MMGist: A Comprehensive Multimodal Benchmark for 2027

The article introduces MMGist, a new multimodal benchmark designed to address limitations in existing vision-language benchmarks. It comprises 7,262 curated items across seven capability dimensions, developed through a rigorous three-stage filtering process, and has been tested on 27 leading large vision-language models (LVLMs). MMGist demonstrates high fidelity in preserving model rankings (Spearman $\rho = 0.98$) while significantly reducing the number of evaluation items by 69% and enhancing cross-model discrimination by 78%, highlighting the importance of visual dependency and discriminative power in evaluating LVLM performance.