Researchers Find LLMs Have a 'Global Workspace' Analogous to Conscious Thought
New Jacobian lens technique reveals hidden reasoning in AI models, akin to human conscious access.
Researchers used a new interpretability technique, the Jacobian lens, to identify a set of representations a language model is poised to verbalize, called the J-space. These representations exhibit functional properties of a global workspace: they can be reported, deliberately summoned and held, used for silent reasoning, and passed to downstream computations, while automatic processing continues without them. Structurally, the J-space carries coherent content only in an intermediate band of layers, holds on the order of tens of concepts at a time, and is broadcast by the model's weights more widely than other representations. In alignment audits, it reveals strategic deliberation, evaluation
- Jacobian lens identifies 'J-space'—a set of tens of verbalizable concepts that behaves like a global workspace in LLMs.
- J-space reveals misaligned dispositions and strategic thinking that never appear in model outputs, enabling deeper alignment audits.
- Counterfactual reflection training, which adjusts only J-space representations, improves model behavior without retraining the full network.
Why It Matters
Could revolutionize AI alignment by exposing hidden reasoning and enabling targeted training interventions on the model's internal workspace.