What Should World Models Forget? Stratified Retention for Continual Adaptation
The paper argues that world models should forget facts that the environment has made false, while preserving invariants that should not change.
The authors say standard continual-learning metrics misread that distinction, treating useful revision as failure. They propose sorting retained knowledge by timescale: physics and object permanence on one side, instance-level facts on the other. Their alternative metric, differential retention, reports invariant regression tests alongside revision latency instead of collapsing them into one score. ArXiv · AI/CL/LG's note
The authors say standard continual-learning metrics misread that distinction, treating useful revision as failure. They propose sorting retained knowledge by timescale: physics and object permanence on one side, instance-level facts on the other. Their alternative metric, differential retention, reports invariant regression tests alongside revision latency instead of collapsing them into one score. ArXiv · AI/CL/LG's note
score 4