Is DocVQA still meaningful?

Known problems

• Labs keep it in technical reports only as a regression check.^{[1][1] Qwen3-VL technical report, arXiv 2511.21631 Primary}
• Four model families scored 85–97% on document, chart, OCR and visual-maths tests by November 2025.^{[1][1] Qwen3-VL technical report, arXiv 2511.21631 Primary}

What to look at instead

ChartQA, OCRBench v1 and MathVista-mini saturated at the same time.^{[1][1] Qwen3-VL technical report, arXiv 2511.21631 Primary} OCRBench v2 stays harder, at 66.8–67.1 in English (Nov 2025).^{[1][1] Qwen3-VL technical report, arXiv 2511.21631 Primary[2][2] OCRBench v2, arXiv 2501.00321 Primary}

• HallusionBench Status:Still useful Yes. Frontier models still score only 60–67, so it separates them on hallucination.
• MMMU-Pro Status:Near ceiling Partly. Scores sit near 85% and differ by source, so compare models within one source only.

Where it appears in the report

Chapter 7, Multimodal, voice and embodied

• Figure 1. Tests Beyond Text Fill Up in Months, Too
• Table 1. Each modality has one number that tells its story
• Table 2. Image and document benchmarks with their best scores in “Go deeper: Image and document benchmarks, full table”