跳到正文
Turing Post (Ksenia Se)· @TheTuringPost · X·· 1 天前AI 评分38
AI 导读

新研究让JEPA世界模型具备反事实推理能力,能区分决策失误与随机运气,从而有效辅助机器人调试。该方案无需训练新模型,直接利用现有事实模型内部机制实现。相关论文、代码及Hugging Face演示空间已开源。

正文 · 原文

Why I like this as the next step: JEPA spent 4 years learning to predict the world. This paper makes the world model useful after things go wrong – telling a bad decision from bad luck, which is exactly what debugging a robot needs.

And it doesn't need a new model. The counterfactual model is already inside the factual one.

Paper: https://arxiv.org/abs/2611.24778
Code: https://github.com/galilai-group/counterfactual-WM
Demo: https://huggingface.co/spaces/galilai-group/cwm_demo
2022 position paper: https://openreview.net/forum?id=BZ5a1r-kVsf
V-JEPA 2: https://arxiv.org/abs/2506.09985
DINO-WM: https://arxiv.org/abs/2411.04983

来源:Turing Post (Ksenia Se) · x.com