DocEng '26: Proceedings of the 2026 ACM Symposium on Document Engineering

With the advent of end-to-end trainable, data-driven deep learning methods, a large number of classical document analysis tasks, such as layout analysis, OCR, handwriting recognition, and information retrieval underwent significant improvements in accuracy. However, at the same time, the “black box” effect of these methods remains challenging for domain experts, who rely on automatic document analysis in their profession, and who have to take responsibility for their observations and conclusions. In this talk, I will present several case studies that highlight benefits of combining deep learning with rule-based methods, with a view to explainability, accuracy, and human annotation effort.

Documents remain one of the primary carriers of knowledge in scientific, enterprise, and governmental settings, yet their complex visual structure—comprising layouts, tables, figures, and multi-column text—poses significant challenges for modern AI systems. Large language models do not natively understand such document structures, while traditional PDF parsers often lose semantic and structural information, leading to noisy or incomplete representations. In this talk, we present Docling, an open-source document processing framework designed to transform heterogeneous documents into high-quality, structured representations suitable for AI applications. Docling combines layout analysis, OCR, table structure recognition, and document assembly into a unified pipeline that produces rich, LLM-friendly outputs such as structured JSON and Markdown while preserving visual grounding and reading order. The framework supports multiple input formats and integrates with modern AI ecosystems, enabling applications including retrieval-augmented generation, schema-based information extraction, and agent-driven workflows. We discuss the architecture of Docling, its emerging ecosystem of models and tools, and its role in enabling scalable, privacy-preserving document intelligence through local execution and open standards. By bridging the gap between visually complex documents and machine-readable knowledge, Docling provides a foundation for reliable document understanding in next-generation AI systems.

Document reuse and accessibility of documents is a major concern for existing and future documents. In the US the (extended) ADA Compliance Deadline April 24, 2027 is approaching. The European Accessibility Act (EAA) already came into effect on June 28, 2025.

Accessible presentations are important in cases where slides are handed out to students or made available on websites. Presentations often use methods such as step-wise discovering of elements or replacement of displayed material—all of which need need special handling to make the presentation accessible.

In this tutorial we will present - for various types of LATEX presentations - how they can be made tagging aware and how the tagging can be checked. We will discuss open issues and future work and the resulting restrictions on supported input.

We will also show tools to inspect and validate well-tagged PDFs.

The participants will be able to compile and test on their own laptops (if they have internet access). A local TEX system is not necessary but if present it should be a current system (an up-to-date TEXlive 2026 or a current MiKTEX).

Modern Intensive Care Unit (ICU) records are not simply multimodal collections of text, tables, and biosignals. They are temporally entangled document systems, in which clinical meaning depends critically on when evidence is recorded, not only what is recorded. Physiological signals evolve continuously, laboratory results arrive with delays, and clinical notes retrospectively summarise events that unfolded hours earlier. These challenges, mainly asynchrony, irregular sampling, and label ambiguity, are not unique to healthcare. They arise wherever document systems capture evolving real-world processes, including legal proceedings, educational records, and financial audit trails. This work uses the ICU as a concrete, high-stakes instance of a general document engineering problem, often how to work through and build systems that respect, rather than bypass, the temporal structure of complex document systems.

This work showcases structured steps serving as an introduction to the temporal anatomy of ICU records and the documentation lag problem as well as handling temporal document-signal alignment under label ambiguity, including event-centred windowing, onset uncertainty, and weak label handling. The final concept of this work will address cross-modal explainability and evaluation, including XAI-based attribution across text, structured data, and biosignals, time-aware explanation methods, and the regulatory context established by the EU AI Act for clinical AI systems. No prior clinical informatics background is required.

We present an overview of the field of malware analysis with emphasis on issues related to document engineering. We will introduce the field with a discussion of the types of malware, including executable binaries, malicious PDFs, polymorphic malware, ransomware, and exploit kits. Malware analysis has been impacted by recent developments in machine learning, especially LLMs, and we will discuss and demonstrate some of these advances. We will conclude with our view of important research questions in the field. This is an updated version of tutorials presented in previous years, with more information about newly-available tools.

