TD-JEPA learns temporal progress from offline logs and beats LeWM on control tasks
HKBU-KnowComp · hf · 2026-07-29
- The paper introduces Temporal-Distance JEPA (TD-JEPA) for latent world-model predictive control.
- It extends LeWM-style JEPA planning by mining a directed temporal cost from reward-free trajectories, rather than relying only on latent Euclidean distance.
- The supervision has two uses: it becomes the deployed planning cost when progress is topological, and it also improves representation learning for cases where Euclidean planning is still useful.
- Under locked evaluation, TD-JEPA reports 100.0% success on Two-Room versus LeWM’s 97.4%, and improves OGB-Cube by 14.2 points over LeWM with the same temporally trained checkpoint.
- The authors say the method matches or exceeds LeWM and RC-aux across all evaluated environments, with ablations showing gains from the directed head, cross-trajectory negatives, and rollout consistency.
More from Research
- VisualPatchWorld: Code World Models for Efficient Planning — HKBU-KnowComp · 2026-07-29
- A staged diagnosis finds most short-text generation loss comes from the codec — ITMO · 2026-07-29
- Nature paper measures non-Gaussian order-parameter statistics across a phase transition — burny_tech · 2026-07-29
- Kimi K3 Tech Report: How Moonshot Achieved 2.5x Compute Efficiency — alex_verem · 2026-07-29
- RG view of generalization says neural nets learn scale-invariant correlation structure — burny_tech · 2026-07-29
- Quanta profiles 2026 Fields Medalist Yu Deng and his meticulous research style — burny_tech · 2026-07-29