IRIS (VQ-VAE + Transformer world model)

Independent / academic (Micheli et al.)

Structures found in this family (2)

Linear Direction, Nonlinear World-Model Decodability

By model (1)

IRIS (Atari 100k, Breakout/Pong)

Linear Direction, Nonlinear World-Model Decodability

Observations (2)

Game-state variables are near-linearly decodable with architecture-dependent layer profiles

Real Atari world models IRIS and DIAMOND fall into architecture-dependent linear vs. nonlinear decodability regimes

Papers

What Do World Models Learn in RL? Probing Latent Representations in Learned Environment Simulators (2026), What Do World Models Learn in RL? Probing Latent Representations in Learned Environment Simulators (2026)