The energetic and environmental impact of AI is becoming significant in all domains of applications. Specific to Document Engineering, AI models have become the predominant components of modern processing pipelines, ranging from document information extraction, transformation, interpretation, translation, etc. As AI workloads are projected to potentially double global data center energy consumption by 2030, driven largely by generative models, there is an urgent need to accurately quantify and mitigate this footprint.

However, measuring this impact is challenged by the “black box” nature of massive model training, opaque hardware-software interactions, and a historical prioritization of accuracy over environmental efficiency. This tutorial addresses these challenges by exploring Lifecycle Assessment (LCA) frameworks and telemetry tools to track operational energy, carbon intensity, and embodied carbon. Participants will gain practical experience in auditing AI pipelines with open-source tools, implementing efficiency techniques such as quantization and distillation, and making strategic deployment decisions to minimize environmental harm. The session includes a live demonstration using a physical benchmark server to measure power drainage in real-time.

This paper presents the DocEng 2026 Competition on Maltese Optical Character Recognition (OCR). The competition challenged participants to develop OCR systems capable of accurately transcribing paragraph images extracted from Maltese-language PDF documents into single-line text suitable for corpus construction. As no annotated OCR training set was provided, participants were required to generate their own synthetic training data, while development and held-out test sets were supplied for validation and evaluation. Three teams participated, exploring approaches based on fine-tuned Tesseract models, transformer architectures, and ensemble methods. The results demonstrate that synthetic training data, combined with effective language-aware modelling and postprocessing, can produce highly accurate OCR systems for Maltese document transcription.

Document image binarization plays a key role in several image processing pipelines such as document transcription and information extraction. The quality and set-up of the digitalization devices, the intrinsic features of the documents and the kind and intensity of the noises present in the image have impact on the performance of the binarization algorithms. This assessment on binarizing photographed documents evaluated the quality, time, space, and performance of twenty-five new algorithms and one hundred and three “classical” algorithms. The test data set is composed of laser and deskjet printed documents, photographed using four widely used mobile devices with the strobe flash on, off, and in auto modes in two different shots.

Automating invoice processing in enterprise settings requires more than high extraction accuracy: it demands a reliable mechanism to decide when an extraction is trustworthy enough to bypass human review. While modern Vision-Language Models (VLMs) achieve state-of-the-art results on document benchmarks, their integration into production pipelines remains hindered by the lack of principled methods to guarantee a bounded False Positive Rate (FPR). In financial workflows, an undetected extraction error (False Positive in production) can trigger irreversible erroneous payments, whereas a False Negative only incurs a reversible human review cost. In this paper, we formalize autonomous invoice validation as a selective classification problem under a field-level FPR constraint. We propose a multi-stage architecture that combines Optical Character Recognition (OCR) and Natural Language Processing (NLP) for field extraction, rule-based rejection, VLM-based correction, and a tabular Machine Learning (ML) scoring module. A key finding is that the VLM's most significant contribution is as an extraction denoiser: by correcting systematic OCR-NLP errors on fields not guarded by business rules, it significantly increases coverage (defined as documents not requiring human review) under a strict 1% field-level FPR target, while also bringing the FPR from above the target to below 1%. Furthermore, we apply a geometric safety layer, providing a tunable “safety knob” that ensures robust FPR guarantees across data folds at a quantifiable coverage cost. Evaluated on 11,253 real-world enterprise invoices, our system achieves 73% auto-validation coverage at a 0.96% field-level FPR, outperforming confidence-based baselines. Our results demonstrate that safe document automation requires not only high extraction accuracy but equally principled risk-constrained validation and the mitigation of label noise in production data.

