PaperlayerView exact revision on arXiv

Source-linked AI summary

A Task-Centric Survey of Long-Document Visual Question Answering: Taxonomy, Evolution, and Future Directions

Yufei Tong, Guanjie Cheng, Tianle Wu, Leilei Zhang, Hang Zheng, Peihan Wu, Jianwei Yin, Shuiguang Deng

arXiv:2610.31385v1cs.CVcs.CLcs.IR

TL;DR

Long-document VQA must acquire and use sufficient evidence across visually rich, lengthy documents under resource constraints, while existing surveys provide limited analysis of how task bottlenecks shape technical choices and research priorities. This survey organizes the field into four task components, develops a fine-grained taxonomy, and traces technical evolution. It finds expanding priorities from multimodal understanding to sufficient, high-fidelity evidence acquisition and sustained evidence management, and outlines future directions.

  • Problem

    Existing surveys emphasize architectures or technical approaches, with limited attention to how long-document VQA bottlenecks shape technical choices and research priorities.

    Source: §1 · p. 2

  • Method

    The survey analyzes over 150 papers through four task components, develops a fine-grained taxonomy, and traces motivations, technical responses, and evolution while emphasizing explicit retrieval and covering direct reading.

    Source: §1 · p. 2, §5 · p. 9

  • Results

    Research priorities expand from multimodal understanding toward sufficient, high-fidelity evidence acquisition and sustained evidence management under limited budgets.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Future directions include evidence sufficiency assessment, verifiable and correctable understanding, budget-aware coordination, and memory reuse and strategy improvement toward continual document understanding.

    Source: §5 · p. 9

    The survey focuses on evidence acquisition and use in long-document VQA, covering foundational recognition and extraction tasks only where they affect question answering.

    Source: Limitations · p. 9

Abstract

from arXiv · show

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.

1 Introduction

Long-document VQA must acquire, understand, and verify sufficient evidence under limited budgets, while existing surveys give less attention to how task bottlenecks shape technical choices and research priorities. This survey analyzes more than 150 papers through a four-component task framework, traces technical evolution, and identifies future directions.

1 Introduction

  • Long-document VQA requires acquiring, understanding, and verifying sufficient evidence within limited budgets as documents grow and evidence spans pages or documents.Exhaustive reading raises context and computational costs, while textualization or visual compression can discard chart, layout, and fine-grained textual information.Source: §1 · pp. 1–2
  • Existing surveys largely organize work by architecture and technical approach, leaving connections among long-document bottlenecks, fine-grained technical analysis, and evolving research motivations underexplored.The paper analyzes over 150 papers to address this gap.Source: §1 · p. 2
  • The survey maps content fidelity, evidence acquisition and use, and reliable evaluation onto four interconnected components: parsing, evidence organization and retrieval, generator organization and learning, and evaluation.Its taxonomy categorizes representations and structure, retrieval modalities and granularity, acquisition strategies, reasoning organization and learning objectives, and evaluation capabilities.Source: §1 · p. 2
  • The survey emphasizes explicit retrieval while covering direct reading, and traces research priorities from multimodal understanding toward sufficient, high-fidelity evidence acquisition and sustained management.It identifies future directions in evidence sufficiency, verifiable and correctable understanding, budget-aware coordination, and continual document understanding.Source: §1 · p. 2

2 Task Formulation and Decomposition

Long-document VQA produces answers from visual documents under resource budgets, with evidence needs shaped by scale, visual complexity, and dependencies across documents. The paper frames the task through four interconnected components and two technical routes: direct reading and explicit retrieval.

2 Task Formulation and Decomposition

  • Long-document VQA takes a question and candidate visual documents and produces an answer under a budget on context, computation, and tool calls.Inputs may be closed-domain, with the document specified, or open-domain, with relevant sources selected from a shared collection.Source: §2 · p. 2
  • The task faces scale-driven search and context costs, visual information loss, and evidence dependencies across pages or documents that make local relevance insufficient.Pages preserve context and provenance, while elements offer finer-grained evidence; cross-page or cross-document relations can link evidence needed jointly.Source: §2 · pp. 2–3
  • Document parsing preserves visual details and structure, while evidence organization and retrieval selects sufficient evidence amid noise and passes it with its sources to the generator.These functions balance representation fidelity against processing cost and address search scale and evidence dispersion.Source: §2 · p. 3
  • Generator organization and learning integrate and verify evidence under context and computation limits, while intermediate judgments can guide acquisition, rereading, or verification.Evaluation assesses correctness, grounding, efficiency, and reliability, using evidence checks and controlled experiments to diagnose errors and compare benefits and costs.Source: §2 · pp. 3–4
  • Direct reading answers from supplied representations, whereas explicit retrieval selects relevant documents, pages, or elements for the generator to read.Reasoning feedback may guide further acquisition, rereading, or on-demand parsing; approximately 80% of surveyed methods use explicit retrieval, which motivates the paper's emphasis on that route while it also covers direct reading.Source: §2 · p. 4

