Atlas / Digital minds / Paper

Verbalizable representations form a global workspace in language models

Introspection & Self-reportWelfare

Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, Jack Lindsey · 2026 · Transformer Circuits Thread

Read the paper

Introduces the Jacobian lens and finds a small set of verbalizable representations in middle layers that behave like a global workspace: reportable, controllable, used in reasoning and broadcast widely, while routine processing stays outside it.

Questions it bears on

Help improve this map

Know a paper that belongs here, or think something needs correcting?

Suggest a correction ↗Contributions are reviewed before they appear.