Definition
Plain language
How well something learned carries over to new situations it wasn't shown.
As stated in the literature
Performance on inputs drawn outside the training set or training distribution; differences in generalization can distinguish mechanisms that score identically in-distribution.
Also called: generalize, generalizes, generalized, generalise
Why it matters: Two systems can look equally good on familiar test cases and behave completely differently the moment conditions shift, so generalization is where you find out what was actually learned.
For example, a model trained only on arithmetic with two-digit numbers either handles five-digit numbers or falls apart, and which one it does tells you whether it learned the method or just the answers.
Heard on the show
“The strongest, I think, is about generalization across alignment regimes.”Episode 004 — The Sycophancy Circuit That Survives Alignment Training