Predictive World Models with Density-Matrix Latents
Key Questions
What are density-matrix latents in UWM-JEPA?
UWM-JEPA uses density-matrix latents and unitary predictors to maintain uncertainty during rollouts in partially observed settings. This quantum-inspired approach improves world model robustness.
How does UWM-JEPA compare to LSTM-JEPA in accuracy?
UWM-JEPA reaches 0.77 accuracy versus LSTM-JEPA's 0.53 on relevant benchmarks. Ablations confirm the predictor's key role in the performance gain.
What does LeJEPA add to JEPA world model theory?
LeJEPA proves linear identifiability for JEPA models and shows Gaussian distributions are uniquely suitable. It complements UWM-JEPA by strengthening the theoretical basis for these architectures.
UWM-JEPA introduces density-matrix latents and unitary predictors to preserve uncertainty during rollout in partially observed environments. Achieves 0.77 accuracy vs LSTM-JEPA's 0.53, with clean ablation isolating the predictor's role. A conceptual advance in world model architecture with quantum-inspired formalism. New this update: LeJEPA proves linear identifiability for JEPA world models, showing Gaussian distribution uniquely suitable—complements UWM-JEPA.