ResearchBenchmarksCompositional Visual Reasoning on CVR 34 (test)Follow99.3Elementary AccuracyPR-A2CL70.778.12585.5592.975Mar 1, 2026Evaluation ResultsMethodMethodLinksElementary AccuracyCompositional AccuracyColor & Position AccuracySize & Contact AccuracyContact & Rotation AccuracyColor & Flip AccuracyCount & Rotation AccuracyPR-A2CLSamples per task=1kSamples per task=1k2026.0399.389.198.395.49488.583.5Human2026.0397.67457.560304545PR-A2CLSamples per task=20Samples per task=202026.0371.839.226.261.565.127.428.5