3 Taxonomy and Technical Evolution

The survey organizes long-document VQA around four connected components and traces how their task bottlenecks shape technical choices and research priorities. Across components, priorities expand from multimodal understanding toward sufficient evidence acquisition, sustained evidence management, and evaluation of capability and cost.

3 Taxonomy and Technical Evolution

  • The survey maps content fidelity, sufficient evidence acquisition, evidence use, and reliable evaluation onto four interconnected components, then traces task-driven technical evolution.Its taxonomy covers document parsing, evidence organization and retrieval, generator organization and learning, and evaluation.Source: §3 · p. 4

3.1 Document Parsing

  • Parsing taxonomies distinguish OCR dependence from deep parsing, which builds semantic overviews or explicit relations such as summaries, trees, and graphs.Research progresses from readable content with layout preservation to original-page visual fidelity, navigable structures, and question- or task-guided parsing on demand.Source: §3.1 · pp. 4–5

3.2 Evidence Organization and Retrieval

  • Retrieval is characterized by matching modality, evidence granularity, retrieval mechanism, and whether feedback expands acquisition beyond the current evidence set.Pages preserve context but can add noise; elements emphasize detail but risk losing context. One-shot acquisition may include fixed multistage retrieval without feedback-driven expansion.Source: §3.2 · p. 5
  • Retrieval research moves from page retrieval toward structural links, fine-grained and iterative acquisition, then explicit evidence states, context, and budget-aware search.More evidence may add noise and redundancy without ensuring sufficient support.Source: §3.2 · pp. 5–6

3.3 Generation and Learning

  • Generator taxonomies distinguish single or multiple runtime agents, fixed or feedback-driven reasoning, and pretraining, supervised fine-tuning, or reinforcement learning.Feedback may acquire new evidence or reread and verify existing evidence without expanding acquisition.Source: §3.3 · p. 6
  • Generator research progresses from pretraining and task fine-tuning to long-input handling, evidence-guided feedback, and evidence-state management.Later approaches use memory, reward memory maintenance, or preserve provenance and reload source pages before reorganizing context.Source: §3.3 · p. 6

3.4 Benchmarks and Evaluation

  • Evaluation dimensions include answer correctness, evidence recall and ranking, spatial localization, efficiency, and robustness, with joint reporting relating gains to end-to-end costs.Benchmark comparisons also consider scale, evidence requirements, and supported evaluation capabilities.Source: §3.4 · pp. 6–7
  • Benchmarks expand from text, layout, and chart understanding to long-document integration, evidence and provenance checks, and joint capability diagnosis.Recent examples assess retrieval and QA together, diagnose performance under varying evidence conditions, or evaluate robustness and controlled document conditions.Source: §3.4 · p. 7

4 Research Trends and Future Directions

The survey traces document VQA’s shift from content understanding toward evidence acquisition and sustained management, then identifies future directions for individual questions and continual document use. It emphasizes coordination across components and across rounds of interaction.

4 Research Trends and Future Directions

  • The section connects component interactions to future directions at two levels: improving individual-question answering and extending the task toward continual document understanding.The appendix develops opportunities for both levels, including sustained interaction and accumulated experience.Source: §4 · pp. 7–8

4.1 Cross Component Evolution

  • Research priorities broadened from multimodal content understanding to sufficient, high-fidelity evidence acquisition and sustained evidence management.The five periods are grouped under three overarching foci: content understanding, evidence acquisition, and sustained evidence management.Source: §4.1 · p. 8, Figure 7 · p. 8
  • Retrieval can omit evidence, while feedback makes parsing and retrieval schedulable actions and generation can maintain evidence states to guide rereading, verification, and stopping.Evaluation should diagnose failures through both answers and evidence and assess gains against end-to-end cost; methods may update selected components rather than replace established techniques wholesale.Source: §4.1 · p. 8

4.2 Challenges and Future Directions

  • Future work for individual questions should assess evidence-set sufficiency, connect source tracing with process verification and error correction, and coordinate component decisions under budget constraints.Local relevance and precise localization may not ensure sufficient support; correct answers may mask unreliable intermediate judgments, and early errors can propagate.Source: §4.2 · p. 8
  • The outlook extends beyond independent questions to reusing evidence and interaction states and improving transferable strategies from cross-task experience.The survey frames these extensions as memory reuse across questions and continual, experience-driven strategy improvement.Source: §4.2 · p. 8

5 Conclusion

The survey connects a four-component account of long-document VQA with technical evolution and future research directions.

