document parsing

This work addresses high-accuracy conversion of unstructured/semi-structured documents (e.g., contracts, academic papers, invoices) into structured, machine-readable data. Method: We systematically survey and empirically compare modular pipeline approaches against end-to-end multimodal large models, proposing a unified framework integrating OCR, layout analysis (LayoutParser), graph neural networks, vision-language models (VLMs), and specialized formula/table recognition. We identify and characterize core bottlenecks—layout understanding, dense text recognition, and cross-modal alignment—for the first time. Contribution/Results: We establish a comprehensive analytical framework covering methodology, challenges, and benchmarks, revealing >32% performance gaps of current SOTA on complex layouts (e.g., multi-column, nested tables). We propose a “dual-driven” evolution path emphasizing both data diversity and scale, and open-source a larger annotated dataset to significantly advance knowledge base construction and training-data generation for large models.

This work addresses the challenges of structured information extraction from web pages, where handcrafted rules are brittle and large language models incur prohibitive computational costs. The authors propose a novel paradigm that treats the HTML DOM as a prunable tree, employing a dedicated pruning mechanism to retain high-information-density subtrees. A lightweight 0.6B-parameter language model is then applied for zero-shot cross-domain extraction. To ensure result reliability, the approach incorporates a traceable Grounded XPath Resolution (GXR) mechanism. Evaluated on the SWDE dataset, the method achieves an F1 score of 88.1%, outperforming several larger, fully trained baseline models while substantially reducing computational and deployment overhead.

This work addresses the limitations of existing methods in question answering over semi-structured documents containing tables, figures, and hierarchical paragraphs, which often suffer from fragmented OCR outputs, inadequate hierarchical modeling, and insufficient cross-region alignment. To overcome these challenges, we propose MoDora, a novel system that reconstructs OCR results through local alignment aggregation, explicitly models inter-component hierarchical and spatial relationships via a Component Correlation Tree (CCTree), and introduces a question-type-aware hybrid retrieval mechanism that fuses semantic and layout information. Evaluated on multiple benchmarks, MoDora significantly outperforms current state-of-the-art approaches, achieving absolute accuracy gains of 5.97% to 61.07% and enabling precise, layout- and semantics-driven document understanding.

This work addresses the limitation of existing document analysis systems that flatten complex documents into plain text, thereby discarding critical hierarchical structures such as sections, tables, and figures, which hinders effective filtering and in-depth analysis. To overcome this, the authors propose a structure-aware document understanding framework that integrates full document hierarchy into semantic indexing and analysis. The approach constructs a hierarchical document tree by parsing the original layout, leverages large language models to generate structure-aware semantic representations, and introduces a multi-view interactive web interface enabling precise natural language–driven retrieval and question answering. Experiments demonstrate significant improvements in retrieval accuracy and question-answering performance on diverse complex documents, including academic papers, technical manuals, and financial reports. The code and a live demo system are publicly released.

To address core challenges in LLM-driven semantic data processing—including user cognitive gaps, prompt inconsistency, and inefficient pipeline construction—this paper proposes a hybrid proactive IDE for semantic data processing. Our approach introduces three novel mechanisms: (1) in-situ user annotations to bridge the gap between user intent and model interpretation; (2) LLM-assisted prompt refinement to enhance prompt robustness; and (3) LLM-guided operation decomposition to support progressive programming with semantic operators (e.g., map, reduce, filter). The system integrates prompt engineering, interactive IDE design, and operation decomposition strategies. Through a 10-participant empirical study and analysis of 1,500+ real-world interaction sessions, we demonstrate significant improvements in users’ strategic evolution capabilities—specifically, their ability to transform open-ended tasks into verifiable classifiers and collaboratively explore data and model boundaries using ambiguous prompts.

This study addresses the lack of efficient, automated methods for structuring heterogeneous real estate questionnaire documents. To this end, the authors propose an end-to-end information extraction framework that first categorizes documents into structural types using K-Means clustering and text classification. Subsequently, it leverages the DeepSeek-R1 large language model enhanced with prompt engineering to accurately extract 35 predefined attributes from complex document formats—including checkboxes and scanned images—and outputs them as structured JSON. Evaluated on a dataset of 2,781 documents, the method produced 2,766 unique property records. Downstream validation demonstrated a Jaccard similarity of 0.82, marking the first high-precision, scalable solution for structured information extraction from such challenging real estate documentation.

This work addresses the limitations of existing unified document parsing models, which rely on full-page autoregressive generation and thus struggle to scale to long documents while failing to exploit global layout structure and parallelism across content blocks. To overcome these challenges, the authors propose a hierarchical parallel decoding paradigm that employs a master layout branch to orchestrate document structure and dynamically allocates content blocks to concurrent decoding branches. This approach is further enhanced by progressive multi-token prediction (P-MTP) to reduce decoding steps. The method represents the first integration of hierarchical parallelism into unified document parsing, achieving a throughput of 4,752 tokens/s on public benchmarks—2.62× faster than the current fastest model and 3.06× faster than a standard autoregressive baseline—while maintaining competitive accuracy.

Existing benchmarks struggle to systematically evaluate the impact of multimodal inputs on structured information extraction. This work proposes the first multimodal extraction benchmark tailored for government forms, generating diverse documents with deterministic ground truth through procedural PDF templates and a reverse annotation pipeline. Each document is provided in four standardized input formats—plain text, layout-preserving text, image, and multimodal—to enable rigorous ablation studies. A three-stage quality control mechanism and compliance-based evaluation of structured outputs ensure high data fidelity and reliable assessment. Experiments reveal that small models (<4B parameters) are primarily limited by structural adherence; fine-tuning a 2B model yields an 81-percentage-point improvement; layout-preserving text consistently outperforms image inputs by 3–18 percentage points; and the benchmark exhibits maximal discriminative power within the 60–95% accuracy range.

This work proposes an intelligent web crawling approach based on multimodal large language models (MLLMs) to overcome the limitations of traditional crawlers, which struggle with dynamic, interactive websites and rely heavily on static HTML parsing and manual customization. The method integrates a specialized toolchain for web interaction and data extraction with a structured five-stage prompting mechanism, enabling fully automated, structured data collection from “index–content” architecture websites. By deeply coupling MLLMs with purpose-built tools, the system autonomously navigates complex user interfaces without human intervention. Experimental results demonstrate that the proposed approach significantly outperforms the Anthropic Computer Use baseline across six news websites and exhibits strong generalization capabilities in e-commerce scenarios.

This work addresses the limitations of existing web text extraction methods, which rely on a single fixed extractor and consequently suffer from insufficient data coverage and low utilization. To overcome this, the authors propose a multi-extractor ensemble strategy that integrates multiple open-source HTML extractors by taking their union, complemented by an optimized data filtering pipeline and a mechanism to preserve structured content such as tables and code. Evaluated on the DCLM-Baseline, this approach increases token yield by up to 71% while delivering performance gains of up to 10 and 3 percentage points on downstream tasks WikiTQ and HumanEval, respectively. These results underscore the critical impact of extractor selection on the effectiveness of structured reasoning tasks.