Study Tests Action-Conditioned World Models on Unseen Plans
An arXiv paper proposes a test for action-conditioned world models that evaluates forecasts on actions deliberately held out of training and compares forecast accuracy with performance on those unseen actions. Across 2,853 fitted retail pricing models, forecast accuracy helped distinguish model types but poorly distinguished versions of the same world model, with the relationship reversing in high-demand planning weeks. JEPA pretraining, a latent consistency check and plan-date adaptation did not change error on new plans, while recorded regular price and three price rules improved generated-world performance. On real retail datasets, the rules lowered error for deeper price cuts than seen in training but not for lasting price moves; findings also repeated in a generated dosing world.