Survey Papers¶

Multimodal Document AI Learning Center

Module: Survey Papers

Version: v4.0

Last Updated: July 2026

本目录整理 Multimodal Document AI 方向最值得阅读的综述论文（Survey Papers）。

建议所有研究生优先阅读综述，再阅读经典论文和最新论文，建立完整的知识体系。

Learning Roadmap¶

Foundation Models
        │
        ▼
Vision-Language Models
        │
        ▼
Document AI
        │
        ▼
Document Parsing
        │
        ▼
DocTags / Docling
        │
        ▼
Benchmarks
        │
        ▼
AI Agent

Reading Strategy¶

Priority	Description
⭐⭐⭐⭐⭐	必读（建议精读）
⭐⭐⭐⭐	建议阅读
⭐⭐⭐	作为扩展阅读

Part 1. Vision-Language Models¶

1. A Comprehensive Survey of Vision-Language Models¶

Priority

⭐⭐⭐⭐⭐

Year

2026

Journal

Information Fusion

Official Paper¶

https://www.sciencedirect.com/science/article/pii/S1566253525006955

Why Read¶

目前最系统的 VLM 综述之一，覆盖：

Vision Encoder
LLM Backbone
Instruction Tuning
Prompt Engineering
Adapter / LoRA
Benchmark
Efficient VLM

非常适合作为实验室学习 VLM 的第一篇综述。

2. Benchmark Evaluations, Applications, and Challenges of Large Vision Language Models¶

Priority

⭐⭐⭐⭐⭐

Year

2025

Paper¶

https://arxiv.org/abs/2501.02189

GitHub¶

https://github.com/zli12321/Awesome-VLM-Papers-And-Models

Why Read¶

重点介绍：

主流 VLM
Benchmark
Evaluation
Applications
Challenges

推荐作为 VLM Benchmark 的参考文献。

3. A Survey on Efficient Vision-Language Models¶

Priority

⭐⭐⭐⭐

Year

2025

Paper¶

https://arxiv.org/abs/2504.09724

GitHub¶

https://github.com/MPSCUMBC/Efficient-Vision-Language-Models-A-Survey

Why Read¶

重点关注：

Mobile VLM
Edge AI
Efficient Inference
Parameter-efficient Fine-tuning

适合关注轻量化 VLM 的学生。

Part 2. Document AI¶

4. Document Intelligence in the Era of Large Language Models¶

Priority

⭐⭐⭐⭐⭐

Year

2025

Paper¶

https://arxiv.org/abs/2510.13366

Why Read¶

近年来 Document AI 最值得阅读的综述之一。

重点讨论：

LLM for Document AI
Multimodal Document Understanding
Document Parsing
Retrieval-Augmented Document AI
Agent-based Document Intelligence

非常契合未来实验室研究方向。

5. Deep Learning based Visually Rich Document Content Understanding¶

Priority

⭐⭐⭐⭐⭐

Year

2026

Journal

Artificial Intelligence Review

Paper¶

https://link.springer.com/article/10.1007/s10462-025-11477-3

Why Read¶

系统总结：

Layout Analysis
OCR
KIE
Table Recognition
Document Parsing
Vision-Language Methods

是目前最新的 VRD（Visually Rich Documents）综述之一。

Part 3. Docling & Structured Documents¶

6. Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion¶

Priority

⭐⭐⭐⭐⭐

Year

2025

Why Read¶

Docling 是当前最重要的开源文档转换框架。

建议重点理解：

Unified Document Representation
DoclingDocument
DocTags
Modular Pipeline
RAG Integration

