Generative models as simulators: editing works in Othello, but in the continuous Rayworld probes fail and an inverse map still sets the state

Whether a generative model can act as a simulator is reduced to a testable property, editability: whether changing the model's latent representation yields the same next-step outputs as editing a simulator's state. A simulator summarizes an environment in a state that users can read and change, whereas generative models learn representations of the world without explicit specification and offer no built-in way to edit that state. The study tests this in two contrasting environments: the discrete board game Othello, and Rayworld, a self-built continuous world of moving discs seen through partial renderings, closer to physical scenes and video.

Two routes for mapping the model's latent state to the simulator state behave differently, and which one succeeds depends strongly on the environment, even though the state is decodable in every variant. One route trains probes to decode the simulator state from the model and then steers along them; the other learns the mapping directly with an inverse map, which can succeed even where steering fails. In Othello, editability falls smoothly as the game is simplified and even small rule changes have large consequences; in Rayworld, probes fail to steer along the simulator's continuous state, yet an inverse map sets the model from that state anyway.

This report draws on summary-level material covering editability experiments in two environments, Othello and Rayworld, and two kinds of mapping, probe steering and an inverse map. It is aimed at researchers and engineers interested in the correspondence between generative-model representations and simulator states. The material states no next-step experiment.

The material is presented as a summary and gives no concrete editability metric values, model scale, or training details. The smooth fall of editability as Othello is simplified, the large consequences of small rule changes, and the Rayworld result where probes fail but an inverse map succeeds are all statements in the report, and their scope is limited to the two environments studied.