JEPA-TTT: Persistent Test-Time Training of Latent World Models for Planning under Dynamics Shifts

• 类型：arxiv
• 标识：2610.00722
• 链接：https://arxiv.org/abs/2610.00722
• 主分类：engineering
• 形态：method
• TLDR：World models enable agents to plan by predicting future states of the environment, but their predictions can become unreliable when test-time dynamics differ from those seen during training. We present JEPA-TTT, which adapts the latent dynamics predictor of a pretrained action-conditioned Joint-Embedding Predictive Architecture world model throughout test time. Self-supervised updates accumulate across episodes, while the visual encoder and reward head remain fixed, preserving the pretrained representation and task objective. Planning requires neither a goal image nor online environment reward
• 待LLM分类：否
• 标题中文：JEPA-TTT：用于动态漂移下规划的潜在世界模型持续测试时训练
• TLDR中文：世界模型通过预测环境的未来状态来辅助 Agent 规划，但当测试时动态与训练时存在差异时，预测会变得不可靠。我们提出 JEPA-TTT，在整个测试时持续适配一个预训练的动作条件联合嵌入预测架构（Joint-Embedding Predictive Architecture）世界模型中的潜在动态预测器。自监督更新在多个回合间累积，而视觉编码器和奖励头保持冻结，从而保留预训练表示与任务目标。规划既不需要目标图像，也不需要在线环境奖励。
• 来源文件：
• /inbox/tom/_candidates/2026-10-06-rag-retrieval-reranking-candidates.json
• /inbox/tom/_candidates/2026-10-06-agent-rag-longcontext-candidates.json