DatBench: Cut VLM eval compute by >10× while INCREASING signal(datologyai.com)

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