Skip to content
Book Open access

Ground Before Answering: Evidence-First Vision-Language Reasoning for Multilingual PDFs (EFVLR)

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 20 references

TL;DR

EFVLR first retrieves candidate pages through hybrid multimodal and lexical retrieval followed by cross-modal reranking, then asks a vision-language model to verify direct evidence independently on each labeled candidate page, while original PDF page numbers are recovered deterministically from metadata.

Abstract

Document visual question answering requires models to understand visually rich documents while identifying the pages that support each answer. We present EFVLR, an evidence-first vision-language reasoning framework for multilingual PDF question answering. EFVLR first retrieves candidate pages through hybrid multimodal and lexical retrieval followed by cross-modal reranking. It then asks a vision-language model to verify direct evidence independently on each labeled candidate page, while original PDF page numbers are recovered deterministically from metadata. The final answer is generated only from the selected evidence-page images with structured parsing and recovery. This design limits expensive multimodal reasoning to a compact page set and avoids generative assignment of original PDF indices. On the LAVA Challenge 2026 test set of 624 questions over 200 PDFs, the submitted system obtains 0.80 on the public leaderboard and 0.82 on the private leaderboard, ranking fourth on the private leaderboard without task-specific fine-tuning. The complete pipeline runs in approximately 1 hour and 45 minutes on a single NVIDIA A100 40 GB GPU.

Read PDF

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