视觉大模型 / VLM - arXivDaily 专题

2411.11706 2026-02-19 cs.CV cs.AI 91%

MC-LLaVA: Multi-Concept Personalized Vision-Language Model

MC-LLaVA：多概念个性化视觉-语言模型

Ruichuan An, Sihan Yang, Renrui Zhang, Ming Lu, Tianyi Jiang, Kai Zeng, Yulin Luo, Jiajun Cao, Hao Liang, Ying Chen, Qi She, Shanghang Zhang, Wentao Zhang

机构 * Peking University（北京大学）； Intel Labs, China（英特尔实验室，中国）

专题命中 VLM训练与架构：vision-language model(title,abstract)；LLaVA(title,abstract)；VLM(abstract)；visual question answering(abstract)

AI总结 MC-LLaVA通过多概念指令微调和个性化提示提升视觉-语言模型的多概念个性化能力。

2602.16590 2026-02-19 cs.CV cs.AI cs.LG 67%

A Contrastive Learning Framework Empowered by Attention-based Feature Adaptation for Street-View Image Classification

一种由基于注意力的特征适应增强的对比学习框架用于街景图像分类

Qi You, Yitai Cheng, Zichao Zeng, James Haworth

机构 * SpaceTimeLab（时空实验室）； University College London（伦敦大学学院）

专题命中 VLM训练与架构：vision-language model(abstract)；分类 cs.CV、cs.AI、cs.LG

AI总结 CLIP-MHAdapter通过引入基于注意力的特征适应机制，提升街景图像属性分类的准确性和效率

2602.16502 2026-02-19 cs.CV 57%

DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images

DressWild: 从真实场景图像生成无需迭代优化的前馈姿态无关服装缝纫图案

Zeng Tao, Ying Jiang, Yunuo Chen, Tianyi Xie, Huamin Wang, Yingnian Wu, Yin Yang, Abishek Sampath Kumar, Kenji Tashiro, Chenfanfu Jiang

机构 * UCLA and Fudan University（UCLA和复旦大学）； University of Utah（犹他大学）

专题命中 VLM训练与架构：vision-language model(abstract)；分类 cs.CV

AI总结 DressWild通过前馈方法从单张真实图像生成物理一致的2D缝纫图案和3D服装，无需多视角输入或迭代优化，实现高效可扩展的服装模拟。

2602.15903 2026-02-19 cs.CV 57%

Detecting Deepfakes with Multivariate Soft Blending and CLIP-based Image-Text Alignment

利用多变量软融合和基于CLIP的图像-文本对齐检测深度伪造

Jingwei Li, Jiaxin Tong, Pengfei Wu

机构 * Zhejiang Gongshang University（浙江工商大学）

专题命中 VLM训练与架构：vision-language model(abstract)；分类 cs.CV

AI总结本文提出MSBA-CLIP框架，通过多变量软融合和CLIP引导的伪造强度估计，提升深度伪造检测的准确性和泛化能力。

2602.12281 2026-02-19 cs.RO cs.AI cs.SY eess.SY 57%

Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment

在视觉-语言-动作对齐中，验证比策略学习更具效率

Jacky Kwok, Xilun Zhang, Mengdi Xu, Yuejiang Liu, Azalia Mirhoseini, Chelsea Finn, Marco Pavone

机构 * Stanford University（斯坦福大学）； NVIDIA Research（NVIDIA研究）

专题命中 VLM训练与架构：VLM(abstract)；分类 cs.AI

AI总结本文提出CoVer-VLA验证方法，通过测试时验证在视觉-语言-动作对齐中实现22%的在分布和13%的分布外收益，同时在现实世界实验中进一步提升45%。

2602.16584 2026-02-19 q-bio.NC 50%

The Representational Alignment Hypothesis: Evidence for and Consequences of Invariant Semantic Structure Across Embedding Modalities

表征对齐假说：跨嵌入模态的不变语义结构的证据及其后果

Akhil Ramidi, Kevin Scharp

专题命中 VLM训练与架构：grounding(abstract)

AI总结该研究提出表征对齐假说，探讨跨模态嵌入的不变语义结构，质疑其根本性，并提出新的哲学视角和区分方法。

Comments 23 pages, 3 figures

1. VLM训练与架构 6 篇

MC-LLaVA: Multi-Concept Personalized Vision-Language Model

A Contrastive Learning Framework Empowered by Attention-based Feature Adaptation for Street-View Image Classification

DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images

Detecting Deepfakes with Multivariate Soft Blending and CLIP-based Image-Text Alignment

Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment

The Representational Alignment Hypothesis: Evidence for and Consequences of Invariant Semantic Structure Across Embedding Modalities