Document Visual Question Answering (VQA) requires models to not only extract accurate textual answers but also precisely localize them within document images, a capability critical for interpretability in high-stakes applications. However, existing systems achieve strong textual accuracy while producing unreliable spatial grounding, or sacrifice performance for interpretability. We present ARIAL (Agentic Reasoning for Interpretable Answer Localization), a modular framework that orchestrates specialized tools through an LLM-based planning agent to achieve both precise answer extraction and reliable spatial grounding. ARIAL decomposes Document VQA into structured subtasks: OCR-based text extraction with TrOCR, retrieval-augmented context selection using semantic search, answer generation via a fine-tuned Gemma 3-27B model, and explicit bounding-box localization through text-to-region alignment. This modular architecture produces transparent reasoning traces, enabling tool-level auditability and independent component optimization. We evaluate ARIAL on four benchmarks (DocVQA, FUNSD, CORD, and SROIE) using both textual accuracy (ANLS) and spatial precision (mAP at IoU 0.50 to 0.95). ARIAL achieves state-of-the-art results across all datasets: 88.7 ANLS and 50.1 mAP on DocVQA, 90.0 ANLS and 50.3 mAP on FUNSD, 85.5 ANLS and 60.2 mAP on CORD, and 93.1 ANLS on SROIE, surpassing the previous best method (DLaVA) by +2.8 ANLS and +3.9 mAP on DocVQA. Our work demonstrates how agentic orchestration of specialized tools can simultaneously improve performance and interpretability, providing a pathway toward trustworthy, explainable document AI systems.

We present a lightweight yet effective pipeline for training vision-language models to solve math problems by rendering LaTeX encoded equations into images and pairing them with structured chain-of-thought prompts. This simple text-to-vision augmentation enables compact multimodal architectures to achieve state-of-the-art reasoning accuracy. Through systematic ablations, we find that rendering fidelity and prompt design are the primary drivers of performance. Despite its simplicity, our approach consistently matches or surpasses both open-source and proprietary math-focused vision-language solvers on widely used benchmarks, while preserving broad general-domain competence - showing gains on tasks such as MMMU, ChartQA, and DocVQA of up to 20%.

In the digital era, the ability to understand visually rich documents that integrate text, complex layouts, and imagery is critical. Traditional Key Information Extraction (KIE) methods primarily rely on Optical Character Recognition (OCR), which often introduces significant latency, computational overhead, and errors. Current advanced image-to-text approaches, which bypass OCR, typically yield plain text outputs without corresponding vision grounding. In this paper, we introduce STNet (See then Tell Net), a novel end-to-end model designed to deliver precise answers with relevant vision grounding. Distinctively, STNet utilizes a unique token to observe pertinent image areas, aided by a decoder that interprets physical coordinates linked to this token. Positioned at the outset of the answer text, the token allows the model to first see--observing the regions of the image related to the input question--and then tell--providing articulated textual responses. To enhance the model's seeing capabilities, we collect extensive structured table recognition datasets. Leveraging the advanced text processing prowess of GPT-4, we develop the TVG (TableQA with Vision Grounding) dataset, which not only provides text-based Question Answering (QA) pairs but also incorporates precise vision grounding for these pairs. Our approach demonstrates substantial advancements in KIE performance, achieving state-of-the-art results on publicly available datasets such as CORD, SROIE, and DocVQA. The code will also be made publicly available.

We perform a missing, reproducible evaluation of all publicly available GPT-4 family models concerning the Document Understanding field, where it is frequently required to comprehend text spacial arrangement and visual clues in addition to textual semantics. Benchmark results indicate that though it is hard to achieve satisfactory results with text-only models, GPT-4 Vision Turbo performs well when one provides both text recognized by an external OCR engine and document images on the input. Evaluation is followed by analyses that suggest possible contamination of textual GPT-4 models and indicate the significant performance drop for lengthy documents.

An important handicap of document analysis research is that documents tend to be copyrighted or contain private information, which prohibits their open publication and the creation of centralised, large-scale document datasets. Instead, documents are scattered in private data silos, making extensive training over heterogeneous data a tedious task. In this work, we explore the use of a federated learning (FL) scheme as a way to train a shared model on decentralised private document data. We focus on the problem of Document VQA, a task particularly suited to this approach, as the type of reasoning capabilities required from the model can be quite different in diverse domains. Enabling training over heterogeneous document datasets can thus substantially enrich DocVQA models. We assemble existing DocVQA datasets from diverse domains to reflect the data heterogeneity in real-world applications. We explore the self-pretraining technique in this multi-modal setting, where the same data is used for both pretraining and finetuning, making it relevant for privacy preservation. We further propose combining self-pretraining with a Federated DocVQA training method using centralized adaptive optimization that outperforms the FedAvg baseline. With extensive experiments, we also present a multi-faceted analysis on training DocVQA models with FL, which provides insights for future research on this task. We show that our pretraining strategies can effectively learn and scale up under federated training with diverse DocVQA datasets and tuning hyperparameters is essential for practical document tasks under federation.

