惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

V
V2EX
V
Vulnerabilities – Threatpost
MongoDB | Blog
MongoDB | Blog
cs.AI updates on arXiv.org
cs.AI updates on arXiv.org
P
Proofpoint News Feed
Know Your Adversary
Know Your Adversary
aimingoo的专栏
aimingoo的专栏
Threat Intelligence Blog | Flashpoint
Threat Intelligence Blog | Flashpoint
C
Cisco Blogs
C
CERT Recently Published Vulnerability Notes
T
Tor Project blog
A
Arctic Wolf
L
LangChain Blog
L
LINUX DO - 热门话题
G
Google Developers Blog
Google DeepMind News
Google DeepMind News
T
Threat Research - Cisco Blogs
Stack Overflow Blog
Stack Overflow Blog
I
Intezer
爱范儿
爱范儿
P
Palo Alto Networks Blog
WordPress大学
WordPress大学
H
Hackread – Cybersecurity News, Data Breaches, AI and More
T
The Blog of Author Tim Ferriss
G
GRAHAM CLULEY
S
Securelist
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
Cisco Talos Blog
Cisco Talos Blog
Security Latest
Security Latest
Martin Fowler
Martin Fowler
AWS News Blog
AWS News Blog
L
Lohrmann on Cybersecurity
C
Cybersecurity and Infrastructure Security Agency CISA
酷 壳 – CoolShell
酷 壳 – CoolShell
Recorded Future
Recorded Future
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
C
CXSECURITY Database RSS Feed - CXSecurity.com
Recent Announcements
Recent Announcements
有赞技术团队
有赞技术团队
Apple Machine Learning Research
Apple Machine Learning Research
V2EX - 技术
V2EX - 技术
CTFtime.org: upcoming CTF events
CTFtime.org: upcoming CTF events
L
LINUX DO - 最新话题
博客园 - Franky
P
Privacy & Cybersecurity Law Blog
Simon Willison's Weblog
Simon Willison's Weblog
W
WeLiveSecurity
Cyberwarzone
Cyberwarzone
The Hacker News
The Hacker News
A
About on SuperTechFans

博客园 - fariver

[PaperWritting] 多模态大模型架构图摘录 每日Paper - 2026-03-06 每日Paper | 2026年3月4日 [Paper Reading] Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking [PaperReading] OneSearch A Preliminary Exploration of the Unified End-to-End Generative Framework for E-commerce Search [PaperReading] OneRec Technical Report [Paper Reading] Tiger: Recommender Systems with Generative Retrieval [PaperReading] Generative Recommendation with Semantic IDs: A Practitioner’s Handbook [PaperReading] GME: Improving Universal Multimodal Retrieval by Multimodal LLMs [Paper Reading] UniME-V2: MLLM-as-a-Judge for Universal Multimodal Embedding Learning [PaperReading] UniME: Breaking the Modality Barrier: Universal Embedding Learning with Multimodal LLMs [PaperReading] Qwen2.5-VL Technical Report [PaperReading] DeepSeek-OCR: Contexts Optical Compression [PaperReading] SAIL-Embedding Technical Report: Omni-modal Embedding Foundation Model [PaperReading] VLM2VEC: TRAINING VISION-LANGUAGE MODELS FOR MASSIVE MULTIMODAL EMBEDDING TASKS [PaperReading] REACT: SYNERGIZING REASONING AND ACTING IN LANGUAGE MODELS [PaperReading] MemGPT: Towards LLMs as Operating Systems [PaperReading] Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution [PaperReading] Mind Search: Mimicking Human Minds Elicits Deep AI Searcher [PaperReading] METAGPT: META PROGRAMMING FOR A MULTI-AGENT COLLABORATIVE FRAMEWORK
[PaperReading] VLM2Vec-V2: Advancing Multimodal Embedding for Videos, Images, and Visual Documents
fariver · 2025-10-17 · via 博客园 - fariver

VLM2Vec-V2: Advancing Multimodal Embedding for Videos, Images, and Visual Documents

link
时间:2025.07
单位:Salesforce Research等
相关领域:多模态表征学习
作者相关工作:Rui MengZiyan Jiang
被引次数:1705
项目主页:https://tiger-ai-lab.github.io/VLM2Vec/

TL;DR

之前的方法主要关注image与text,模态种类太少,未使用到视频与视觉文档模态信息,VLM2VecV2解决了该问题。同时,扩充了MMEB Benchmark,也设计了新的统一模型架构VLM2VecV2。

Method

Q:VLM2Vec-V2与原始VLM2Vec算法有什么区别?

  1. VLM2Vec仅支持文本与图像两种模态,而V2支持了更多,包括:视频、PDF、网页截图、语音(预留接口);
  2. Data Sampling Strategies:
  • on-the-fly batch mixing: 为每个数据集分配特定采样概率;
  • Interleaved Sub-batching: 将大批次划分为更小的​​子批次,每个子批次独立从单一数据源采样;
  1. multi-modal data formatting
    \(q_{inst}\)=inst+[VISUAL_TOKEN]+q
    inst: 构造出来的指令,例如,"Find a video that contains this image
    VISUAL_TOKEN:用来声明后续q的类型,具体值可以是 image_pad声明“图像”、video_pad声明“视频” 以及 无 声明“文本”

Benchmark

MMEB-V2 Benchmark与MMEB有什么区别? => 蓝色为V1的,红色为V2相对于V1扩充的。
image

Q&A

Q:CLS, QA, RET, GD, Overall这些指标是什么含义?

  • CLS指标​​指导特征学习优化
  • RET指标​​指导嵌入空间对齐
  • ​QA指标​​指导语义理解深化
  • GD指标​​指导细粒度感知提升

Q:训练使用了多少训练集?

Experiment

image

Ablation Study: 增加更多视觉模态是能继续涨点的
image

sub-size超参数的Ablation,最佳值为64
image

image

参考链接

总结与思考

相关链接

https://zhuanlan.zhihu.com/p/1932043359238726091