Action-conditioned reasoning on MindCube
36.85AccuracyWorld2VLM-GRPO
Evaluation results
| Method | Links | |
|---|---|---|
| World2VLM-GRPOSetting=HY-WorldPlay as WM2026.04 | 36.85 | |
| World2VLM-SFTSetting=HY-WorldPlay as WM2026.04 | 36.57 | |
| World2VLM-GRPOSetting=SVC as WM2026.04 | 34.86 | |
| Qwen2.5-VL-7B + WMSetting=MindJourney-style[55]2026.04 | 33.85 | |
| World2VLM-SFTSetting=SVC as WM2026.04 | 33.14 | |
| Qwen2.5-VL-7BSetting=Base VLM2026.04 | 29.26 |