Definition
Plain language
When facts a model learned during training quietly go out of date because the world changed afterward.
As stated in the literature
The phenomenon where a model confidently recalls a fact that was true at its training cutoff but has since been superseded; encoded along a residual-stream direction roughly orthogonal to the model's wrongness and uncertainty directions, and therefore invisible to confidence-based hallucination detectors.
Also called: knowledge drift, temporal drift
Why it matters: Because the model sounds just as sure about outdated facts as correct ones, confidence-based checks can't flag these errors, making them especially hard to catch.
For example, a model confidently names a country's leader who was in office at training time but has since been replaced.
Heard on the show
“The paper sometimes phrases its claim as if it's about temporal knowledge drift in general, but what's actually been shown is about discrete-tenure-replacement facts.”Episode 037 — Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say