AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Uni-DocRobust: Universal Plug-and-Play Robustness Enhancement for Multi-modal LLMs via Feature Restoration

Yuxuan Zhou, Baole Wei, Xingjian Hu, Haowei Chen, Yu Li, Xingyue Lin, Liangcai Gao, Zhi Tang

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models

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 Uni-DocRobust, a plug-and-play feature-restoration module that improves MLLM robustness to real-world document degradations (noise, blur, low resolution).

Problem

MLLMs degrade substantially on document understanding under real-world degradations such as noise, blur, and low resolution, due to scarce large-scale aligned training data and a lack of transferable restoration solutions across models.

Approach

A universal, plug-and-play robustness module that restores degraded features so it can transfer across MLLMs without retraining the base model; details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Uni-DocRobust feature restoration moduleInternVL3.5-2B · DocRobust-VQA ChartQA (corrupted)accuracy: 35.71 -> 59.45Improved
add Uni-DocRobustLLaVA1.5-7B · WildDocaccuracy: 28.71 -> 34.25Improved
add Uni-DocRobust instead of LoRA fine-tuningLLaVA1.5-7B · DocRobust-VQA ChartQAaccuracy: 41.95 -> 42.61Improved
add LoRA fine-tuning on top of Uni-DocRobustInternVL3.5-2B · DocRobust-VQA ChartQAaccuracy: 59.45 -> 61.55Mixed
add Uni-DocRobust under white-box adversarial attackInternVL3.5-2B · Adversarial ChartQAaccuracy: 12.72 -> 17.00Improved
use feature-level restoration instead of pixel-level DocResInternVL3.5-2B · DocRobust-VQA ChartQAaccuracy: 28.76 -> 59.45Improved
add Phase I multi-teacher distillation (vs random-init, full budget) (ablation)Uni-DocRobust · DocRobust-VQA averageavg score: 61.06 -> 62.72Improved
add Phase I distillation (vs random-init under fast adaptation) (ablation)Uni-DocRobust · DocRobust-VQA averageavg score: 52.29 -> 62.72Improved
use high-capacity Uni-DocRobust instead of standard low-capacity adapter (ablation)InternVL3.5-2B · DocRobust-VQA averageavg score: 58.36 -> 62.72Improved
increase number of teacher encoders in Phase I (one to three) (ablation)Uni-DocRobust · DocRobust-VQA averageavg score: 55.88 -> 62.72Improved
use InternViT teacher instead of CLIP teacher (ablation)Uni-DocRobust · DocRobust-VQA averageavg scoreImproved
scale base MLLM from 2B to 14B and add Uni-DocRobustInternVL3.5-14B · DocRobust-VQA ChartQAaccuracy: 50.32 -> 64.79Improved

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