AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

One Patch Doesn’t Fit All: Adaptive Patching for Native-Resolution Multimodal Large Language Models

Wenzhuo Liu, Weijie Yin, Fei Zhu, Shijie Ma, Haiyang Guo, yi chen, Xiao-Hui Li, LiangXiao, ChaoFeng, Cheng-Lin Liu

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Benchmarks and analysis of MLLMs

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 adaptive patching for native-resolution MLLMs so that patch granularity varies with image content, preserving fine detail for documents and charts while keeping cost low for simple images.

Problem

Native-resolution MLLMs must handle images of varying resolution; uniform patching wastes compute on simple images or loses fine detail on dense documents and charts.

Approach

Adaptive patching that chooses patch sizes per image region so that high-resolution detail is preserved only where needed, rather than using one fixed patch size for all inputs.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add adaptive patch size selection (training-free PI-resize)Qwen2.5-VL-3B · MMEscore: 2135.90 -> 2210.41Improved
add adaptive patch size selectionQwen2.5-VL-3B · OCRBenchscore: 821 -> 845Improved
add adaptive patchingOvis2.5-2B · DocVQA valaccuracy: 93.90 -> 93.22Hurt
add adaptive patchingQwen2.5-VL-3B · LLaVABenchscore: 66.30 -> 65.34Hurt
add adaptive patching across model scales (ablation)Qwen2.5-VL 3B/7B/32B, Ovis2.5 2B/9B · MMEscore: 2313.17 -> 2345.96 (7B)Improved
add adaptive patching instead of resizing images to a fixed range (ablation)Qwen2.5-VL-3B · MMEscore: 2165.01 -> 2210.41Improved
use adaptive patch sizes during SFT (training-based)SAIL-VL-2B · ChartQA test (LLaVA1.5-665K)accuracy: 47.96 -> 54.84Improved
use PI-resize rather than bilinear or area interpolation for patch embedding conversion (ablation)Qwen2.5-VL-3B · pixel-range benchmarksaccuracyImproved
increase alpha (resolution sensitivity) in patching law (ablation)AdaPatch · Qwen2.5-VL-3B hyperparameter sweepaccuracyHurt
use adaptive patching at high resolutionAdaPatch · inference timelatencyMixed

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