Document AI (or Intelligent Document Processing) is a rapidly evolving field widely used to parse complex documents, yet its computational and environmental costs remain largely unquantified. To address this gap, this paper empirically evaluates the tradeoffs between accuracy and energy consumption across the three most representative categories of state-of-the-art open-source Document AI systems: pipeline tools, general-purpose Vision-Language Models (VLMs), and specialized VLMs. Using the reference benchmark OmniDocBench v1.5 and a controlled measurement protocol validated by physical wattmeters, we assess parsing performance alongside key sustainability metrics, including throughput, energy usage, and CO2eq emissions. Our findings reveal no strong correlation between parsing accuracy and energy consumption. Notably, specialized VLMs emerge as the most compelling category, with GLM-OCR achieving the best balance between accuracy and energy usage. In contrast, larger general-purpose models consume significantly more energy without necessarily delivering superior results. Moreover, pipeline tools demonstrate lower accuracy despite matching the energy footprint of specialized models. Finally, our extensive profiling shows that efficiency is heavily influenced by the hardware platform (CPU vs. GPU), the inference backend, and document characteristics, specifically length and complexity.

Two archives that compress the same original files will have different on-disk byte representations if they are created with different archive formats (e.g., ZIP, TAR) or compression algorithms (e.g., LZMA, DEFLATE). The ability to summarize the content and metadata of a compressed archive for later content or metadata identity analyses can help systems operate more efficiently; scan-repack-and-forward caching internet middleboxes, information processing systems, and file management and deduplication systems can all benefit from the offline ability to identify similar archives despite differences in on-disk byte representations. Many compressed archive formats are in use today, each with its own binary format, metadata fields, and often inconsistent implementations, so comparison and summarization must be flexible and format agnostic. To enable similarity comparisons between compressed archives, we propose a format-agnostic tiered canonical digest with 8 content and metadata parts. The digest's output size is independent of archive format, compression algorithm, and archive size. The digests reveal identity and similarity between archives and gracefully supports future extensions. They also enable new features such as flexible integrity checks, smart file caching, remote file system attestation, and forensic similarity analysis. We implement a reference canonical digest tool that supports 13 modern and historical compressed archive formats for digest creation, validation, and similarity analysis. Performance measurements show that digests are typically 350-440 bytes, can be computed in less than 1.5ms per KB of compressed archive, and use 300-500 KB of memory per KB of compressed archive.

A key task for better understanding and maintaining source code as a code document is semantic code document segmentation, i. e., dividing the code into coherent blocks of functional intent. Despite its potential to enhance code comprehension, navigation, and reuse, this task remains underexplored due to the lack of semantically segmented datasets. Prior work generally relied on syntactic signals (e. g., AST-based features) or manual heuristics, limiting scalability and generalization. We propose two language-model-based strategies: (i) a line-by-line approach that classifies each line of code separately before grouping the results into functional units, and (ii) a range-based approach that aims to directly determine groups of code lines from the input. The latter is particularly suitable for generative language models as they can take an entire code file as the input context. Furthermore, we release two expert-annotated datasets from real-world scientific code in both a low-resource language, R, and the widely used Python language. Experiments show that line-by-line strategy with a local context of K surrounding lines generally outperforms the range-based approach for both programming languages. Fine-tuning smaller models like CodeBERT and CodeT5+ for line-by-line classification generally outperforms larger, generative language models, even without R-specific pretraining. On a single GPU, the runtime of CodeBERT is 100-170x faster than those of the best competing LLMs, supporting the practical integration of semantic segmentation into modern development environments. The code, prompts, and datasets are available at: https://github.com/Dahouabdelhalim/CodeSeg

Operational runbooks increasingly function as living documents within operational workflows: they are maintained by people, used in incident support, and continuously revised as organizational knowledge changes. Yet little is known about how such document collections evolve over time in production settings, or which interpretable signals are useful for monitoring document change. We analyze 17 weeks of version-controlled runbook snapshots produced during post-incident revision with machine-assisted review and human oversight. We quantify document evolution using lightweight linguistic and structural signals grounded in prior work on procedural language, technical communication, and document organization. Correlation analysis shows mixed patterns of association, including both positive and negative relationships, with effect sizes ranging from negligible to moderate and varying across corpora. Temporal summaries further distinguish comparatively stable signals from revision-sensitive ones, supporting monitoring of document maintenance workflows rather than one-shot quality judgments. An illustrative analysis on public GitHub product documentation shows that signal behavior differs across corpora, reinforcing the need for context-aware interpretation. These results identify which metric classes provide stable anchors and which provide edit-sensitive indicators for managing operational documentation as an engineered document resource.

