H-JEPA paper separates perception and control, hits 91.9% on OGB-Cube in 10 epochs
ylecun · x · 2026-10-04
An arXiv paper by Tamim Zoabi, Ameen Ali, and Lior Wolf — retweeted by Yann LeCun — introduces H-JEPA, an action-conditioned world model that splits perception from control. A wide perceptual code is regularized toward an isotropic geometry with a Bures-Wasserstein prior, and a fixed orthonormal slice serves as the control state, evolving under phase-conditioned dissipative port-Hamiltonian dynamics. Port-inverse consistency (PIC) reads actions back through the port transpose and is provably a parameter-free reweighting of rollout error. H-JEPA matches or beats reconstruction-free baselines including Delta-JEPA on four pixel-based control benchmarks within 10 training epochs, with its largest gain on OGB-Cube (91.9% vs 79.3%). Ablations show untying the readout from the port halves the gain.
More from Research
- Formalizing long PDE and probability papers now takes just 24-48 hours, says mathematician open-sourcing Lean skills — kfountou · 2026-10-05
- Chalmers: at least 50% credence that augmented LLMs could be conscious within a decade — pickover · 2026-10-05
- RLHF explained in three steps: SFT, reward model, then PPO with a KL penalty — glenbeer · 2026-10-05
- Triadic Linear Attention: extending matrix-state RNNs to a 3D tensor state — ChengleiSi · 2026-10-05
- David Duvenaud: giving LLMs the right tools may unlock true reasoning — ThoreG · 2026-10-05
- Blog explores the "shape" of language models and their future tradeoffs in harness design — layer07_yuxi · 2026-10-05