What Generation Knows: Unlocking Fine-Grained Understanding in Unified Multimodal Models

One-line takeaway

This paper systematically analyzes what generation-side representations in unified multimodal models offer for fine-grained visual understanding, finding them more sensitive to local attributes and spatial relations, and uses native generation-side features directly to strengthen understanding rather than only as auxiliary supervision.

这篇论文系统分析统一多模态模型里“生成侧”表示对细粒度视觉理解的作用，发现它们对局部属性和空间关系更敏感，并直接利用原生生成侧表示来补强理解，而不是只把生成当作辅助监督。