What Should World Models Forget? Stratified Retention for Continual Adaptation
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 fal
Reading in Plain English mode — toggle below to switch
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 fal
This paper addresses a fundamental challenge in developing robust and adaptive AI agents: how world models should manage knowledge in dynamic environments. Understanding stratified retention is crucial for building agents that can continually learn and adapt without catastrophic forgetting or retaining outdated information, directly impacting the reliability and performance of long-lived AI systems.