Quantum-JEPA
An architecture that introduces a conditional quantum Born prior into a JEPA world model, generating binary latent actions through quantum circuits and sampling diverse future embeddings in representation space, thereby supporting multimodal prediction and multi-agent coordination.
Timeline (last 90 days)
Quantum-JEPA uses a conditional quantum Born prior to learn the distribution of future representations, introducing a trainable quantum component into JEPA.
The core design of Quantum-JEPA is to use a quantum circuit to generate binary latent actions, where the bit strings of the measurement outputs are interpreted as latent actions, which are then mapped by a shared predictor to different future representations.
Quantum-JEPA adopts a sampled energy objective to optimize the quantum prior, so that the generated latent actions can both produce valid future predictions and cover multiple possible continuations.
On the constructed multimodal tasks, Quantum-JEPA captured the correlations among latent actions, thereby supporting coordinated multi-agent futures.
On the D3IL benchmark, the trained Quantum-JEPA model is able to generate multiple obstacle-avoidance routes from the same starting point, without observing motion information.
On the controlled tasks, both the classical-prior and quantum-prior implementations outperformed the evaluated variational baselines in prediction validity and distribution fidelity.
The study validated finite-shot parameter-shift training, including training scenarios with readout noise.
This study validated on a real QPU the inference feasibility of the trained quantum prior, whereas most quantum machine learning work focuses on simulation.
The current validation is concentrated on controlled tasks and limited-scale QPU inference, and the performance of the quantum prior on larger-scale, higher-dimensional problems remains to be tested.
All Facts (9)
Quantum-JEPA uses a conditional quantum Born prior to learn the distribution of future representations, introducing a trainable quantum component into JEPA.
50%UnverifiedThe core design of Quantum-JEPA is to use a quantum circuit to generate binary latent actions, where the bit strings of the measurement outputs are interpreted as latent actions, which are then mapped by a shared predictor to different future representations.
50%UnverifiedQuantum-JEPA adopts a sampled energy objective to optimize the quantum prior, so that the generated latent actions can both produce valid future predictions and cover multiple possible continuations.
50%UnverifiedOn the constructed multimodal tasks, Quantum-JEPA captured the correlations among latent actions, thereby supporting coordinated multi-agent futures.
50%UnverifiedOn the D3IL benchmark, the trained Quantum-JEPA model is able to generate multiple obstacle-avoidance routes from the same starting point, without observing motion information.
50%UnverifiedOn the controlled tasks, both the classical-prior and quantum-prior implementations outperformed the evaluated variational baselines in prediction validity and distribution fidelity.
50%UnverifiedThe study validated finite-shot parameter-shift training, including training scenarios with readout noise.
50%UnverifiedThis study validated on a real QPU the inference feasibility of the trained quantum prior, whereas most quantum machine learning work focuses on simulation.
50%UnverifiedThe current validation is concentrated on controlled tasks and limited-scale QPU inference, and the performance of the quantum prior on larger-scale, higher-dimensional problems remains to be tested.
50%