A Task-Centric Survey of Long-Document Visual Question Answering: Taxonomy, Evolution, and Future Directions
Authors: Yufei Tong, Guanjie Cheng, Tianle Wu, Leilei Zhang, Hang Zheng, Peihan Wu, Jianwei Yin, Shuiguang Deng
Organizations: College of Computer Science and Technology, Zhejiang University, China · School of Software Technology, Zhejiang University, China · Polytechnic Institute, Zhejiang University, China · BWTON Technology Co., Ltd., China
Visual document understanding is essential for acquiring knowledge from visually rich documents and is attracting growing research interest. Existing surveys emphasize architectures or technical approaches, with limited attention to how task bottlenecks shape technical choices and research priorities. We present a task-centric survey that decomposes long-document visual question answering (VQA) into four components: document parsing, evidence organization and retrieval, generator organization and learning, and evaluation. For each, we develop a fine-grained taxonomy and trace motivations, technical responses, and cross-component connections. We focus on explicit retrieval while covering direct reading. Our temporal analysis traces expanding priorities from multimodal content understanding to sufficient, high-fidelity evidence acquisition and sustained evidence management. We outline future directions in evidence sufficiency, verifiable and correctable understanding, and budget-aware coordination, with an outlook on continual document understanding.
Figures & tables
Figure 1: Core challenges of long-document VQA and publication trends of this field.
Figure 2: Task decomposition and two technical routes for long-document VQA. Colored bands indicate component functions, and italic labels identify the challenges they address. Both routes share parsing and generation, with evaluation throughout. Evidence sources and abstention are task-dependent outputs.
Method
Parsing
Evidence Organization and Retrieval
Generation & Learning
Name
Venue
OCR
Deep
Modality
Gran.
Mechanism
Acquisition
Agent
Reasoning
Learning
Retrieval-only methods
DSE ( 2024a ) (24.06)
EMNLP’24
✗
✗
Uni(V)
P
Visual t
One-shot
–
–
–
ColPali ( 2025 ) (24.06)
ICLR’25
✗
✗
Uni(V)
P
Visual t
One-shot
–
–
–
LMS-Retrieval ( 2026 ) (26.08)
ICDAR’26
✓
Tree
Uni(T)
E
Text t
One-shot
–
–
–
CMDR-Embed ( 2026 ) (26.07)
ECCV’26
✗
✗
Uni(V)
P
Visual t
One-shot
–
–
–
Table 1: Representative explicit-retrieval methods. (YY.MM): first release. Within task groups, methods follow final conference submission deadlines, then first release within each conference; preprints/journal papers are inserted by first release. Venue omits tracks. Full details: Appendix C , Table 6 .
Method
Venue
OCR
Model
Learning
Single-page VQA
Donut ( 2022 ) (21.11)
ECCV’22
✗
BART
PT+SFT
LayoutLMv3 ( 2022 ) (22.04)
MM’22
✓
RoBERTa
PT+SFT
Pix2Struct ( 2023 ) (22.10)
ICML’23
✗
Pix2Struct
PT+SFT
DocVLM* ( 2025 ) (24.12)
CVPR’25
✓
Qwen-VL
SFT
Multi-page VQA
Table 2: Taxonomy of representative direct-reading methods. Full comparison: Appendix B , Table 5 .
Figure 3: Four stages of document parsing: bottlenecks in black, technical responses in teal.
Figure 4: Four stages of evidence organization and retrieval: bottlenecks in black, technical responses in blue.
Figure 5: Four stages of generation and learning: bottlenecks in black, technical responses in purple.
Benchmark
Data Size
Question Type
Evidence and Evaluation
Name
Venue
Domain
#Pages
#QA
MD
MP
SP
Unans.
Source
Gran.
Evaluation
Retrieval-focused benchmarks
ViDoRe ( 2025 ) (24.06)
ICLR’25
Open
1.0k
3.81k
✗
✗
✓
✗
TXT/L/C/TAB/I
P
Ans. Ret.
REAL-MM-RAG ( 2025 ) (25.02)
ACL’25
Open
2.1k
4.6k
✗
✗
✓
✗
TXT/C/TAB/I
P
Ans. Ret.
