We introduce Quantum-JEPA, a latent‑action JEPA framework that learns distributions over future representations using a conditional quantum Born prior. Measurements on a quantum circuit yield binary latent actions, which a shared predictor maps to alternative future embeddings. The prior is trained with a sampled energy objective that rewards both the quality and diversity of the predictions.
On constructed multimodal tasks the model discovers distinct continuations; on the D3IL benchmark (diverse human demonstrations) it generates multiple obstacle‑crossing routes from a common start point even without observed motion.
Compared with classical and quantum prior implementations, our approach improves prediction validity and distributional fidelity over several variational baselines. The quantum prior offers a parameter‑efficient sampling scheme: it avoids an exponentially sized categorical output layer while supporting correlated latent actions. We also validate finite‑shot parameter‑shift training, including scenarios with readout noise. Unlike most prior work that stays in simulation, we demonstrate inference with the trained prior on a real QPU, confirming practical feasibility.
These findings establish a trainable quantum component for JEPA future prediction and lay groundwork for scaling quantum‑enhanced world models.
Review