DocLens : A Tool-Augmented Multi-Agent Framework for Long Visual Document Understanding

TL;DR

DocLens is a multi-agent framework that improves long visual document understanding by effectively localizing evidence and generating accurate answers, surpassing previous models and even human experts on benchmark datasets.

Contribution

Introduces a tool-augmented multi-agent framework that enhances evidence localization and answer generation in long visual documents, addressing fundamental limitations of existing Vision-Language Models.

Findings

Achieves state-of-the-art performance on MMLongBench-Doc and FinRAGBench-V datasets.

Outperforms existing models and even human experts on vision-centric and unanswerable queries.

Demonstrates superior evidence localization capabilities in long visual document comprehension.

Abstract

Comprehending long visual documents, where information is distributed across extensive pages of text and visual elements, is a critical but challenging task for modern Vision-Language Models (VLMs). Existing approaches falter on a fundamental challenge: evidence localization. They struggle to retrieve relevant pages and overlook fine-grained details within visual elements, leading to limited performance and model hallucination. To address this, we propose DocLens, a tool-augmented multi-agent framework that effectively ``zooms in'' on evidence like a lens. It first navigates from the full document to specific visual elements on relevant pages, then employs a sampling-adjudication mechanism to generate a single, reliable answer. Paired with Gemini-2.5-Pro, DocLens achieves state-of-the-art performance on MMLongBench-Doc and FinRAGBench-V, surpassing even human experts. The framework's superiority is particularly evident on vision-centric and unanswerable queries, demonstrating the power of its enhanced localization capabilities.