ViDoRe V2 ( 2025 ) (25.03)
arXiv’25
Open
1.1k
327
✓
✓
✓
✗
TXT/L/C/TAB/I
P
Ret.
MMDocIR ( 2025 ) (25.01)
EMNLP’25
Closed
65.1
1.66k
✗
✓ (18.9%)
✓ (81.1%)
✗
TXT/L/C/TAB/I
P, E
Ret. Loc.
Table 3: Representative benchmarks grouped by evaluation goal and input scale. Full details: Appendix A , Table 4 .
Figure 6: Four stages of benchmarks and evaluation: bottlenecks in black, technical responses in amber.
Figure 7: Research foci and technical evolution across four components of document VQA. Five periods span three overarching foci: content understanding, evidence acquisition, and sustained evidence management.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Challenges and future directions for answering individual questions in long-document VQA. The three panels correspond to Secs. D.1 – D.3 , covering evidence sufficiency, verifiable and correctable understanding, and budget-aware coordination toward reliable answers under limited budgets.
Figure 9: Two complementary directions from individual question answering toward continual document understanding. Across questions, systems reuse, supplement, and revise evidence and interaction memory. Across tasks, they propose strategies from experience, validate them on independent tasks before updating agents and skills, and continue accumulating experience.
Benchmark
Data Size
Question Type
Evidence and Evaluation
Name
Venue
Domain
#Docs
#Pages
#Tokens
#QA
MD
MP
SP
Unans.
Source
Gran.
Eval.
Retrieval-focused benchmarks
Wiki-SS ( 2024a ) (24.06)
EMNLP’24
Open
–
1.27M
–
52.7k
✗
✗
✓
✗
TXT
Page
Ans. Ret.
SlideVQA-open ( 2024a ) (24.06)
EMNLP’24
Open
–
50.7k
–
12.4k
✗
✓
✓
✗
TXT/L/C/TAB/I
Page
Ret.
ViDoRe ( 2025 ) (24.06)
ICLR’25
Open
–
1.0k
–
3.81k
✗
✗
✓
✗
TXT/L/C/TAB/I
Page
Ans. a Ret.
REAL-MM-RAG ( 2025 ) (25.02)
ACL’25
Open
163
2.1k
–
4.6k
✗
✗
✓
✗
TXT/C/TAB/I
Page
Ans. Ret.
Appendix
Table 4: Complete benchmark catalog: 55 records covering dataset size, question types, evidence annotations, and evaluation capabilities. Ordering and first-release dates follow Table 1 . – denotes unreported or inapplicable information.
Method
Parsing
Evidence Organization and Retrieval
Generation & Learning
Eval.
Name
Venue
OCR
Deep
Modality
Gran.
Mechanism
Acq.
Agent
Reasoning
Model
Learning
Retrieval-only methods
DSE ( 2024a ) (24.06)
EMNLP’24
✗
✗
Uni(V)
Page
DSE t (P V )
One-shot
–
–
–
–
Ret. Eff.
ColPali ( 2025 ) (24.06)
ICLR’25
✗
✗
Uni(V)
Page
ColPali t (P V )
One-shot
–
–
–
–
Ret. Eff.
RobTabCol ( 2025 ) (25.02)
ACL’25
✗
✗
Uni(V)
Page
RobTabCol t (P V )
One-shot
–
–
–
–
Ret. Rob.
Light-ColPali ( 2025b ) (25.06)
ACL’25 F
✗
✗
Uni(V)
Page
ColPali t (P V )
One-shot
–
–
–
–
Ret. Eff.
Appendix
Table 6: Complete comparison of methods with explicit retrieval: 20 retrieval-only and 90 document VQA methods. Ordering and date conventions follow Table 1 ; venue superscripts follow Table 4 . Mechanism specifies retrievers, evidence mappings, and retrieval training; Model gives the main generator model family. + denotes combinations and / alternatives; – denotes unreported, unconfirmed, or inapplicable information.
School of Software Engineering, Huazhong University of Science and Technology · MiLM Plus, Xiaomi Inc. · School of Computer Science and Technology, Huazhong University of Science and Technology +1