ResearchBenchmarksVisual Question Answering on TextVQA OCR n=5000 (val)For AI assistantsMarkdownContent indexVisual Question Answering on TextVQA OCR n=5000 (val)Follow0.8212Answer LossΠ30.8196680.8300090.840350.850691May 7, 2026Evaluation resultsMethodMethodLinksAnswer LossDelta vs Π3Π3best family=OCR gate +...best family=OCR gate + fallback2026.050.82120Π1best family=CART-d4best family=CART-d42026.050.84450.0233Π0best family=always_ocrbest family=always_ocr2026.050.84630.0251Π2best family=logistic C...best family=logistic C=0.32026.050.85950.0383