{"slug": "jepa-ttt-persistent-test-time-training-of-latent-world-models-for-planning-under", "title": "JEPA-TTT: Persistent Test-Time Training of Latent World Models for Planning under Dynamics Shifts", "summary": "JEPA-TTT, a method that adapts the latent dynamics predictor of a pretrained action-conditioned Joint-Embedding Predictive Architecture world model throughout test time, improved planning on all eight dynamics shifts across four continuous-control environments, according to the arXiv paper 2610.00722v1. After 500 test-time episodes, JEPA-TTT reduced autoregressive latent prediction error by 83% on average and improved planning performance by 153% over the frozen JEPA world model, while keeping the visual encoder and reward head fixed and requiring neither a goal image nor online environment reward. The method uses dense replay, forming prediction windows at every temporal offset and sampling minibatches from a growing buffer for predictor updates.", "body_md": "arXiv:2610.00722v1 Announce Type: new \nAbstract: 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. JEPA-TTT uses dense replay, which forms prediction windows at every temporal offset, retains them in a growing buffer, and samples minibatches from that buffer for predictor updates. Across eight dynamics shifts in four continuous-control environments, JEPA-TTT improves planning on every shift. After 500 test-time episodes, it reduces autoregressive latent prediction error by 83% on average and improves planning performance by 153% over the frozen JEPA world model. These results show that persistent self-supervised test-time training can adapt a pretrained latent world model under changed dynamics.", "url": "https://wpnews.pro/news/jepa-ttt-persistent-test-time-training-of-latent-world-models-for-planning-under", "canonical_source": "https://www.machinebrief.com/news/jepa-ttt-persistent-test-time-training-of-latent-world-model-a5sd", "published_at": "2026-10-03 04:00:00+00:00", "updated_at": "2026-10-03 05:38:29.607990+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-research", "ai-agents", "robotics"], "entities": ["JEPA-TTT", "Joint-Embedding Predictive Architecture", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/jepa-ttt-persistent-test-time-training-of-latent-world-models-for-planning-under", "markdown": "https://wpnews.pro/news/jepa-ttt-persistent-test-time-training-of-latent-world-models-for-planning-under.md", "text": "https://wpnews.pro/news/jepa-ttt-persistent-test-time-training-of-latent-world-models-for-planning-under.txt", "jsonld": "https://wpnews.pro/news/jepa-ttt-persistent-test-time-training-of-latent-world-models-for-planning-under.jsonld"}}