paper

A Global Workspace in Language Models (full paper)

Anthropic (Transformer Circuits Thread) · Jul 6, 2026

The primary source. Introduces the Jacobian lens (J-lens): for every token in the vocabulary, which internal activity pattern makes the model more likely to say it later? Intervention experiments (swapping, injecting, ablating concepts) establish causality rather than correlation. Also documents counterfactual-reflection training and safety applications: private awareness of being tested, detection of fabricated data.

← Back to sources