HybriDoc: An Adaptive Multi-path Framework for End-to-End Document Structure Extraction

End-to-end Document Structure Extraction (DSE) from visually rich, unstructured documents such as PDFs is a foundational challenge in unlocking vast knowledge for downstream applications like Retrieval-Augmented Generation (RAG). However, prevailing monolithic parsing paradigms, constrained by a fixed inductive bias, often exhibit a lack of generalization when confronted with the high heterogeneity of document types. To address this, we move beyond a one-size-fits-all approach and propose HybriDoc, an adaptive, multi-path parsing framework. At the core of HybriDoc is a hierarchical decision engine that rapidly diagnoses a document’s archetype and dynamically dispatches it to an optimal expert parser. Furthermore, we introduce a cross-model fusion mechanism to leverage the complementary strengths of different parsers and a terminal LLM-based calibration module to ensure global semantic coherence. On the rigorous and comprehensive CCKS2025 End-to-End Document Structuring Challenge benchmark, HybriDoc establishes new state-of-the-art performance, achieving a final score of 83.10 and securing first place. Our framework demonstrates exceptional robustness and adaptability, particularly in the development (83.98) and general (87.71) document sub-categories. This work empirically validates that an adaptive, “divide-and-conquer” parsing strategy is a highly effective pathway for solving real-world document structuring problems.