5 Conclusion

  • The survey links a four-component decomposition of long-document VQA and a fine-grained taxonomy to task bottlenecks, technical evolution, and future directions.It emphasizes explicit retrieval while also covering direct reading, and traces priorities toward sufficient, high-fidelity evidence acquisition and sustained management under limited budgets.Source: §5 · p. 9

Limitations

The survey focuses on higher-level document VQA, particularly evidence acquisition and use in long documents.

Limitations

  • Foundational recognition and extraction tasks are covered only where they affect question answering, rather than reviewed in detail.The scope includes text and layout recognition and field and entity extraction.Source: Limitations · p. 9

Use of AI-Assisted Tools

The authors used AI assistance for language editing and LaTeX formatting, then reviewed the content and accepted responsibility for the final manuscript.

Use of AI-Assisted Tools

A Benchmark Catalog

The catalog groups benchmarks by evaluation goal and input page count, and emphasizes that scale, evidence span, and processing stage shape what results measure.

A Benchmark Catalog

  • Benchmark selection should distinguish input scale, evidence span, and the stage at which processing begins.Page count alone does not establish relative difficulty, and open-corpus retrieval differs from answering over a given document.Source: A Benchmark Catalog · p. 18
  • MP-DocVQA tests single-page evidence selection among candidate pages, whereas SlideVQA requires evidence across pages.These benchmark settings illustrate why longer inputs do not necessarily require cross-page evidence combination.Source: A Benchmark Catalog · p. 18
  • SPIQA supplies pre-extracted elements, so its results primarily measure understanding and evidence selection over supplied material rather than full-page parsing.Evidence annotation granularity also determines whether retrieval, region localization, or answer quality is being verified.Source: A Benchmark Catalog · p. 18
  • The catalog covers 55 benchmark records, including long-input, cross-page, fine-grained acquisition, source-verification, visual-degradation, and evidence-condition evaluations.Cost comparisons should specify input conditions and budgets and account for parsing, indexing, and online inference, not answer scores alone.Source: A Benchmark Catalog · p. 18

B Methods without Explicit Retrieval

Direct-reading methods organize supplied content without explicit evidence retrieval, balancing input scale, information organization, and evidence use.

B Methods without Explicit Retrieval

  • Direct-reading methods organize supplied content without explicit evidence retrieval in the answering pipeline.Table 5 groups methods into single-page and multipage categories by primary design and evaluation settings.Source: B Methods without Explicit Retrieval · p. 18
  • Multipage methods use hierarchical representations, document tokens for cross-page communication, or compressed high-resolution visual features to control input scale.Hi-VT5, GRAM, and mPLUG-DocOwl2 respectively illustrate these design choices.Source: B Methods without Explicit Retrieval · p. 18
  • M3D and DocR1 strengthen the link between evidence identification and answering through evidence-and-answer supervision and evidence-page rewards, respectively.Direct reading allocates resources among input capacity, detail fidelity, and cross-page relations; compression can lose critical text or visual detail.Source: B Methods without Explicit Retrieval · p. 18
  • Comparisons should consider answer quality, evidence selection, and cost together, because broader page coverage and higher resolution increase reading cost.Direct reading and explicit retrieval are complementary: retrieval can control what evidence is passed to multipage reading.Source: B Methods without Explicit Retrieval · p. 18

C Methods with Explicit Retrieval

The survey catalogs explicit-retrieval methods and connects their parsing, retrieval, generation, and evaluation choices to future challenges in evidence sufficiency and reliable answering.

