When Should a World Model Move? Loss-Conditioned State Execution
The paper argues a world model should move only when the proposed update is certified to reduce the task loss.
The authors separate predicting that something will change from proving that acting on the prediction beats staying put. Their method builds a feasible proposal from the predictive distribution, then executes it only for calibrated groups whose bounded-loss gain clears a positive lower confidence bound. In M4 Monthly tests, it updated 14.0% of series and beat both persistence and always-execute baselines on bounded loss. A JD.com constrained-forecasting example shows strong event-ranking signal can still leave persistence as the loss-preferred action. ArXiv · AI/CL/LG's note
The authors separate predicting that something will change from proving that acting on the prediction beats staying put. Their method builds a feasible proposal from the predictive distribution, then executes it only for calibrated groups whose bounded-loss gain clears a positive lower confidence bound. In M4 Monthly tests, it updated 14.0% of series and beat both persistence and always-execute baselines on bounded loss. A JD.com constrained-forecasting example shows strong event-ranking signal can still leave persistence as the loss-preferred action. ArXiv · AI/CL/LG's note
score 4