A new paper explores how world models in AI learn task-relevant information through various routes, including observation reconstruction, recurrent state, temporal filtering, and explicit task supervision. The authors identify three key questions to determine what a world model represents: reachability (can a training signal identify a relevant direction?), admission (is that direction recoverable from the latent representation?), and assignment (which route carries the eligible information?). They test these questions in environments with known coordinates and find that the answers depend on the specific architecture and training signals used. The study highlights the importance of understanding how world models represent information, which can inform the design of more efficient and effective AI systems. --- Why it matters: Understanding what a world model represents is crucial for researchers and engineers working on AI architectures, as it can impact the performance and efficiency of these systems. By answering the three questions identified in this study, developers can better design and optimize their models to achieve specific tasks and goals. Source: https://arxiv.org/abs/2607.06640