arXiv cs.AI· Jaeyong Ko, Hankyeol Kim, Pilsung Kang·· 14 小时前AI 评分38
Oddin：统一多模态模型通过图像生成自蒸馏实现潜在视觉推理
Oddin: Latent Visual Reasoning in Unified Multimodal Models via Self-Distillation from Image Generation
AI 导读
Oddin 提出一种在统一大语言多模态模型（ULMMs）中实现潜在视觉推理的框架，用紧凑的 latent token 替代中间图像、全部交由理解分支处理，生成分支只负责输出最终图像。模型通过离线自蒸馏训练，结合最终图像生成损失、文本蒸馏损失和特殊 token 的交叉熵损失三项目标。在七个 VBVR-Pro 视觉推理任务上，Oddin 平均表现超过交错生成基线和闭源 VLM。
正文 · 原文
Abstract:Unified large multimodal models (ULMMs) are able to perform visual reasoning through interleaved text-image generation, but generating explicit intermediate images incurs significant computational overhead. While visual latent reasoning addresses this, most existing methods rely on vision-language models (VLMs) incapable of visual generation. We propose Oddin, a novel framework enabling latent visual reasoning in ULMMs. Oddin replaces intermediate images with compact latent tokens processed entirely by the understanding branch, reserving the generation branch for the final output image. We train the model via offline self-distillation with three objectives, final image generation losses, text distillation losses, and cross-entropy losses for special tokens. Oddin surpasses interleaved generation baselines and closed-source VLMs on average across seven VBVR-Pro visual reasoning tasks. Qualitative analysis reveals that latent tokens attend to relevant regions in an input image, demonstrating that ULMMs can perform complex visual reasoning without generating intermediate images.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.33062 [cs.CV] |
| (or arXiv:2610.33062v1 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33062 arXiv-issued DOI via DataCite |
Submission history
From: Hankyeol Kim [view email]
[v1]
Mon, 19 Oct 2026 14:07:28 UTC (1,010 KB)
来源：arXiv cs.AI · arxiv.org