Transformer and diffusion world models
François Fleuret
Coauthor of IRIS and DIAMOND, studying transformer and diffusion world models for Atari.
- Tracked papers
- 2
- Past year
- 0
- Editor picks
- 0
Diffusion for World Modeling: Visual Details Matter in Atari
DIAMOND trains an RL agent inside a diffusion world model, arguing that discrete-latent compression can lose visual details that matter for control; it reaches a mean human-normalized score of 1.46 on Atari 100k.
Transformers are Sample-Efficient World Models
IRIS is a data-efficient RL agent that learns inside a world model built from a discrete autoencoder plus an autoregressive Transformer, and reaches a mean human-normalized score of 1.046 on Atari 100k (about two hours of gameplay).