL-7B · VQA average (W4A8)average accuracy (%): 69.93 -> 73.85Improved
ImpQuant instead of SmoothQuantInternVL2-8B · VQA average (W3A16)average accuracy (%): 71.39 -> 71.87Improved
ImpQuant instead of SmoothQuantInternVL2-8B · POPE average accuracy (W3A16)accuracy (%): 86.82 -> 88.18Hurt
ImpQuant instead of RTNQwen2.5-VL-7B · latency speedup over FP16 (batch size 1)speedup: 1.47x -> 1.45xNo clear effect
ImpQuant instead of SmoothQuantLLaVA-OneVision-7B · OCRBench (W3A16)accuracy (%): 39.30 -> 57.20Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 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