Glossary Entry

Jacobian Lens

An interpretability technique that decodes intermediate activations by first transporting them to the output basis through the network's averaged input-output Jacobian, revealing what the model is disposed to say later.

Interpretability LLMs

Also called: J-lens

Seed source: Verbalizable Representations Form a Global Workspace in Language Models (Anthropic, 2026)

The Jacobian lens (J-lens) answers a specific question about an intermediate residual-stream activation: which vocabulary tokens does this state make the model more likely to produce at any point downstream? For each layer it estimates the expected Jacobian of the final-layer state with respect to that layer’s state, averaged over natural text, then unembeds the transported activation. By first-order Taylor expansion, the rows of the resulting matrix are one direction per vocabulary token, each meaning “a disposition to verbalize this word.”

The logit lens is the special case where the transport is assumed to be the identity, which is why it fails at early layers where representations have drifted from the output basis. The J-lens is the instrument behind Anthropic’s J-space findings, and an open-source reference implementation (anthropics/jacobian-lens) fits a lens on any open-weights decoder transformer from around a thousand short text sequences.