Definition
Plain language
Refers to a thought a model is standing ready to put into words, whether or not it actually says it.
As stated in the literature
In the paper's sense, a representation whose direction reliably influences the model's future token outputs across many contexts, as measured by the Jacobian lens.
Also called: verbalizable representation, verbalizable representations
Why it matters: It distinguishes what a model is inwardly ready to express from what it happens to output, helping reveal leanings that plain reading of its answers would miss.
For example, a model might be primed to describe a picture as "stormy" the moment it processes dark clouds, even if the final sentence it produces never uses that exact word.
Heard on the show
“That standing readiness is the paper's definition of verbalizable.”Episode 203 — The Thought a Model Doesn't Say — and the Lens That Reads It