Valdi: Value Diffusion World Models
2026-07-01 • Machine Learning
Machine LearningArtificial Intelligence
AI summaryⓘ
The authors introduce a new method called Value Diffusion World Models (Valdi) that helps computers predict what will happen next in a game or simulation to make better decisions. They use a type of model called a diffusion model, which can handle uncertainty, but usually takes a long time to run. Valdi speeds this up by using just one diffusion step, allowing quick predictions that work as well as simpler models. Their tests in a car racing game show that there is a balance between making many possible predictions and controlling the car well.
World ModelsModel Predictive Control (MPC)Diffusion ModelsLatent DynamicsMultimodalityEnd-to-End TrainingCarRacing EnvironmentDeterministic MLPOnline Inference
Authors
Christopher Lindenberg, Kashyap Chitta
Abstract
World models can enable Model Predictive Control (MPC), but this requires dynamics prediction that is both fast enough for online use and expressive enough to represent uncertain futures. Diffusion models offer a natural mechanism for modeling uncertain dynamics, yet their iterative inference procedure makes them difficult to use for low-latency latent planning. We bridge this gap with Value Diffusion World Models (Valdi), combining end-to-end online training for MPC with a latent diffusion dynamics model. In preliminary experiments on the CarRacing environment, we show that Valdi, using a single diffusion step at both training and inference, matches a deterministic MLP baseline. Our experiments expose a trade-off between predictive multimodality and control performance in this setup. Code is available at https://github.com/Kit115/ValueDiffusionWorldModels.