TL;DR
This survey provides a task-centric framework for long-document Visual Question Answering (VQA), decomposing the process into four key components: parsing, retrieval, generation, and evaluation. It moves beyond model architectures to analyze how task bottlenecks—like visual information loss and search scale—drive technical evolution, ultimately pointing toward a future of verifiable, budget-aware, and continual document understanding.
What Problem Does This Paper Address?
Visually rich documents, such as financial reports, research papers, and legal forms, are the lifeblood of modern information. However, teaching AI to understand them is exceptionally difficult. Unlike standard VQA which might focus on a single image, long-document VQA must handle documents spanning dozens or hundreds of pages.
The authors identify three primary challenges:
- Input and Search Scale: Reading every page is computationally prohibitive. Systems must efficiently find a "needle in a haystack."
- Visual Content Complexity: Vital information is often hidden in small text, complex charts, or specific spatial layouts that are easily lost during simple text conversion.
- Evidence Dispersion: Answers often require integrating disparate pieces of evidence scattered across multiple pages or even different documents.
While previous surveys focused on specific model architectures, this paper takes a task-centric approach, looking at the functional bottlenecks that force researchers to make specific technical choices.
Methods
The authors propose a unified framework that decomposes the long-document VQA pipeline into four interconnected components:
- Document Parsing: Converting raw pages into accessible content. This ranges from basic OCR to "Deep Parsing" that builds semantic structures like trees or graphs to preserve relationships.
- Evidence Organization and Retrieval: This is the "search" phase. The paper highlights two main routes: Direct Reading (processing everything, often with compression) and Explicit Retrieval (selecting relevant pages/elements first).
- Generator Organization and Learning: How the model (often a Multimodal Large Language Model or MLLM) integrates the retrieved evidence to form an answer. This includes agentic reasoning where the model can "re-read" or "verify" its findings.
- Evaluation: Moving beyond simple answer accuracy to measure retrieval quality, localization precision, and computational efficiency.
Datasets
The survey catalogs a massive list of 55 benchmarks, categorizing them by their evaluation goals (retrieval-focused vs. answer-quality) and input scale.
- Classic Benchmarks: DocVQA, ChartQA, and InfographicVQA for single-page understanding.
- Long-Document Benchmarks: SlideVQA, MP-DocVQA, and the more recent MMLongBench-Doc and M-LongDoc, which test multi-page integration.
- Retrieval-Focused: ViDoRe (Visual Document Retrieval Benchmark) and MMDocIR, which specifically evaluate the ability to find the right page in a large collection.
- Diagnostic Benchmarks: New entries like DocScope and SynthDocBench are designed to diagnose exactly where a system fails—whether it's in parsing, retrieval, or reasoning.
Key Insights and Ideas
- The Rise of Explicit Retrieval: Since 2024, there has been a decisive shift toward explicit retrieval. Because reading long documents is so expensive, the "main route" today involves finding the right visual evidence first rather than trying to cram everything into a model's context window.
- Visual Fidelity vs. Textualization: Early methods converted everything to text (OCR), but this lost layout cues. Modern "visual-first" retrievers (like ColPali) index page images directly to preserve every visual detail.
- From One-Shot to Iterative Reasoning: We are moving from systems that take one guess to "Agentic" systems. These agents can look at a document, realize they are missing information, and go back to "acquire new evidence" or "re-read" a specific region.
- The Verifiability Gap: A major "aha moment" in the survey is that correct answers can often mask flawed reasoning. Future systems must provide provenance—linking every part of an answer to a specific region in the source document.
Key Results
The survey analyzes over 150 papers, revealing a field in explosive growth. The authors note that explicit retrieval now accounts for approximately 80% of surveyed methods, cementing its status as the dominant technical route for handling long contexts.
The temporal analysis shows a clear evolution of research priorities:
- 2021–2023: Focused on basic multimodal content understanding.
- 2024–2025: Shifted toward high-fidelity evidence acquisition (finding the right pages).
- 2026+ (Projected): Moving toward "sustained evidence management," where systems maintain a memory of what they've read to handle complex, multi-turn questions.
Limitations and Future Work
The authors acknowledge that their survey focuses on VQA as a high-level task; they don't dive deep into foundational low-level tasks like OCR engine design unless it directly impacts VQA performance.
Future Directions include: * Evidence Sufficiency: Developing models that can judge if the evidence they've found is enough to answer the question, or if they need to keep searching. * Budget-Aware Coordination: AI agents that can balance answer quality against the "cost" (computation/time) of further retrieval. * Continual Document Understanding: Moving beyond independent questions to systems that "remember" a document over time, reusing evidence and refining strategies across multiple tasks.
Summary updated