Reward Fidelity on RoboTwin videos (held-out evaluation)
88Accuracy vs HumanCreFlow
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CreFlowReward Type=Binary2026.05 | 88 | 87.4 | 91 | — | 89.3 | |
| EVAReward Type=Cont. (IDM)2026.05 | 74 | 73.2 | 78 | 0.815 | 79.2 | |
| VLM-as-judgeReward Type=Binary2026.05 | 71 | 70.3 | 74 | — | 71.8 | |
| WMRewardReward Type=Cont. (JEPA)2026.05 | 70 | 69.1 | 73 | 0.758 | 74.1 | |
| VideoAlignReward Type=Cont. (Appearance)2026.05 | 63 | 62.4 | 65 | 0.652 | 62.1 |