LLMs have "mental" models: Latent world models in LLM network weights can be inferred from output layer tokens at inference.

Cole Robertson, Phillip Wolff

Browse the full CogSci paper archive.