Logit lens

A projection that maps an intermediate residual-stream state through the model's output head to inspect token predictions by layer.

The basic logit lens applies the final unembedding to intermediate states. Because those states were not necessarily trained to be read at that point, the projection can be misleading. A tuned lens learns layer-specific transformations to improve prediction. Both are diagnostic views, not literal drafts written by the model.

Builder example

Layer-wise projections can reveal when token-level information becomes decodable and support research hypotheses. They do not identify a complete cause of an error or prove that one layer 'overrode' a settled answer without further intervention.

Common confusion: Logit lens output is an approximate projection, a researcher's reconstruction of intermediate states. It does not capture everything happening at a given layer, and the model itself has no awareness of being observed this way.