LLM-driven construction can populate a knowledge graph (KG) with large volumes of triples extracted from documents, yet hallucination leaves their quality far from guaranteed. Existing verification methods return only a single label, one that fails to indicate which part of the triple is at fault and offers no evidence to check against. We present RA-QGQA, which recasts triple verification as a question-driven, corpus-grounded task. For each triple, a Question Generation (QG) module poses one question per slot (head, relation, and tail); a Retrieval-Augmented Generation (RAG) Question Answering (QA) module retrieves from the corpus and answers each independently; the answers are then matched against the triple to reach a verdict. The verdict pinpoints the faulty slot and carries a passage-level evidence chain the user can trace back to the source. We demonstrate RA-QGQA as an interactive web system in which users import a KG and its source corpus, verify all triples in a single pass, and, for any flagged triple, inspect the passages that justify its verdict. The system thus turns large-scale KG verification into a scalable, evidence-backed process.

This demonstration presents a Retrieval-Augmented Generation system which extracts design parameters from the technical reports of planning applications. It is implemented in partnership with an industry partner, utilising 145 documents and four projects of one of their clients. Testing of the various system features found two workflows that demonstrate up to 5x improvement over a baseline scenario.

We present MilitAI, a hybrid query platform designed for military registers from the 18^th century digitised by the French Mémoire des Hommes initiative. MilitAI answers natural-language and template questions about soldiers, their origins, their families and their service in three complementary modes over a knowledge graph and a vector store, returning exact, source-linked results. Its distinctive feature is measurability: as few registers, once transcribed by handwritten-text recognition systems, carry a ground truth, we evaluate each query mode against a synthetic corpus, generated in a complete and a masked condition. The masked condition reproduces the sparsity of the transcribed corpus, so that a single ground truth separates the accuracy of the query engine from the analytical cost of incomplete annotation. We demonstrate both interactively.

We demonstrate Dash, a hypermedia workspace in which AI agents operate as first-class documents on a freeform spatial canvas, running entirely on-device without transmitting user data to external servers. Building on the Dash hypermedia system—which treats all document types uniformly within a richly linked spatial canvas— this work describes the addition of an agentic framework powered by open-source models running locally via Ollama. The accompanying video illustrates three core capabilities: context provision through native bidirectional linking, local multi-agent workflow orchestration, and a security model guaranteeing that sensitive data never leaves the device. A medical scenario walkthrough demonstrates the system's applicability to HIPAA-regulated clinical documentation, though the same agentic workflows apply equally to legal document review, defense intelligence analysis, financial modeling, and academic research.

Extreme multi-label classification (XML) and hierarchical text classification (HTC) address closely related multi-label prediction problems, but have largely developed as separate research areas. XML focuses on very large label spaces and typically evaluates ranked label lists, while HTC assumes a human-curated label hierarchy and commonly reports classification-based F1 scores. This paper studies the transferability of representative methods across these two settings. We evaluate XML models on HTC benchmarks by flattening the human-defined hierarchy, and HTC models on XML benchmarks by inducing synthetic hierarchies. To make the comparison meaningful across both communities, we report classification metrics, ranking metrics, and R-Precision as a label-cardinality-aware bridge between them. Our results show a clear asymmetry: XML methods transfer well to HTC datasets and are often competitive with specialised HTC models, especially on ranking-based metrics. In contrast, HTC methods struggle on XML datasets, either due to scalability limits or substantially lower performance. These findings suggest that XML methods may be considered strong baselines for HTC, and that future HTC models should also be evaluated for scalability and ranking performance. Our code is available at https://github.com/FloHauss/XMC_HTC.

