Skip to content
Book Open access

PageRAG-H: Page-Level Visual Retrieval-Augmented Generation with Grounded Reasoning Harness for Multilingual PDF Question Answering

Nov 2026 · 0 citations

Abstract

Multilingual PDF question answering requires systems to reason over long, visually rich documents and ground each answer to its supporting pages. Recent multimodal language models have shown strong document understanding capabilities. However, directly applying them to multi-page PDFs remains challenging. The model must identify sparse supporting evidence from visually complex pages containing tables, figures, and scanned regions. We propose PageRAG-H, a page-level visual retrieval-augmented generation framework equipped with a grounded reasoning harness. PageRAG-H retrieves visually relevant pages through late-interaction visual matching and guides the multimodal language model to perform page-wise evidence checking, evidence selection, answer derivation, and structured output generation. The harness further enforces image-to-page grounding and evidence-page validity, and is refined by a score-guided model-in-the-loop agent using errors on labeled LAVA training examples and leaderboard feedback. In the LAVA 2026 official final evaluation, PageRAG-H achieved a score of 0.8930 and ranked first, demonstrating that compact visual evidence retrieval and controlled grounded reasoning are effective for multilingual PDF question answering.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.