Year: 2026
Type: preprint
Abstract: Document Visual Question Answering (DocVQA) over heterogeneous domains requires precise extraction and reasoning over documents whose formats range from business reports and scientific papers to maps, comics, and engineering drawings. No single vision-language model (VLM) performs uniformly well across such diverse content. We present a multi-agent document-reasoning system submitted to the DocVQA 2026 competition at ICDAR 2026, where it placed first in Task 1 (General category, \((>)\)35B-parameter class) with 60.00% ANLS (95% Wilson-score CI: \((\pm)\)7 pp; \((n = 160)\)), 13.75 pp ahead of the second-placed system (46.25%). The system combines three components: (1) IDA, a layout-aware OCR engine that produces structured Markdown output as a deterministic text anchor; (2) four frontier VLMs (Gemini 3.1 Pro, Gemini 2.5 Pro, Claude Sonnet 4, Qwen3.5) performing question-guided focused reads with model-adapted question reformulations; and (3) a Claude Opus 4.6 reasoning agent that synthesizes perspectives through domain-specific trust hierarchies and cross-perspective conflict detection. A first submission reached 55.63% ANLS; a refinement pass using the first run's reasoning outputs as prior context lifted the score to 60.00% (\((+4.37)\) pp). The system improves over the strongest single-model zero-shot baseline on the leaderboard (Gemini 3.1 Pro: 37.50%) by \((+22.5)\) pp and over a naive mixture-of-experts estimate (\((\sim)\)40%) by approximately \((+20)\) pp. We attribute roughly \((+7)\) pp to IDA's deterministic text extraction and \((+13)\) pp to multi-agent orchestration; these are attributed estimates from internal experiments rather than controlled-ablation measurements, and we state this limitation explicitly. Per-domain, the system scores highest on text-heavy documents (business reports 0.75, scientific papers 0.70 ANLS), where deterministic extraction anchors the answer, and lowest on comics (0.45), with maps and science posters at 0.50.
Source: Research Square
Author: Welf Wustlich
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226033018
