DatBench fixes VLM evals: 70% blindly solvable, 42% mislabeled, 35% prod gap(datologyai.com)

https://www.datologyai.com/blog/datbench-discriminative-faithful-and-efficient-vision-language-model-evaluations?source=HN