Maintaining up-to-date code documentation is difficult in fast-moving repositories because design knowledge is scattered across source files and pull requests. We present CODENS, a system that turns pull requests into living, accessible, and queryable documentation for production codebases. CODENS incrementally builds a typed software knowledge graph from pull requests, enriches components through schema-driven semantic extraction, derives typed relations between them, and exposes the resulting knowledge through three retrieval modes, including agent-guided graph traversal for repository-level question answering. The system also preserves semantic change history across pull requests and integrates both answer-quality and operational evaluation metrics. We evaluate CODENS on a client Ruby on Rails project in production. Results show that CODENS produces highly relevant and well-grounded answers, while qualitative feedback highlights a remaining challenge in concise, documentation-oriented synthesis.

Culture shapes reasoning, values, prioritization, and strategic decisionmaking, yet large language models (LLMs) often exhibit cultural biases that misalign with target populations. As LLMs are increasingly used for strategic decision-making, policy support, and document engineering tasks such as summarization, categorization, and compliance-oriented auditing, improving cultural alignment is important for ensuring that downstream analyses and recommendations reflect target-population value profiles rather than default model priors. Previous work introduced a survey-grounded cultural alignment framework and showed that culture-specific prompting can reduce misalignment, but it primarily evaluated proprietary models and relied on manual prompt engineering. In this paper, we validate and extend that framework by reproducing its social sciences survey based projection and distance metrics on open-weight LLMs, testing whether the same cultural skew and benefits of culture conditioning persist outside closed LLM systems. We then introduce use of prompt programming with DSPy for this problem—treating prompts as modular, optimizable programs—to tune cultural conditioning by optimizing against cultural-distance objectives. In our experiments, we show that prompt optimization often improves upon cultural prompt engineering, suggesting prompt compilation with DSPy can provide a more stable and transferable route to culturally aligned LLM responses.

Public procurement refers to the process by which public sector institutions, such as governments, municipalities, and publicly funded bodies, acquire goods and services. Swiss law requires the integration of ecological, social, and economic sustainability requirements into tender evaluations in the format of criteria that have to be fulfilled by a bidder. However, translating high-level regulations into concrete, verifiable, and sector-specific procurement criteria (such as selection criteria, award criteria, and technical specifications) remains a labor-intensive and error-prone manual task, requiring substantial domain expertise in several groups of goods and services and considerable manual effort. This paper presents a configurable, LLM-assisted pipeline that supports automated document workflows for generating and evaluating sustainability-oriented procurement criteria catalogs in Switzerland. Our prototype system SwissSPC integrates in-context prompting, interchangeable LLM backends, and automated output validation to enable auditable criteria generation across different procurement sectors. As a proof of concept, we instantiate the pipeline using official sustainability guidelines published by the Swiss government and the European Commission, which are ingested as structured reference documents. We evaluate the system through a combination of automated quality checks, including an LLM-based evaluation component, and expert comparison against a manually curated gold standard. Our results demonstrate that the proposed pipeline can substantially reduce manual drafting effort while producing criteria catalogs that are consistent with official guidelines, thereby highlighting key considerations for integrating generative AI into public sector procurement workflows.

Compared with binaries and decompiled code, malware source code more directly reflects the attackers' original intent. However, the scarcity of source code and the high cost of manual review make such datasets difficult to build and maintain. We propose MASCOT-Android, a curated dataset of Android malware source code and an automated collection framework for scalable malware source code discovery on GitHub. A key finding of our work is that repository-level documentation alone provides a strong signal for malware source code collection. Our model extracts character-level TF-IDF features from 8,772 malware and 25,747 benign README documents and trains a LinearSVC classifier to distinguish malware repositories. This README-only model achieves an accuracy of 96.28% and an FPR of 1.06% in local evaluation. In addition, the model outputs confidence scores, allowing users to adjust the decision threshold to balance FPR and coverage, which is practical in real-world malware source code collection.

