[论文] Latent Reasoning in Continuous Space for Unified Multimodal Models

论文概要

研究领域: NLP 作者: Byungwoo Jeon, Yoonwoo Jeong, Hyunseok Lee, Minsu Cho, Jinwoo Shin 发布时间: 2026-10-30 arXiv: 2610.42149

中文摘要

统一多模态模型（UMMs）在多种多模态理解与生成任务上取得了显著成功，但它们在有效利用基于语言的思维链推理进行图像生成方面仍显吃力。因此，现有的推理增强型UMM往往无法继承大语言模型中观察到的测试时扩展能力和细粒度可控性。我们将这种失败归因于一个根本性的模态不匹配：推理在离散的token空间中进行，而图像生成在连续空间中操作。为弥合这一差距，我们引入了连续空间中的潜在推理框架（LARC），它通过在连续隐状态上进行推理来缓解token级瓶颈。具体而言，该训练框架包含两个阶段：(i) 课程式监督微调（SFT）；(ii) 信息增益强化学习（RL）。首先，课程SFT阶段逐步将token级推理转化为潜在推理；随后，信息增益RL阶段在没有真实标签监督的情况下自我进化潜在推理。在多种文本到图像生成基准上的实验表明，LARC持续提升生成质量、提示遵循度和组合保真度。值得注意的是，LARC在GenEval上取得了最先进的性能，比BAGEL提升了7%p，并展现出清晰的测试时扩展行为，在并行和顺序测试时扩展策略下均优于基于token的生成。

原文摘要

Unified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models. We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space. To bridge this gap, we introduce a Latent Reasoning framework in Continuous space (LARC), which alleviates the token-level bottleneck by reasoning over continuous hidden states. Specifically, the training framework consists of two stages: (i)...

#论文 #arXiv #NLP #小凯