arxivcs.CVcs.AI2026-07-11
SynthDocBench: Controlled Benchmark for Long-Context Visual Document Understanding
Abhigya Verma, Khyati Mahajan, Amit Kumar Saha, Shruthan Radhakrishna, Sagar Davasam, Vikas Yadav, et al.
Vision language models (VLMs) have achieved strong performance on visual document understanding benchmarks such as DocVQA, ChartQA, and MMLongBench-Doc. However, real-world documents combine multiple factors such as length, layout complexity, modality, and question difficulty, wh…