We conducted two case studies. First, we constructed an Android malware code-reuse graph and combined it with LLM-based code detection to assess traces of LLM assistance in malware development. The results suggest that LLMs are already contributing, at least to some extent, to the development and propagation of malware. The second study performs symbolic information ablation experiments in which we gradually remove different types of symbolic information from malware source code to assess their impact on malware detection performance. This study shows that import statements contain highly informative signals because they are related to API usage, whereas comments and class names have limited discriminative value. In summary, we present a curated dataset of Android malware source code and an automated collection model, and our case studies highlight the value of source code for studying both LLM-assisted malware development and the role of symbolic information in malware detection.

RVL-CDIP is a popular dataset for benchmarking document classifiers. However, the dataset contains ample amounts of label errors as well as non-trivial amounts of test-train overlap, both of which may impact model performance metrics. In this paper, we address these two problems by (1) finding and fixing label errors, and (2) detecting and addressing test-train overlap. We produce several variations of RVL-CDIP with label error and test-train overlap fixes, and benchmark document classification performance on these new RVL-CDIP variations. Our rigorous analysis of RVL-CDIP finds that the corpus contains 12% label error and approximately 35% test-train duplication. Remediation sees improvements in classification accuracy when errors are removed, but sees decreases in accuracy when duplicates are removed. We additionally evaluate models on RVL-CDIP-N, an out-of-distribution benchmark, finding that training on error-corrected data substantially improves OOD generalization, with supervised models gaining an average of 8.1 percentage points in accuracy and improvements as large as 14 percentage points.

Question answering over visually structured documents remains difficult when evidence is distributed across prose, tables, figures, captions, visual layout, and document structure. We present DOCLAYOUT-MM-RAG, a layout-aware annotation framework for grounded question answering over documents. The framework links each question-answer instance to supporting layout elements, preserving element-level provenance for annotation, retrieval, citation, generation, and evaluation. We instantiate the framework on annual reports and release an initial curated corpus of 650 accepted grounded question-answer instances across 30 documents. The corpus captures evidential complexity, with 36.6% of instances requiring cross-page support and 40.8% requiring multimodal support. Exploratory analyses compare flat-text, structure-aware, and layout-derived multimodal retrieval representations, and show how element-level provenance enables retrieval-to-generation and oracle-evidence analysis. DOCLAYOUT-MM-RAG provides a concrete basis for studying provenance-preserving retrieval-augmented generation over visually structured documents.

Software architecture diagrams communicate software structure and relationships, but automatically estimating their connectivity remains difficult. We study whether Vision Large Language Models, e.g., LLaVA, Gemma, Mistral, and Qwen-VL, can estimate connectivity through a localized proxy: counting arrowheads, usually denoting dependencies, control flows, or communication links. We evaluate local and cloud-based VLLMs on annotated architecture diagrams, showing that the strongest models capture useful connectivity information: Qwen3-VL-Cloud achieves the best overall error-based performance (MAE 1.73, RMSE 3.65), while Gemma4 obtains the highest exact-match accuracy (48.1%). However, performance degrades on large and dense diagrams. To address this limitation, we introduce an image-slicing strategy counting arrowheads in sub-regions; a 2 × 2 partition improves tolerance-based accuracy from 65.8% to 73.3%. Overall, while estimating diagram connectivity and complexity remains challenging, preliminary results suggest VLLMs combined with simple diagram-aware preprocessing are promising for arrowhead localization and graph reconstruction.

Piping and Instrumentation Diagrams (P&IDs) are essential engineering documents, but many remain available only as scanned PDFs, limiting their integration into digital workflows. Automatic extraction of structured information from these drawings is challenging due to large document sizes, small text annotations, and ambiguous symbol-tag relationships.

