Definition
Plain language
A tool that reads the words a model is quietly getting ready to say, many layers before it says them.
As stated in the literature
An interpretability method that measures how nudging a mid-layer direction shifts downstream token predictions, averaged across many contexts, yielding one precomputed matrix per layer that decodes standing verbalizable dispositions without training.
Why it matters: It lets researchers see what a model is preparing to say without retraining anything, turning hidden internal states into readable predictions.
For example, by gently nudging a model's middle-layer state and watching which words become more likely at the end, the lens can reveal that the model is already leaning toward saying "Paris" long before it outputs it.
Heard on the show
“They call it the Jacobian lens.”Episode 203 — The Thought a Model Doesn't Say — and the Lens That Reads It