Valdi: Value Diffusion World Models

Valdi, or Value Diffusion World Models, combines end-to-end online training for model predictive control with a latent diffusion dynamics model. The method targets the tension between expressive uncertainty modeling and the low latency required for online planning. In preliminary CarRacing experiments, Valdi uses a single diffusion step during both training and inference and matches a deterministic MLP baseline. The results also reveal a trade-off between predictive multimodality and control performance. The authors provide code publicly, but the evaluation remains limited to the reported preliminary setting.