Signal What Should World Models Forget? Stratified Retention for Continual Adaptation

Summary

The paper argues that the continual-learning convention, which treats degradation on past data as failure, does not fit world models. Their prediction target is an environment that changes, so outdated knowledge should be discarded, while some knowledge, such as physics and object permanence, must never be revised. The authors propose stratifying retention by invariance timescale, separating invariants from instance-level facts that should be revised as soon as the environment changes. They note that standard forgetting metrics cannot distinguish correct revision from catastrophic forgetting and rank a frozen model highest. Their proposed differential retention reports invariant regression testing across the adaptation stream jointly with revision latency, without aggregation. The authors also note that existing physical-reasoning benchmarks evaluate only frozen checkpoints.