"ChartVRBench"

"ChartVRBench is a comprehensive benchmark meticulously designed to isolate and evaluate the genuine visual reasoning capabilities of Multimodal Large Language Models (MLLMs) in chart understanding. Unlike existing datasets that often conflate visual perception with textual OCR and arithmetic deduction, ChartVRBench focuses exclusively on the foundational task of numerical value estimation on non-annotated charts. To ensure robustness and real-world applicability, the benchmark features a highly diverse collection of charts spanning 38 distinct topics and seven primary chart types, encompassing both rigorously constrained synthetic data and manually curated real-world charts from authoritative platforms like Statista. By forcing models to interpret visual geometry, structural scales, and spatial relationships without relying on explicit text labels, ChartVRBench serves as a precise diagnostic tool to assess true multimodal comprehension and uncover visual hallucinations."