Paper 2606.08063
Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding?
- Published
- Jun 2026
- Research lab
- Independent
- Citations
- 1
- GitHub
- 363 stars
01 In brief
Summary
Multimodal Large Language Models (MLLMs) degrade under real-world visual corruptions.
Existing robustness methods are limited: black-box feature alignment lacks interpretability, and text-based reasoning cannot restore pixel-level details.
This paper proposes Robust-U1, a framework that equips MLLMs with explicit visual self-recovery capability.
It uses a three-stage pipeline: supervised fine-tuning on ImageNet-C for initial reconstruction, reinforcement learning with dual rewards (pixel-level SSIM and semantic CLIP similarity) for high visual quality, and multimodal reasoning that jointly considers corrupted and recovered images.
Experiments on R-Bench show Robust-U1 achieves state-of-the-art robustness (overall 0.7398 vs.
0.5770 for BAGEL and 0.5017 for Robust-R1).
On adversarial corruptions of MMMB, MMStar, and RealWorldQA, it maintains superior performance with minimal drops (e.g., MMMB 100% corruption: 83.18 vs.
78.48 for BAGEL).
Ablations confirm that high-quality visual recovery directly enhances reasoning, and that both rewards are essential.
The work establishes self-recovery as a critical mechanism for robust visual understanding.
02 From the paper
Abstract
Multimodal Large Language Models (MLLMs) have demonstrated remarkable success in visual understanding, yet their performance degrades significantly under real-world visual corruptions. While existing robustness enhancement approaches exist, they are limited: black-box feature alignment lacks interpretability, and white-box text-based reasoning cannot restore lost pixel-level details. This work investigates a fundamental research question: Can MLLMs recover corrupted visual content by themselves? To address this, we propose Robust-U1, a novel framework that equips MLLMs with explicit visual self-recovery capability for robust understanding. The approach comprises three core stages: supervised fine-tuning for initial reconstruction, reinforcement learning with dual rewards (pixel-level SSIM and semantic-level CLIP similarity) for aligning high visual quality, and multimodal reasoning that jointly considers both the corrupted input and the recovered image. Extensive experiments demonstrate that Robust-U1 achieves state-of-the-art robustness on the real-world corruption benchmark and maintains superior performance under adversarial corruptions on general VQA benchmarks. Analysis confirms that high-quality visual recovery directly enhances reasoning performance, establishing self-recovery as a critical mechanism for robust visual understanding. The source code is available at https://github.com/jqtangust/Robust-U1.