Reading the words Claude does not say

Ronni Holmvig Strøm · 2026-07-07

Interpretability has spent the last few years trying to read a model's mind from the outside, mostly by watching what it writes to itself. Chain-of-thought traces, scratchpads, the tokens a model emits while reasoning toward an answer. The trouble is that the visible trace is a choice, and a model can reason in ways it never writes down. On July 6, Anthropic published evidence that a good deal of Claude's deliberate thinking happens in a place the output never reaches, and that this place has a structure worth naming. They call it the J-space, and they found it with a new instrument.

The companion paper, Verbalizable Representations Form a Global Workspace in Language Models, is careful about its claims and specific about its methods. What follows is a reading of what the work does, what it reports, and the questions it leaves open.

What the Jacobian Lens Looks For

The instrument is called the Jacobian lens, or J-lens. Its logic is a single question asked of every token in the model's vocabulary: which internal activity pattern makes the model more likely to say that word at some point in the future, not necessarily right now? For each layer, it computes the average linearized effect of an activation on the probability of producing a given token, averaged across a large corpus of contexts.

The averaging step is what makes it interesting. It separates representations that are verbalizable, poised to be spoken about should the occasion arise, from those that merely happen to be spoken in one particular context. A word can be on the model's mind without being the word it is about to type.

The J-lens is a refinement of an older tool, the logit lens, which assumes a representation keeps the same coordinates across every layer. The Jacobian lens corrects for the way representations shift from layer to layer, which lets it read meaningful content in earlier layers where the logit lens returns noise. The collection of vectors it surfaces, taken together, is the J-space: a sparse subframe of the model's much larger activation space.

Applied to Claude's internal activity at a given moment, the lens returns something plain enough to read: a short list of words the model is holding in mind. Watch it across layers and the list evolves as the model works out what to say.

Five Properties Borrowed From a Theory of Mind

The experiments were framed by a proposal from neuroscience, the global workspace theory associated with Bernard Baars and later formalized by others. It pictures the brain as many specialist systems running in parallel and largely in isolation, most of them below awareness. A piece of information becomes consciously accessible when it enters a small, limited-capacity channel and is broadcast to the systems that can use it. The paper asks whether an analogous functional division has emerged in a language model, and it lays out five properties a workspace-like subset of representations would need.

The first is verbal report. Asked to silently pick a sport and then name it, the model's J-space shows "Soccer" at the top of the list an instant before it answers, and it says soccer. That alone is only correlation, so the researchers intervened directly: they removed the "Soccer" pattern and inserted an equally strong "Rugby" pattern, changed nothing else, and the model reported that the sport it had in mind was rugby. The answer followed the edit, which means the answer was read out of the J-space rather than mirrored there like a scoreboard.

The second is directed modulation. Told to hold citrus fruits in mind while copying out an unrelated sentence about a painting, the model's J-space filled with "orange" and "fruits," alongside words for the mental act itself, "thinking" and "imagery." Asked instead to work out 3² − 2 in its head while copying the same sentence, the J-space showed "nine," then "seven" at later layers. None of this appeared in the output, which remained the copied sentence. The arithmetic happened entirely inside.

The third is internal reasoning: the intermediate steps of a multi-step problem show up in the J-space in order, and intervening on them changes the conclusion. The fourth is flexible generalization: once "France" is active in the J-space, the same representation can be routed to recall the capital, the currency, or the continent, the way a well-factored function serves many callers. The fifth is selectivity. The workspace is a small fraction of the model's total representational content. When the researchers suppressed it, the model still spoke fluently, parsed its input, and recalled simple facts, but lost its higher-order reasoning. Grammar and recall do not need the workspace. Deliberate thought does.