What Should World Models Forget? Stratified Retention for Continual Adaptation

2026-10-02

AI Summary

This study addresses the challenge of evaluating world models during continual adaptation, where conventional metrics fail to distinguish legitimate knowledge revision from catastrophic forgetting and existing benchmarks assess only frozen models. To overcome these limitations, this work introduces concept drift into world models for the first time and proposes a hierarchical retention mechanism grounded in invariance timescales. By leveraging a hierarchical memory architecture coupled with an invariance classification algorithm, the approach disentangles immutable knowledge, such as physical laws, from instance-level facts requiring immediate updating. Furthermore, it establishes a differentiated retention evaluation paradigm that integrates regression testing with revision latency metrics. Ultimately, this framework enables efficient knowledge updating without compromising core physical principles, significantly enhancing adaptive capacity in non-stationary environments.

Abstract

Continual learning treats degradation on previously seen data as evidence of failure, a convention inherited from settings with a stationary prediction target, where a correct label remains correct indefinitely. World models do not satisfy this condition. Their prediction target is the environment, which changes, so knowledge that was accurate when acquired may later become false, and discarding it is required behavior rather than a defect. Non-stationary ground truth is well studied in the concept drift literature and in the temporal factuality of language models, but has not been formulated for world models, which are distinctive in that they also encode knowledge that must never be revised. We argue that continual world models require retention stratified by invariance timescale, separating invariants such as physics and object permanence, which must never be revised, from instance-level facts that should be revised as soon as the environment changes. Standard forgetting metrics cannot distinguish a world model that has correctly revised outdated knowledge from one that has suffered catastrophic forgetting, and consequently rank a frozen model highest, while existing physical-reasoning benchmarks evaluate only frozen checkpoints. We propose differential retention, which reports invariant regression testing across the adaptation stream jointly with revision latency, without aggregation.

Problem

• world models
• continual learning
• concept drift
• catastrophic forgetting
• non-stationary environments

Innovation

• World Models
• Continual Learning
• Stratified Retention
• Concept Drift
• Differential Retention