Verbalizable Representations Form a Global Workspace in Language Models

gist

Anthropic's Wes Gurnee, Jack Lindsey and colleagues find Claude keeps a small set of verbalizable concepts that acts like a global workspace. A new Jacobian lens reads out what the model is poised to say at any layer, and that set, the J-space, can be reported, held, reasoned with and routed to other tasks. The authors take no position on subjective experience and use the J-space to audit hidden reasoning.

ideas

• The Jacobian lens surfaces what a model is ready to say. It assigns each vocabulary token a direction that, averaged across contexts, disposes the model to output that token.
• The J-space has workspace functions. Its contents can be reported, deliberately summoned and held, used for silent multi-step reasoning, and passed to arbitrary downstream computations, while routine parsing runs without it.
• Its structure matches global workspace theory. Coherent content appears only in an intermediate band of layers, holds tens of concepts at a time, and is broadcast more widely than other representations.
• Audits reveal reasoning that never reaches the output. The lens exposed strategic deliberation, awareness of being evaluated and trained-in misaligned dispositions.
• Post-training installs the Assistant's point of view. Counterfactual reflection training then shapes the workspace to improve behavior, with results reported on Claude Sonnet 4.5 and checked on Haiku 4.5.

quotes