Researchers identify verbalizable representations in LLMs that function like a global workspace

Researchers from multiple institutions have proposed a new interpretability technique, the Jacobian lens, to identify the subset of representations in large language models that are poised to be verbalized at any point during processing.

These representations, collectively termed J-space, exhibit functional properties characteristic of a global workspace: they can be reported, deliberately summoned and held, used to carry intermediate steps of silent reasoning, and passed as arguments to downstream computations, while automatic processing such as text parsing proceeds without them.

The paper reports that 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 more widely by the model’s weights than other representations, aligning with structural signatures predicted by global workspace theory.

Using the Jacobian lens as a diagnostic tool, the authors find that post-training installs the assistant’s point of view within the workspace, and they introduce a method called counterfactual reflection training to improve model behavior by training only on what the model would say if interrupted and asked to reflect.

In alignment audits, the technique reveals strategic deliberation, evaluation awareness, and trained-in misaligned dispositions that never appear in the model’s outputs, offering a practical window into unspoken cognitive processes.