This paper presents an end-to-end framework for automatic instrument-tag extraction and association from scanned P&IDs. A tiled OCR strategy with graph-based text merging improves text completeness and reconstructs fragmented engineering tags. For symbol detection, an RF-DETR model fine-tuned on the Dataset-P&ID benchmark achieves 99.96% mAP@50 and 99.97% precision, while SAHI-based inference slicing improves performance on large drawings.

To automate symbol-tag association, we formulate the problem as a minimum-cost bipartite matching task that combines geometric, semantic, and spatial cues and solves it globally using the Hungarian algorithm. Results on Dataset-P&ID demonstrate accurate tag reconstruction, highly reliable symbol-tag associations, and a substantial reduction in manual annotation effort, providing an effective foundation for large-scale P&ID digitization. The code is available at https://github.com/dimitri009/STA.

Detecting and extracting sensitive information from documents is essential for privacy and regulatory compliance. Existing approaches either require training on large labeled datasets or rely on brittle, costly-to-curate pattern matching, while Large Language Models (LLMs) offer a promising alternative. We present a systematic evaluation of several proprietary and open-source LLMs for sensitive entity extraction from documents. Because large datasets of completed forms containing personal information are unavailable, we also introduce a form-filling pipeline that uses vision-capable LLMs to label form fields, generate realistic synthetic personas, and fill real blank forms, enabling reproducible evaluation across diverse layouts. Evaluating on these forms and the public RVL-CDIP dataset, we find performance is uneven across entity types—and that LLMs fall short of a simple pattern-based baseline on Social Security numbers.

Privacy policies are text documents intended to inform consumers about the data practices of apps and websites, but they can be challenging to read. Some organizations try to address the obstacles by organizing their privacy policies as question-answer pairs. We use a combination of automated and manual analysis methods to compare two corpora: PrivaSeerQA, a corpus we create and release consisting of 1.2 million question-answer pairs extracted from 345,073 privacy policies, and PrivacyQA, a previously released corpus of 1,750 questions people ask about app privacy. We find a mixture of thematic overlaps (e.g., data types and data uses) and divergences (e.g., levels of granularity and the issue of location tracking). We also find that privacy policies have a greater focus on the visibility of data, while users have a greater focus on their data being sold. The contrasts suggest that QA-structured privacy policies still do not fully meet consumers' needs, although the structure clarifies what specific topics of user interest are missing.

OCR transcription errors in historical archives often hinder digital search and retrieval. While Large Language Models (LLMs) can correct many of these errors, applying them indiscriminately is costly and may negatively affect already-clean text. We propose a three-tier collaboration framework that routes each text segment to one of: (1) No Correction, (2) LLM Correction, or (3) Human Correction. We introduce a regression-guided routing approach that prioritizes segments by predicted CER improvement, paired with a safeguard layer that detects harmful LLM corrections and routes uncertain segments to human review. With only <5% of the corpus reviewed by human experts, our safeguard achieves a 14% relative reduction over the All-LLM baseline, and substantially outperforms standard confidence-based approaches. By dynamically routing degraded segments to humans and fixable errors to the LLM, the collaborative framework outperforms either corrector in isolation.

This paper presents pandoc-lecturenotes, a system for personal single-source publishing based on pandoc and Markdown, which allows authors to maintain a single document containing both lecture notes (prose content) and slides, and lets them generate different types of documents (slides and lecture notes with and without embedded slides) from this source document. While conceptually and technically simple, it significantly simplifies the creation and maintenance of related documents. The paper also reports on the author's experiences with this system for creating course material and other scholarly documents over a period of almost 5 years.

Newspaper page layout is a highly constrained, information-dense design task whose production remains largely manual and labor-intensive. Its boundary conditions—high information density, strict spatial limits, and journalistic hierarchy—differ markedly from the document classes for which existing computational aesthetic measures were developed: generic measures penalize density and reward decorative whitespace, misaligning with professional editorial practice. In this paper we focus on two newspaper-specific aesthetic measures: a Text Density measure that penalizes large, uninterrupted text regions, and a set of Leading Article Measures that capture editorial hierarchy and the dominant visual entry point of a page. We operationalize both as computable algorithms and evaluate them on 4,624 professionally designed pages from a major German daily. Rather than treating measures as fitness functions to be maximized, we adopt a descriptive validation approach based on distributional concentration (excess kurtosis, coefficient of variation) to identify which measures reflect enforced professional norms. The results offer a compact, empirically grounded basis for incorporating domain-specific aesthetics into automated newspaper layout.

