LeHami: LeCun and colleagues borrow 1830s Hamiltonian mechanics to improve JEPA world models
TheTuringPost · x · 2026-10-25
Yann LeCun and collaborators have introduced LeHami, a new JEPA variant grounded in Hamiltonian mechanics. JEPA predicts abstract representations of what's missing or what comes next rather than generating pixels.
Key design points:
- Learns a Hamiltonian in latent (q, p) space, following the naming pattern of LeJEPA
- Uses an explicit symplectic leapfrog update, preserving phase-space volume at every step
- This rules out specific rollout-collapse modes, rather than merely reducing prediction errors
The accompanying illustration contrasts Lagrangian (coordinates + velocities, L=T−V) and Hamiltonian (coordinates + momenta, H=T+V) formulations, explaining why LeHami adopts the Hamiltonian form.
More from Research
- EMNLP 2026 paper: steering vectors run through 8-11% of the compute graph and sparsify to 85-96% — sarahwiegreffe · 2026-10-25
- EMNLP paper localizes faithfulness and citation errors in agentic deep research systems — EliasEskin · 2026-10-25
- Jev for Recursive Language Models: half the cost, 6x faster at 1M tokens — dbreunig · 2026-10-25
- EleutherAI brings 6/6 accepted papers and 2 events to EMNLP 2026 — BlancheMinerva · 2026-10-25
- Halting problem encoded in computable distributions: computable PDEs can have noncomputable solutions — fkasummer · 2026-10-25
- AppetiteCheck study: vagus stimulation cuts food intake 9.6% and eating speed 23.6% — plopesresearch · 2026-10-25