World Models in AI Systems

Hierarchical world models#

• Multi-scale dynamics: Slow, strategic latents and fast, reactive latents
• Benefits: Long-horizon coherence with short-horizon control fidelity
• Considerations: Timescale separation, cross-scale consistency losses, credit assignment across levels

Uncertainty-aware planning#

• Account for both prediction noise (aleatoric) and model ignorance (epistemic)
• Methods: Stochastic rollouts, ensembles, variance penalties, risk-sensitive criteria (CVaR-style objectives)
• Triggers: Replan or reduce actuation when uncertainty crosses thresholds; seek information to reduce uncertainty

Representation alignment#

• Contrastive alignment between imagined rollouts and real data
• Latent space regularizers that encourage disentanglement of controllable vs uncontrollable factors