The proposal to organize an academic workshop dedicated to Legal Document Engineering has been submitted to, and accepted by the 2026 ACM Symposium on Document Engineering (DocEng'26) committee. The WoLDE'26 workshop takes place on August 28th, 2026, in Fribourg. The organizers are the authors of this abstract.

A technical program committee of 8 people has been created and a call-for-papers has been formally emitted ¹. Each submitted paper has received at least 3 reviews. Four papers have been accepted for publication and presentation.

Next to formal paper presentations, the workshop also features three keynotes talks and several work-in-progress presentations.

Accurate extraction of argumentative information from legal documents allows transforming legal documents into structured representations that can support downstream applications, such as information retrieval, document analysis, and legal decision support. Legal Argument Mining addresses the identification and classification of argumentative structures and their relations in legal texts. Such a task can be decomposed into multiple interdependent reasoning steps. In this work, we investigate the implementation of an argument mining pipeline using LLM agents, each one dedicated to a separate task. We evaluate our system under three different settings that progressively expose the impact of error propagation, ranging from independent evaluation of each agent to a realistic end-to-end scenario. We perform an experimental evaluation on the Demosthenes corpus, comparing an open-weight model against a commercial one. Results show that cascading errors substantially degrade downstream performance, making the first stages of the pipeline a significant bottleneck. Our findings highlight the importance of evaluating modular LLM systems in their end-to-end behavior in addition to individual task performance.

Akoma Ntoso was first developed twenty years ago as an XML framework for parliamentary and legislative documents within United Nations initiatives on legal information interoperability. Since then, it has become an OASIS standard and has progressively expanded its scope to judicial, administrative and general legal documents. This paper reflects on the technical and legal evolution of Akoma Ntoso, focusing on the architectural choices that have enabled its longevity: a regular structural model, persistent identification of legal resources, semantic metadata, multilingual and multicultural support, and controlled extensibility. We argue that these features, originally designed for open access and interoperability, have become increasingly relevant in the era of legal knowledge graphs, digital-ready policymaking and AI-based legal document processing.

Legal research traditionally relies on qualitative¹ analysis of legal texts, including case law. Yet, the growing number of judicial decisions makes systematic manual review increasingly difficult. This paper demonstrates how lightweight computational techniques, accessible to researchers with limited technical expertise, can support doctrinal legal research through the systematic collection, enrichment, and delimitation of legal corpus. Using asylum case law of the Swiss Federal Administrative Court (SFAC) as a case study, we design a transferable pipeline producing a structured database covering most decisions since 2007. Focusing on sexual orientation and gender identity (SOGI) asylum claims, the pipeline first identifies a SOGI-related corpus, representing around 1.25% of the collected database, which is then refined through citation network reconstruction. The pipeline enables the identification of structurally relevant decisions for qualitative analysis while producing data suitable for quantitative and computational research.

Reinforcement learning with verifiable rewards (RLVR) has been established as a viable paradigm for the post-training of Large Language Models (LLMs), including downstream tasks, such as Neural Machine Translation (NMT). With the latest research indicating that RLVR could be the preferred training method for translating legal documents due to the induced reasoning capabilities, it raises the question whether it is really attributed to the reasoning or more generally to the training paradigm. We investigate the importance of including the model's reasoning trace in the generated responses during both training and inference by systematically omitting it from one of the phases. Our experiments show that including the reasoning, specifically during inference, has a positive effect on the overall translation quality. Furthermore, we recognise that the reasoning leads to an increase in output tokens, hence we study the cost-quality tradeoff between the increased computational demands and the improved translation quality.