ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models

查看原文 / View Original 编辑 / Edit

Computer Science > Computer Vision and Pattern Recognition arXiv:2610.20781 (cs) [Submitted on 14 Oct 2026] Title:ReasonUni: Reasoning-Augmented Reconstruction Advances Unified Multimodal Models Authors:Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan, Li Yuan.

计算机科学 > 计算机视觉与模式识别 arXiv:2610.20781 (cs) [提交于 2026 年 10 月 14 日] 标题：ReasonUni：推理增强重构推进统一多模态模型作者：Feize Wu, Zhiyuan Yan, Yuhuan Hong, Yuwei Niu, Zongjian Li, Zheyuan Liu, Bin Lin, Jinbo Yan, Jie Qin, Xudong Mao, Limeng Qiao, Guanglu Wan, Li Yuan。

Abstract: Image-to-text (I2T) understanding and text-to-image (T2I) generation are dual tasks in unified multimodal models (UMMs). Reconstructive post-training couples these tasks, but existing reconstruction-based methods remain perception-centric and provide limited supervision for reasoning about structure, intent, and transformation.

摘要：图像到文本（I2T）理解和文本到图像（T2I）生成是统一多模态模型（UMM）中的对偶任务。重构后训练将这些任务耦合在一起，但现有的基于重构的方法仍然以感知为中心，且在推理结构、意图和转换方面提供的监督有限。

We propose ReasonUni, a reasoning-augmented reconstructive post-training framework that extends this coupling beyond perception. Given an original image and an editing instruction, the UMM first reasons about what and how to edit and describes the target state, then generates the corresponding target image conditioned on this understanding output.

我们提出了 ReasonUni，这是一个推理增强的重构后训练框架，它将这种耦合扩展到了感知之外。给定一张原始图像和编辑指令，UMM 首先推理出编辑的内容和方式并描述目标状态，然后根据此理解输出生成相应的目标图像。

To jointly optimize both modules, we introduce TreeGRPO, a tree-structured reinforcement learning method that samples multiple reasoning paths and candidate images per path, then aggregates visual rewards to update both modules under a unified objective. Experiments across multiple benchmarks demonstrate consistent gains in both understanding and generation, showing that reasoning-augmented reconstruction yields genuine mutual gains for UMMs.

为了联合优化这两个模块，我们引入了 TreeGRPO，这是一种树状结构的强化学习方法，它为每条路径采样多个推理路径和候选图像，然后聚合视觉奖励，在统一目标下更新这两个模块。在多个基准测试上的实验证明了在理解和生成方面的一致性提升，表明推理增强重构为 UMM 带来了真正的互惠收益。