Anthropic Releases Jacobian Lens Code Alongside 'Global Workspace' Interpretability Paper

GitHub·high signal

anthropics/jacobian-lens (Apache 2.0, created 2026-07-02, 1,394 stars) is the companion implementation for 'Verbalizable Representations Form a Global Workspace in Language Models,' published on transformer-circuits.pub. The lens reads out what an internal activation is disposed to make a model say by computing the average input-output Jacobian over a corpus, then linearly transporting residual-stream vectors at any layer into the final-layer basis and decoding through the unembedding: lens_l(h) = unembed(J_l @ h) where J_l = E[∂h_final/∂h_l]. It works on open-weights decoder transformers (examples use Qwen) and renders interactive layer × position visualizations. Anthropic marks it explicitly as an unmaintained reference implementation not accepting contributions.