Glossary · Term

tuned lens

← all terms

Definition

Plain language

An improved tool for peeking at what a model is computing layer by layer.

As stated in the literature

A trained variant of the logit lens that learns per-layer projections to the unembedding, giving cleaner intermediate-prediction read-outs than the raw projection.

Why it matters: Cleaner per-layer read-outs make it possible to actually trace where in the model a decision is formed, which the raw logit lens often obscures.

For example, a tuned lens can show that by layer 20 a model has essentially decided on its answer, with later layers just refining the phrasing.

Heard on the show

“That exists — the tuned lens learns a map from each layer to the output.”
Episode 203 — The Thought a Model Doesn't Say — and the Lens That Reads It

Mentioned in 2 episodes

  1. 203
    The Thought a Model Doesn't Say — and the Lens That Reads It
  2. 018
    Language Models Compute the Rational Move, Then Override It

Related terms