ResearchBenchmarksHuman Preference Evaluation on PhysicsIQ 1.0 (test)Follow54.9Physics Plausibility Win RateMAGI-144.70847.3545052.646Jan 15, 2026Evaluation ResultsMethodMethodLinksPhysics Plausibility Win RatePhysics Plausibility AccuracyVisual Quality Win RateVisual Quality AccuracyPrompt Alignment Win RatePrompt Alignment AccuracyOverall Win RateOverall AccuracyMAGI-1Sampling Strategy=WMRe...Sampling Strategy=WMReward(∇+BoN)2026.0154.958.752.957.155.764.954.560vLDMSampling Strategy=WMRe...Sampling Strategy=WMReward(∇+BoN)2026.0153.154.854.758.851.353.25355.7vLDMSampling Strategy=vanillaSampling Strategy=vanilla2026.0146.945.245.341.248.746.84744.3MAGI-1Sampling Strategy=vanillaSampling Strategy=vanilla2026.0145.141.347.142.944.335.145.540