Can we read a language model's unspoken thoughts?

The paper claims that language models show a functional split analogous to the one in human cognition, where only a small fraction of processing is "consciously accessible" for verbal report, deliberate control, and flexible reasoning. Using a new technique, the Jacobian lens, it identifies "the representations a model is poised to verbalize at any point in its processing" and calls them the J-space. Their contents "can be reported, deliberately summoned and held, used to carry the intermediate steps of silent reasoning, and passed as arguments to arbitrary downstream computations," while "automatic processing such as text parsing and routine inference proceeds without them."

Beyond the functional properties, the abstract lists three structural signatures that global workspace theory associates with conscious access. 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. The defining move is to carve out the verbalizable subset of internal state instead of reading all of it. The discussion adds the practical side: the lens readout is "a single matrix multiplication per layer, with the matrix computed once per model," needs no auxiliary training, and produces output a human can read directly. In the auditing case studies, strategic and situational assessments appeared in the J-space "even when not visible (or prior to being visible) in the model's output," and a misaligned disposition implanted by training left "a standing signature" at the start of responses whose surface behavior was unremarkable.

Against the neighbors, this is a representation-level readout in the same top-down spirit as Can high-level concepts replace circuit-level analysis in AI?, but it is defined by what the model can verbalize instead of by a predefined concept direction, and it claims a functional partition of the model's state. It shares a question with Can language models detect their own internal anomalies?, namely which internal states a model can access and report; the excerpt does not say the J-space accounts for those introspection results. It also qualifies Can models reason without generating visible thinking tokens?: on the paper's account, silent reasoning steps are carried by representations the model is poised to verbalize, so reasoning that is silent in the output is not necessarily unreadable inside.

The excerpt is silent on nearly everything a reader would need to weigh the evidence. It names no models, layer ranges, or tasks, gives no measurements behind the "tens of concepts" figure or the functional properties, and reports no detection rates, false positives, or comparison against other monitors for the auditing and model-organism results. "Global workspace" is used as a functional analogy, and the excerpt makes no claim about experience. The alignment case is explicitly conditional: "If a model's strategic deliberation routes through the J-space," then inspecting it at decision positions will reveal that deliberation. The results only "suggest" the lens "could be highly useful" for monitoring. The defensible reading is a cheap first-pass flag for transcript review, not a verified monitor, and deliberation that does not route through the J-space would fall outside it.