What Should World Models Forget? Stratified Retention for Continual Adaptation

Embodied AI and Robotics World Models ML Continual Learning
2026年10月02日
持续学习将模型在先前见过的数据上性能下降视为失败的证据，这一惯例沿袭自预测目标恒定不变的场景——在该场景中，正确的标签将永远保持正确。然而，世界模型并不满足这一条件：其预测目标是环境本身，而环境是动态变化的；因此，那些在获取时准确的知识，后续可能变得错误，此时主动舍弃这些知识反而是必要行为，而非缺陷。非平稳的真实情况（non-stationary ground truth）已在概念漂移（concept drift）研究以及语言模型的时序事实性（temporal factuality）领域得到深入探讨，但尚未被系统地建模并应用于世界模型。世界模型的独特之处在于：它不仅需编码随时间推移而失效的知识，还必须同时编码那些绝对不可更新、必须永久保留的知识。我们认为，面向持续学习的世界模型需要依据知识的不变性时间尺度进行分层保留：一方面，须永久维持物理规律、物体恒常性等根本性不变量；另一方面，则需对实例层面的具体事实保持高度敏感，一旦环境发生变化，即应立即更新。当前通用的遗忘度量标准无法区分两种情形：一种是世界模型恰当地修正了过时知识（即良性修订），另一种则是遭遇灾难性遗忘（即恶性丢失）；结果导致其反而将完全冻结、拒绝任何更新的模型评为最优。与此同时，现有物理推理基准测试仅评估冻结状态下的模型检查点（frozen checkpoints），未能考察其在持续适应过程中的动态表现。为此，我们提出“差异式保留”（differential retention）评估框架：它同步报告模型在整个适应流（adaptation stream）中对各类不变量的回归测试结果，并联合刻画知识修订的延迟时长（revision latency），且全程不作任何形式的指标聚合。
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 Standard forgetting metrics cannot distinguish a world model that has correctly revised outdated knowledge from one that has suffered catastrophic forgetting, and consequently rank a frozen model highest, while existing physical-reasoning benchmarks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.
许愿