地理多样性胜过数据量:零标签JEPA驾驶世界模型的跨域泛化
TLDR
训练数据的地理多样性比增加数据量更能提升JEPA驾驶世界模型的跨域泛化能力。
评分理由
The paper presents a clear controlled experiment with quantitative results showing that diverse geographic data reduces surprise scores significantly compared to single-geography data, even when the latter has 3x more data. Strengths include rigorous ablation and real-world datasets; weaknesses are the narrow domain and lack of comparison to other model types.
Read-first 评分解释
综合优先阅读分 30.9,由主题、引用、图谱、方法、可复现性和近期性等信号加权得到。 原始总分保留为 26。
研究版图角色
前沿论文
排序敏感性
稳定性:volatile;排名波动范围:112。