7. SmolDocling¶

Priority

⭐⭐⭐⭐⭐

Year

2025

Paper¶

https://arxiv.org/abs/2503.11576

Official Papers¶

https://www.docling.ai/papers/

Why Read¶

SmolDocling 首次提出：

End-to-End Document Conversion
DocTags Generation
Ultra-compact Document VLM

代表新一代 Document Foundation Model。

Part 4. Benchmark Surveys¶

8. OCRBench v2¶

Priority

⭐⭐⭐⭐⭐

Paper¶

https://arxiv.org/abs/2501.00321

GitHub¶

https://github.com/Yuliang-Liu/MultimodalOCR

Why Read¶

了解：

OCR Evaluation
Visual Grounding
OCR Reasoning
LMM Evaluation

是当前 OCR Benchmark 的代表工作。

9. MMDocBench¶

Priority

⭐⭐⭐⭐⭐

Paper¶

https://arxiv.org/abs/2410.21311

Project¶

https://mmdocbench.github.io/

Why Read¶

重点理解：

OCR-free Evaluation
Fine-grained Grounding
Multi-task Evaluation
Vision-Language Benchmark

适合研究 Qwen3-VL、InternVL 等模型。

Part 5. Recommended Resource Collections¶

Awesome Lists¶

Resource	URL
Awesome Vision-Language Models	https://github.com/zli12321/Awesome-VLM-Papers-And-Models
Docling Papers	https://www.docling.ai/papers/
Papers with Code	https://paperswithcode.com
Hugging Face Papers	https://huggingface.co/papers

Conference Proceedings¶

Conference	URL
CVPR Open Access	https://openaccess.thecvf.com
ICCV Open Access	https://openaccess.thecvf.com
ECCV Proceedings	https://eccv.ecva.net
ACL Anthology	https://aclanthology.org
NeurIPS Proceedings	https://papers.nips.cc
ICML Proceedings	https://proceedings.mlr.press

Laboratory Recommended Reading Order¶

建议实验室统一采用如下阅读路线：

Survey Papers
      │
      ▼
Foundation Models
      │
      ▼
Vision-Language Models
      │
      ▼
Document AI
      │
      ▼
Docling
      │
      ▼
DocTags
      │
      ▼
Benchmarks
      │
      ▼
Research Projects

Learning Objectives¶

完成本章节后，应能够：

系统理解 Vision-Language Model 的发展脉络；
掌握 Document AI 的主要研究方向与关键挑战；
理解 Docling、DocTags 等新一代结构化文档技术；
熟悉主流 Benchmark 与评价体系；
能够快速定位高质量论文，并独立阅读 CVPR、ICCV、ECCV、ACL、EMNLP、ICML、NeurIPS 等顶级会议的最新研究成果。

返回上一级下一章

Week	Reading
Week 1	VLM Survey（Paper 1）
Week 2	Efficient VLM Survey（Paper 3）
Week 3	Document AI Survey（Paper 4）
Week 4	VRD Survey（Paper 5）
Week 5	Docling（Paper 6）
Week 6	SmolDocling（Paper 7）
Week 7	OCRBench v2（Paper 8）
Week 8	MMDocBench（Paper 9）

Survey Papers¶

Learning Roadmap¶

Reading Strategy¶

Part 1. Vision-Language Models¶

1. A Comprehensive Survey of Vision-Language Models¶

Official Paper¶

Why Read¶

2. Benchmark Evaluations, Applications, and Challenges of Large Vision Language Models¶

Paper¶

GitHub¶

Why Read¶

3. A Survey on Efficient Vision-Language Models¶

Paper¶

GitHub¶

Why Read¶

Part 2. Document AI¶

4. Document Intelligence in the Era of Large Language Models¶

Paper¶

Why Read¶

5. Deep Learning based Visually Rich Document Content Understanding¶

Paper¶

Why Read¶

Part 3. Docling & Structured Documents¶

6. Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion¶

Paper¶

IBM Research¶

Official Documentation¶

GitHub¶

Why Read¶

7. SmolDocling¶

Paper¶

Official Papers¶

Why Read¶

Part 4. Benchmark Surveys¶

8. OCRBench v2¶

Paper¶

GitHub¶

Why Read¶

9. MMDocBench¶

Paper¶

Project¶

Why Read¶

Part 5. Recommended Resource Collections¶

Awesome Lists¶

Conference Proceedings¶

Recommended Reading Plan¶

Laboratory Recommended Reading Order¶

Learning Objectives¶