
ICML 2026:视觉自恢复 + 双奖励强化学习,提升受损图像理解
Quick Answer
Robust-U1 enhances multimodal understanding by integrating visual self-recovery and dual-reward reinforcement learning, significantly improving performance on R-Bench with a score of 0.7398, compared to 0.5770 for BAGEL.
Quick Take
This approach allows models to generate and utilize restored images for better reasoning in degraded visual contexts.
Key Points
- Robust-U1 utilizes a unified for visual recovery and reasoning.
- Achieved a PSNR increase from 14.37 to 21.49, indicating improved image quality.
- The model's dual-reward system includes pixel structure and semantic consistency checks.
- Performance remains robust even under high degradation levels in visual tasks.
- Training required 1920 GPU hours for the recovery supervision phase.
Source Excerpt
原文作者:公众号“Today读什么”原文链接:https://mp. weixin. qq. com/s/BrsWJJAv22qHVa_gfv2cpg一张照片被压缩、噪声、暗光和模糊破坏后,多模态模型仍然可以写出一段逻辑完整的分析。 但分析越流畅,不代表它看到的证据越充分:车头朝向已经模糊,模型仍能解释车辆为何“直行”;公交车轮廓已经重叠,它依然可以自信地数出三辆。 过去的方法通常让视觉编码器适应噪声,或者让模型先用文字分析图像受到了什么破坏。 Robust-U1 换了一条路线:**既然关键视觉信息已经丢失,就先生成一张恢复后的图像,再让模型同时查看原图与恢复图完成回答。 **方法由图像恢复监督训练、带双重视觉奖励的强化学习,以及双图联合推理组成。 实验结果表明,性能提升并不只是来自“生成了一张更好看的图片”,而是恢复出的视觉证据确实帮助模型回答了原本看不清的问题。 MotivationRobust-U1 要处理的核心矛盾是:多模态模型可以继续推理,却无法仅凭语言补回已经消失的视觉细节。 视觉鲁棒性研究此前主要沿着两条路线展开。
特征对齐方法让干净图像和受损图像在视觉编码空间中尽量接近,能够提高预测稳定性,却很难解释模型究竟恢复了什么信息。 Robust-R1 这样的工作把过程显式化,让模型描述压缩、噪声或暗光如何影响画面,再依据描述进行推理。 文字能够提醒模型“这里可能看不清”,却无法重建车辆朝向、物体数量、边缘位置或细小文字。 视觉证据已经被破坏时,再长的推理链也可能只是围绕错误观察组织出一段连贯解释。 …
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from 雷峰网机器人
See more →
WAIC深度观察:具身数据,进入「开箱即用」时代
The WAIC forum highlights the urgent need for 'Model-Ready Data' in embodied intelligence, emphasizing high precision, multi-modality, and diversity as essential for effective physical interaction. Companies like JianZhi Robotics are pioneering solutions to bridge existing data gaps, ensuring that data is directly usable without further processing.

