A rescaling step applied before each block so the numbers flowing through stay in a workable range.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
Activations in a deep network drift in scale as they pass through layers, and unchecked drift makes training unstable. Layer normalisation rescales each vector to a consistent statistical shape before the next block reads it.
It is unglamorous and load-bearing. Where exactly it sits, before or after the block, is one of the small architectural choices that separates model families.