C Methods with Explicit Retrieval

  • Table 6 catalogs 110 explicit-retrieval methods, recording evidence representation, matching, returned units, generator pathways, agent organization, feedback, and training where applicable.Parsing, retrieval, generation, and evaluation attributes should be read together; the same retriever can support different generation pipelines.Source: C Methods with Explicit Retrieval · p. 18
  • Modality describes evidence-side representations used to assess relevance, while granularity describes the evidence units actually received by the generator.For retrieval-only methods, granularity refers to returned units.Source: C Methods with Explicit Retrieval · p. 18

D Challenges and Research Opportunities

  • Future directions address evidence sufficiency, verifiable and correctable understanding, and budget-aware coordination for individual questions, alongside memory reuse and strategy improvement across tasks.Figure 8 maps its three panels to evidence sufficiency, verifiable and correctable understanding, and budget-aware coordination under limited budgets.Source: D Challenges and Research Opportunities · p. 19, Figure 8 · p. 19

D.1 Evidence Sufficiency

  • Finding relevant content does not ensure sufficient answer support: cross-page context or question-relevant regions can still omit essential dependencies, such as a table header.Task-oriented sufficiency asks whether current evidence supports an answer and what is missing.Source: §D.1 · p. 19
  • Future work should construct question-driven evidence sets by modeling required facts and dependencies, selecting complementary evidence, and testing judgments of sufficiency and gaps.These judgments can provide testable grounds for adaptive acquisition and stopping.Source: §D.1 · pp. 19–20

D.2 Verifiable and Correctable Understanding

The paper proposes traceable, revisable evidence states to improve verification and correction, and budget-aware decisions that weigh end-to-end quality against cost.

D.2 Verifiable and Correctable Understanding

  • Correct answers can coexist with incomplete evidence chains, motivating checks that acquired information is used faithfully and errors can be corrected.The discussion raises whether parsing or summarization errors can enter memory and be repeatedly reinforced.Source: §D.2 · p. 20
  • Future work should connect source tracing, process verification, and correction to maintain traceable, correctable evidence states across rounds.Revisable memory should retain sources and dependencies, retract erroneous records, and update dependent judgments; controlled evidence changes can test support and robustness.Source: §D.2 · p. 20
  • Under cross-page dependencies, distractors, visual degradation, or unanswerable conditions, systems should be evaluated on reparsing, further acquisition, answer revision, abstention, and error recovery.Source: §D.2 · p. 20

D.3 Decision Making under Budget Constraints

  • Local savings from compression or dynamic processing may be offset by auxiliary calls, interaction rounds, and state maintenance, so costs and quality require full-pipeline assessment.The paper proposes using evidence states and remaining budgets to decide whether to search, inspect, reparse, verify, or answer.Source: §D.3 · p. 20
  • Budget-aware evaluation should compare quality–cost curves across document lengths, evidence dispersion, and budgets, including preprocessing, indexing, auxiliary models, and online inference.Budget allocation can adapt parsing scope, resolution, retrieval depth, and reasoning rounds to question difficulty and evidence gaps.Source: §D.3 · p. 20

D.4 Towards Continual Document Understanding

This section examines directions toward sustained document understanding through memory reuse across interactions and experience-driven strategy improvement. It also defines the table’s mechanism notation and explains its evaluation dimensions.

D.4 Towards Continual Document Understanding

  • Continual document understanding combines memory reuse across questions with experience-driven improvement of processing strategies.Systems should solve the current question, exploit connections across successive interactions, and improve later processing through experience across tasks.Source: §D.4 · pp. 20–21

Memory reuse across questions

  • Multiturn systems can combine document content and interaction history to resolve references, retain comparison conditions, incorporate corrections, and reuse applicable evidence.MemoryDocDataSet explores conversational memory and reasoning over long text documents (Xie et al., 2026a); visual-document systems could link dialogue to source regions and revise memories after corrections or document updates.Source: §D.4 · p. 21

Continual, experience-driven strategy improvement

  • Systems could learn transferable retrieval, reading, and verification strategies from task experience, validate proposed strategies on independent tasks, then update skills or control procedures.MemSkill revises memory skills through difficult cases (Zhang et al., 2026b), and AIDE2 enables updated agents to participate in their own improvement (Srikanth et al., 2026); transfer requires establishing applicability conditions and feedback grounds.Source: §D.4 · p. 21

D.4 Towards Continual Document Understanding

  • Mechanism labels distinguish document, page, and element granularity and text, visual, or joint evidence, with arrows marking retrieval-to-generation evidence flow.The left side of an arrow is the evidence returned by retrieval or selection; the right side is the generation input.Source: §D.4 · p. 26
  • Evaluation dimensions are included only when quantitatively reported, with Ret. covering evidence retrieval and Loc. measuring within-page localization or coverage.The passage also notes exceptions to these measures for several methods.Source: §D.4 · p. 26
Loading 2610.31385v1…