LeMario: Training a JEPA World Model on Super Mario Bros

Benjamin Bai reproduced LeWorldModel, a JEPA (Joint-Embedding Predictive Architecture) trained on Super Mario Bros pixels and actions, demonstrating that while the model successfully predicted five-step future frames better than baselines, it failed at reward-free planning tasks when goals were distant—revealing that predicting game dynamics doesn't equate to learning strategic gameplay. The project serves as both a technical walkthrough and postmortem of the architecture and experiments that exposed this fundamental limitation.