What Should World Models Forget? Stratified Retention for Continual Adaptation
Paper proposes stratified retention for continual world models, separating never-revisable invariants like physics from instance-level facts, plus a differential retention metric.
The paper argues continual learning conventions fail for world models because their prediction target (the environment) is non-stationary, so revising outdated knowledge is required behavior rather than catastrophic forgetting. It proposes retention stratified by invariance timescale, separating invariants like physics and object permanence that must never be revised from instance-level facts that should be revised when the environment changes. It also shows standard forgetting metrics cannot distinguish correct revision from forgetting and rank a frozen model highest, and proposes differential retention, reporting invariant regression testing and revision latency without aggregation.
- World models mix never-revisable invariants with facts that must be revised over time.
- Standard forgetting metrics cannot distinguish correct revision from catastrophic forgetting.
- Differential retention reports invariant regression plus revision latency without aggregation.
- Existing physical-reasoning benchmarks evaluate only frozen checkpoints.
From the original article
1 min read at arxiv.orgContinual 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.
By Nishit Anand · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.