According to AIatMeta, WildArtifactBench uses win rates and Elo from human and agent judges to benchmark multimodal agents across 10 real tasks.