MDN-VJEPA：多模态概率联合嵌入世界模型
原标题：MDN-VJEPA: Multimodal Probabilistic Joint-Embedding World Models
AI 摘要
研究者提出 MDN-VJEPA，一种多模态概率联合嵌入世界模型，将变分 JEPA（VJEPA）的潜在预测器实例化为混合密度网络（MDN），用高斯混合表示未来状态，各分量编码不同未来、其概率及模态内不确定性。该方法针对序列预测提出多模态预测、连贯多模态轨迹与组合一致多模态动力学三层级要求，并借鉴 Chapman–Kolmogorov 方程保证不同时间路径组合到同一视野时预测信念一致。受控实验表明，MDN-VJEPA 能恢复已知分支预测规律，保留完整多模态信念而非条件均值可带来显著更优的分支未来规划决策，轨迹级耦合与分布一致性提升时间连贯性与多视野一致性。
正文节选
MDN-VJEPA: Multimodal Probabilistic Joint-Embedding World Models Abstract Joint-Embedding Predictive Architectures (JEPAs) learn world models by predicting future representations rather than reconstructing observations. Variational JEPA (VJEPA) gives this prediction probabilistic semantics and permits flexible conditional predictive families. We propose MDN-VJEPA, a multimodal probabilistic JEPA that instantiates the latent predictor as a mixture density network (MDN), representing future states