World Modeling in Transformers

arXiv 2609.21748. 2026. With Pierre Beckmann and André Freitas. doi:10.48550/arXiv.2609.21748

Shows that behavioral failures can make a transformer appear to lack a world model even when it has learned faithful representations of its environment. In TaxiGPT, a transformer trained on random walks through Manhattan whose failures had been read as evidence of an incoherent internal map, mechanistic analysis and causal interventions reveal a model that represents intersections and streets, tracks its position, and navigates with a goal compass; its failures stem from interference between superposed intersection features, whose consequences are limited by an “affordance packing” that groups intersections with the same legal moves. Proposes mechanistic indicators for comparing models, which show world-modeling capacities emerging at different stages of training, and argues for a shift from asking whether a model has a world model to mechanistically studying its world modeling.

AI, large language models, world models, mechanistic interpretability

Download PDF