ResearchBenchmarksVision-Language Compositionality on WhatsUp A 1 × 4 groupsFollow56.8GroupScoreTTM29.531236.610643.6950.7694Oct 9, 2025Evaluation ResultsMethodMethodLinksGroupScoreAbsolute Gain (Δ)Relative GainError ReductionTTMBackbone=CLIP-B32, Alg...Backbone=CLIP-B32, Algorithm=Algorithm 12025.1056.826.285.737.7CLIP-B32Backbone=CLIP-B32Backbone=CLIP-B322025.1030.58———