A Task-Centric Survey of Long-Document Visual Question Answering: Taxonomy, Evolution, and Future Directions
This survey tackles long-document visual question answering, where answers require locating and integrating evidence scattered across many pages and modalities including charts, tables, and fine-grained layout. Rather than grouping papers by model architecture, the authors decompose the task into four functional components: document parsing, evidence organization and retrieval, generator organization and learning, and evaluation. The core argument is that task bottlenecks—visual fidelity, evidence dispersion, and budget limits—should drive taxonomy design, not just technical novelty. This reframing matters because the field has grown explosively since 2024 and prior surveys have organized literature by approach rather than by the functional requirements that long documents impose.
The paper is a thorough and structurally clear survey that successfully reframes a crowded literature through a task-centric lens. Its four-component decomposition is intuitive and the temporal analysis provides a useful narrative arc for understanding how bottlenecks shape technical choices. However, the survey's heavy reliance on very recent and often unreviewed preprints—many dated 2025 and 2026—means its taxonomies rest on a rapidly shifting foundation, and some categorical distinctions flatten meaningful heterogeneity.
The strongest contribution is the fine-grained, function-driven taxonomy. Tables 1–3 meticulously map methods across parsing, retrieval, and generation dimensions including modality, granularity, mechanism, acquisition, reasoning, and learning, making the landscape scannable. The temporal evolution figures (Figs. 3–7) connect technical trends to expanding bottlenecks, offering a concrete historical narrative that goes beyond simple chronology and ties retrieval innovations back to the need for sufficient evidence.
A primary concern is that the taxonomy may be premature given the literature it surveys: with explosive growth since 2024 and many citations to 2026 preprints, the field is still in flux, risking rapid obsolescence of the categories themselves. The binary Deep Parsing flag (✓/✗) collapses structurally different approaches—summaries, trees, and graphs—into a single dimension, which flattens meaningful heterogeneity. Additionally, the future directions section remains high-level, calling for evidence sufficiency and budget-aware coordination without offering concrete technical roadmaps or prioritization criteria.
While the authors attempt balanced coverage, explicit retrieval dominates at approximately 80% of surveyed methods, rendering the direct-reading route comparatively thin. Parsing is also evaluated ambiguously: the authors correctly note that parsing value should be measured by downstream evidence completeness and cost, yet the tables still mark presence or absence (✓/✗) rather than measuring those outcomes, creating a tension between the stated functional philosophy and the actual comparative format.
The evidence is extensive: over 150 papers are synthesized into comparison tables and temporal analyses. The authors justify their taxonomic dimensions by linking them to task bottlenecks—input scale, visual complexity, and evidence dispersion—which grounds the framework better than a purely architectural survey. That said, claims about evolving research priorities rely on qualitative synthesis rather than quantitative bibliometric analysis beyond publication counts in Figure 1. Comparisons to existing surveys are fair but brief, and the paper could more directly demonstrate where prior architectural surveys misclassify or overlook methods that its task-centric view captures.
The authors share a project page with GitHub resources, which supports reproducibility of the supplementary material and benchmark catalog. The taxonomy dimensions and method comparisons in Tables 1–3 are detailed enough to serve as a public reference. However, the main text omits explicit survey methodology (search protocol, inclusion/exclusion criteria, date ranges), making it unclear how the 150 papers were selected and whether an independent reviewer could reproduce the literature scope. Hyperparameters and experimental details are not applicable, but the lack of a reproducible search strategy is a notable gap for a survey positioning itself as systematic.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.