DocVAL: Validated Chain-of-Thought Distillation for Grounded Document VQA

Document visual question answering requires models not only to answer questions correctly, but also to precisely localize answers within complex document layouts. While large vision-language models (VLMs) achieve strong spatial grounding, their inference cost and latency limit real-world deployment; on the other hand, compact VLMs are efficient but suffer substantial localization degradation under standard fine-tuning or distillation. To address this gap, we propose DocVAL, a validated chain-of-thought (CoT) distillation framework that transfers explicit spatial reasoning from large teacher models to compact, deployable student VLMs. DocVAL combines (1) teacher-generated spatial CoT supervision, (2) a rule-based dual-mode validator that filters low-quality training signals and provides fine-grained, pixel-level corrective feedback, and (3) a validation-driven two-stage training procedure with iterative refinement. Text detection is used only as training-time scaffolding for supervision and validation, enabling the final student to operate as a pure VLM without OCR or detection at inference. Across multiple document understanding benchmarks, the proposed DocVAL yields consistent improvements of up to 6-7 ANLS points over comparable compact VLMs. We further introduce mean Average Precision (mAP) as a localization metric for document question answering and report strong spatial grounding performance under this new evaluation. We release 95K validator-verified CoT traces and show that high-quality, validated supervision is more effective than scaling unfiltered data, enabling efficient and trustworthy document grounding. Code/Data: GitHub.

Many important documents, such as receipts, invoices, forms, medical records, and legal files, contain information in specific places on the page. When people ask an AI system a question about a document, it is not enough for the system to give the right answer; it should also show where that answer came from, so that a human can quickly check it so as to increase the trustworthiness of the system. Large AI models are often good at this kind of document understanding, but they can be expensive and slow to run. Smaller models are faster and easier to deploy, but they often struggle to point to the correct location in the document. In this work, we introduce DocVAL, a training method that helps smaller models learn from stronger models while carefully checking the quality of what they learn. DocVAL teaches a smaller model not just the final answer, but also the step-by-step reasoning needed to find the answer in the document. It also uses automatic checks during training to remove poor examples and correct location mistakes. After training, the final model can answer document questions and point to the answer location without needing extra text-detection tools. We also release 95K carefully verified training examples to support future research on reliable document understanding.