Turing Post (Ksenia Se)· @TheTuringPost · X·· 1 天前AI 评分34
AI 导读
使用冻结编码器的世界模型在四个模拟器上将反事实R²从38–46%提升至86–96%。该方案在BlueROV2水下车辆测试中达到90–94%,并扩展至DINOv3、SigLIP 2等四种额外冻结编码器,得分达81–95%。此外,其回溯规划能力覆盖35–69%的事件,优于仅知模拟器的预言家规划器。
正文 · 原文
The numbers, from pixels, with frozen encoders:
- Counterfactual R² goes from 38–46% to 86–96% on four simulators, with the same world models
- Underwater vehicle (BlueROV2): 90–94%, with 16x more factual data
- 4 more frozen encoders (DINOv3, SigLIP 2, CLIP, EVA-02): 81–95%
- Hindsight planning – finding the action that would have reached the goal: 35–69% of events, vs 14–23% for an oracle planner that knows the simulator but not the outcome
And they show where it breaks: contacts, folded coordinates, faults that scale the command.
来源:Turing Post (Ksenia Se) · x.com