When Ink Meets Pixel: Intent-Aware OCR with Vision-Language Models for Human-Corrected Financial Documents

Human annotations on documents—such as arrows, carets, strikethroughs, and insertions—express corrective intent, yet existing OCR and document understanding systems treat these marks as visual noise rather than semantic edit operations. We introduce intent-aware OCR, the task of generating corrected transcriptions by jointly reasoning over original document content and visual intent cues (handwritten or digital) that convey edit intent. To support this setting, we present DocIntent, a multimodal dataset of 1,600 financial documents with over 5,700 annotated visual intent cues, each paired with layout-aware, intent-resolved OCR targets. We further introduce IntentBench, a 216-document benchmark designed to evaluate model robustness under diverse real-world annotation conditions. Leveraging DocIntent, we fine-tune DocIntentOCR, a suite of 3B and 7B parameter vision-language models, under direct intent supervision on corrected transcriptions. On IntentBench, the fine-tuned 3B and 7B models achieve F1-scores of 0.81 and 0.83, respectively—substantially outperforming commercial closed-source models (e.g., GPT-4.1 at 0.63) and open-source models (e.g., LLaMA-3.2-Vision-90B at 0.50). This demonstrates that intent-aware supervision enables compact domain-specific models to outperform substantially larger general-purpose systems on this task. This work establishes intent-aware OCR as a novel human-in-the-loop document understanding task and provides foundational resources (dataset, benchmark, and models) to advance research in interactive document processing.