arXiv cs.CL· Byungwoo Jeon, Yoonwoo Jeong, Hyunseok Lee, Minsu Cho, Jinwoo Shin·· 11 小时前AI 评分38
LARC：面向统一多模态模型的连续空间隐式推理框架
Latent Reasoning in Continuous Space for Unified Multimodal Models
AI 导读
研究者提出连续空间隐式推理框架 LARC，让统一多模态模型在连续隐藏状态上推理，而非离散 token 空间，以解决语言链式推理难以迁移到图像生成的模态错配问题。训练分课程式 SFT 与信息增益 RL 两阶段，后者无需真值监督即可自我演化。在多个文生图基准上，LARC 在 GenEval 上较 BAGEL 提升 7%p 并达到 SOTA，且在并行与串行测试时扩展策略下均优于基于 token 的生成。
正文 · 原文
Abstract: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) curriculum supervised fine-tuning (SFT), and (ii) information-gain reinforcement learning (RL). First, the curriculum SFT stage gradually converts token-level reasoning into latent reasoning. Subsequently, the information-gain RL stage self-evolves latent reasoning without ground truth supervision. Experiments across diverse text-to-image generation benchmarks show that LARC consistently improves generation quality, prompt following, and compositional fidelity. Notably, LARC achieves state-of-the-art performance on GenEval with a 7%p gain over BAGEL and demonstrates clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies.
| Comments: | Published as conference proceeding for NeurIPS 2026. Last two authors advised equally |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.42149 [cs.CV] |
| (or arXiv:2610.42149v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.42149 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Byungwoo Jeon [view email]
[v1]
Fri, 30 Oct 2026 17:59:43 UTC (1,098 KB)
来源：arXiv cs.CL · arxiv.org