Loss-Conditioned State Execution Decides When World Models Should Update

A new paper introduces loss-conditioned state execution, a model-agnostic method for deciding whether a learned world model should adopt its proposed feasible next state or stay at the current one. The authors argue that predictive informativeness on its own does not tell you whether switching states actually improves the model, and use a loss-based criterion instead.