We introduce Instruction Document Visual Question Answering (iDocVQA) dataset and Large Language Document (LLaDoc) model, for training Language-Vision (LV) models for document analysis and predictions on document images, respectively. Usually, deep neural networks for the DocVQA task are trained on datasets lacking instructions. We show that using instruction-following datasets improves performance. We compare performance across document-related datasets using the recent state-of-the-art (SotA) Large Language and Vision Assistant (LLaVA)1.5 as the base model. We also evaluate the performance of the derived models for object hallucination using the Polling-based Object Probing Evaluation (POPE) dataset. The results show that instruction-tuning performance ranges from 11X to 32X of zero-shot performance and from 0.1% to 4.2% over non-instruction (traditional task) finetuning. Despite the gains, these still fall short of human performance (94.36%), implying there's much room for improvement.

Document Visual Question Answering (DocVQA) is a fast growing branch of document understanding. Despite the fact that documents contain sensitive or copyrighted information, none of the current DocVQA methods offers strong privacy guarantees. In this work, we explore privacy in the domain of DocVQA for the first time. We highlight privacy issues in state of the art multi-modal LLM models used for DocVQA, and explore possible solutions. Specifically, we focus on the invoice processing use case as a realistic, widely used scenario for document understanding, and propose a large scale DocVQA dataset comprising invoice documents and associated questions and answers. We employ a federated learning scheme, that reflects the real-life distribution of documents in different businesses, and we explore the use case where the ID of the invoice issuer is the sensitive information to be protected. We demonstrate that non-private models tend to memorise, behaviour that can lead to exposing private information. We then evaluate baseline training schemes employing federated learning and differential privacy in this multi-modal scenario, where the sensitive information might be exposed through any of the two input modalities: vision (document image) or language (OCR tokens). Finally, we design an attack exploiting the memorisation effect of the model, and demonstrate its effectiveness in probing different DocVQA models.

Recent advances in Large Language Models (LLMs) have stimulated a surge of research aimed at extending their applications to the visual domain. While these models exhibit promise in generating abstract image captions and facilitating natural conversations, their performance on text-rich images still requires improvement. In this paper, we introduce Contrastive Reading Model (Cream), a novel neural architecture designed to enhance the language-image understanding capability of LLMs by capturing intricate details that are often overlooked in existing methods. Cream combines vision and auxiliary encoders, fortified by a contrastive feature alignment technique, to achieve a more effective comprehension of language information in visually situated contexts within the images. Our approach bridges the gap between vision and language understanding, paving the way for the development of more sophisticated Document Intelligence Assistants. Through rigorous evaluations across diverse visually-situated language understanding tasks that demand reasoning capabilities, we demonstrate the compelling performance of Cream, positioning it as a prominent model in the field of visual document understanding. We provide our codebase and newly-generated datasets at https://github.com/naver-ai/cream .

Layout-aware pre-trained models has achieved significant progress on document image question answering. They introduce extra learnable modules into existing language models to capture layout information within document images from text bounding box coordinates obtained by OCR tools. However, extra modules necessitate pre-training on extensive document images. This prevents these methods from directly utilizing off-the-shelf instruction-tuning language foundation models, which have recently shown promising potential in zero-shot learning. Instead, in this paper, we find that instruction-tuning language models like Claude and ChatGPT can understand layout by spaces and line breaks. Based on this observation, we propose the LAyout and Task aware Instruction Prompt (LATIN-Prompt), which consists of layout-aware document content and task-aware instruction. Specifically, the former uses appropriate spaces and line breaks to recover the layout information among text segments obtained by OCR tools, and the latter ensures that generated answers adhere to formatting requirements. Moreover, we propose the LAyout and Task aware Instruction Tuning (LATIN-Tuning) to improve the performance of small instruction-tuning models like Alpaca. Experimental results show that LATIN-Prompt enables zero-shot performance of Claude and ChatGPT to be comparable to the fine-tuning performance of SOTAs on document image question answering, and LATIN-Tuning enhances the zero-shot performance of Alpaca significantly. For example, LATIN-Prompt improves the performance of Claude and ChatGPT on DocVQA by 263% and 20% respectively. LATIN-Tuning improves the performance of Alpaca on DocVQA by 87.7%. Quantitative and qualitative analyses demonstrate the effectiveness of LATIN-Prompt and LATIN-Tuning. We provide the code in supplementary and will release it to facilitate future research.