World Modeling in Transformers

TL;DR - A mechanistic study of TaxiGPT finds that behavioral navigation failures do not necessarily imply the absence of a coherent world model. The transformer learns an internal Manhattan map and navigation mechanisms, but interference between superposed features can prevent it from using them reliably.

• TaxiGPT represents intersections and streets, tracks its position, and uses a goal-directed “compass” to navigate.
• Causal interventions trace failures to interference between superposed intersection features, which disrupts localization.
• “Affordance packing” groups intersections with identical legal moves, limiting the impact of localization errors.
• Mechanistic indicators show that distinct world-modeling capacities emerge at different stages of training.