In multi-page documents, a table may span more than one page, known as a cross-page table. Since many downstream tasks rely on the structural integrity of such tables, their reconstruction has become a critical subtask in multi-page document parsing. However, existing multi-page document parsing methods either ignore cross-page table reconstruction entirely or implement it via heuristic rules, leading to erroneous fragmentation or poor generalization. In particular, real-world cross-page tables often exhibit challenging patterns—such as vertical splitting, header repetition, cross-page cells, and dimension mismatch—which rule-based methods struggle to resolve robustly. To overcome these limitations, we propose TabStitcher, an LLM-based, two-stage framework for cross-page table reconstruction that can be seamlessly integrated with existing single-page parsers. It first identifies table fragments across consecutive pages and then merges them into a coherent, unified table structure. Furthermore, to address the scarcity of costly manually annotated data for training, we propose a data synthesis framework that can generate large-scale cross-page table instances with challenging patterns, enabling the model to learn to handle such cases effectively. Experiments on real-world benchmarks demonstrate TabStitcher’s effectiveness in both detection (98.3 F1-score) and merging tasks (96.6 TEDS). Besides, its integration with diverse off-the-shelf single-page parsers yields substantial improvements (at least 11.9 CPTEDS higher than the best baseline) in end-to-end cross-page table reconstruction. To the best of our knowledge, this work presents the first systematic study of cross-page table reconstruction in multi-page document parsing and offers a practical solution to this challenging task.
Tamper localization and document recovery are critical challenges in digital document forensics, where ensuring the authenticity and integrity of documents such as contracts, certificates, invoices, and identification records is essential. With the rapid advancement of document editing tools and AI-driven manipulation techniques, including content splicing, copy-move forgery, text alteration, signature replacement, and synthetic document generation, safeguarding digital documents has become increasingly complex. We propose Doc-Protector, a deep learning–based self-healing framework that enables both tamper localization and semantic-consistent document reconstruction. The method extracts hierarchical latent representations using a ResNet-34 backbone with a Feature Pyramid Network (FPN), capturing stroke-level, text-structure, and layout-level information. These compressed latents are embedded into the document via a mask-guided U-Net–based residual embedding network to ensure imperceptibility in text-sensitive regions. Given a tampered document, an extraction network recovers the embedded payload, which is decoded using a reverse-FPN hierarchical decoder to reconstruct the document in a coarse-to-fine manner. Tamper localization is obtained through discrepancy analysis between the received and reconstructed documents. Experiments under diverse manipulation scenarios demonstrate strong robustness in both in-domain and cross-domain settings, while reconstructed documents preserve structural and textual consistency under severe attacks .
Aramaic incantation bowls from Late Antiquity present a highly challenging visual domain characterized by severe material degradation, irregular layouts, and ambiguous letter morphology. Under such conditions, supervised learning is limited by the scarcity and instability of reliable paleographic labels. Self-supervised learning (SSL) offers an alternative to learning visual representations without explicit annotations; however, its effectiveness depends critically on the invariance assumptions encoded by data augmentations. In this work, we systematically evaluate seven SSL frameworks on a curated letter-level dataset derived from 32 Aramaic incantation bowls and organized around seven manually annotated letter classes. We formulate augmentation strategies as explicit hypotheses about visual identity: H1 (appearance-based transformations), H2 (structure-aware modifications), and H3 (historically and materially motivated degradations). Using controlled classification probes and quantitative geometric metrics, we analyze how these hypotheses shape embedding topology, cluster compactness, and inter-class separability. Our results show that the augmentation design substantially influences the quality of the representation. In particular, structural and historically grounded degradation hypotheses (H2, H3) consistently outperform appearance-only transformations (H1) and, in several cases, exert an effect comparable to or stronger than the choice of the SSL algorithm itself. These findings highlight the central role of domain-aware invariance design as an inductive bias for representation learning in degraded historical scripts. Project page: https://ssl-aramaic-bowls.github.io/ .
Automated information extraction from form-based historical records presents unique challenges, including highly variable page layouts, obsolete typefaces, and diverse handwriting styles. We introduce ADV-FORMS, a particularly complex, openly available dataset of form-based historical documents with benchmarks for layout analysis, handwritten text recognition (HTR), and optical character recognition (OCR). ADV-FORMS comprises 1,200 questionnaires in multiple layout templates, collected in the early 1930s, which are filled in with handwritten, typewritten and stamped content, as well as non-textual elements like underscores, drawings, and selection marks. Handwritten text occurs as German Kurrent and Latin script. ADV-FORMS contains manual annotations using bounding boxes capturing 17 recurrent, visually distinct layout categories as well as manual transcriptions of handwritings and all other typewritten or stamped texts. Baseline evaluations using fine-tuned YOLOv11 for layout detection and out-of-the-box TrOCR for HTR/OCR provide valuable insights, with layout detection achieving good precision/recall (>0.9) for most categories. HTR/OCR performance showed high word error rates, which can be addressed with post-filtering, as evidenced by high BERT scores (>0.8). In summary, our dataset provides benchmarks for further model development with each improvement directly bringing benefit to the analysis of historians and research in the Digital Humanities.
Long document understanding remains challenging for Vision-Language Models (VLMs) due to the need to locate relevant information across dozens of pages with diverse content types like text, tables, and figures. Existing embedding-based retrieval methods can miss relevant pages when the query–evidence relation is compositional. Embedding-free approaches avoid this by feeding all pages to VLMs, but incur high inference cost. Neither exploits the inherent hierarchical structure of documents for efficient navigation. In this paper, we propose AdaNav, an adaptive tree-based navigation framework for visual question answering in long documents. AdaNav constructs a multimodal document tree that hierarchically organizes sections, pages, and visual elements, enabling structured navigation without embedding retrieval models. We design the navigator to adapt the retrieval granularity to the question ensuring that content queries navigate from sections to pages in a coarse-to-fine manner. To support this, we build a multi-agent system with a question router and specialized navigators that differ in their action spaces, enabling traversal of the multimodal tree at the appropriate granularity. Experiments on MMLongBench-Doc demonstrate that AdaNav outperforms existing open-source VLM-based agent systems by over 5
