DocTrace Enhances Traceable Long Document VQA with Evidence Graphs

Multimodal Large Language Models (MLLMs) struggle with Long Document Visual Question Answering (LongDocVQA) because they often lack explicit mechanisms to show how they arrive at an answer from diverse document elements. DocTrace addresses this by reframing LongDocVQA as an explicit evidence graph reasoning problem. This hierarchical framework first localizes relevant evidence across multiple pages, then performs structured document parsing, and finally constructs an evidence graph. This graph explicitly links pieces of evidence, allowing for transparent and verifiable reasoning. Through a two-stage training process, DocTrace significantly outperforms existing baselines and proprietary MLLMs, providing not just better answers but also clear provenance for those answers.