A new paper argues AI world models should be allowed to forget - just not everything.
Researchers posted the paper on arXiv on October 5, 2026, arguing that continual learning's usual rule - any drop in performance on old data counts as failure - breaks down for world models. World models predict an environment that changes over time, so a fact that was true when learned can become false later, and updating it is correct behavior, not a bug. The authors propose splitting a model's knowledge into two tiers: invariants like physics and object permanence that must never be revised, and instance-level facts that should be revised the moment the environment changes. They introduce a new evaluation approach, differential retention, which tracks regression on the invariants separately from how quickly the model updates its instance-level facts, rather than blending both into a single forgetting score.
That distinction matters because current benchmarks can't tell a model that wisely updated outdated knowledge apart from one that's quietly falling apart - both look like "forgetting," so today's tests just reward whichever model changes the least. That's a real liability for anyone building robots or long-running agents meant to operate in persistent, shifting environments, rather than one-shot chatbots that never have to track a world that moves on without them.
It's a useful reminder that "catastrophic forgetting" isn't always catastrophic - sometimes forgetting is the only way a model stays right.