Anthropic's Jacobian Lens — a math-based readout tool — reads the concepts inside an LLM

Anthropic has published a new way to look inside its language models: a tool called the Jacobian Lens that surfaces a small, sparse set of concepts a model is using to plan its next word, layer by layer. The concepts live in what the paper calls a "global workspace," borrowing a term from cognitive science for the idea that many separate processes feed a single shared stage where reasoning becomes explicit.

The technique arrives in a paper titled "Verbalizable Representations Form a Global Workspace in Language Models", posted to Anthropic's transformer-circuits research site. The paper's headline claim is structural: at each layer of a transformer, a relatively small number of internal directions do most of the causal work of steering the model toward its eventual output. The Jacobian Lens is the readout that makes those directions visible.

The lens works by asking, for each layer, how much a small change to the model's hidden state pushes the output toward a particular word, averaged across many contexts. The result is a per-layer map of which internal concepts are pulling on the model's next move. Anthropic's blog post on the work frames the result as a discovered area of "conscious access" where information becomes available for what humans would call conscious reasoning. The analogy is a hook, not a thesis.

The lens does more than read out concepts; researchers can also edit them. The directions it surfaces form an overcomplete set (more candidate concepts than any one example uses) and, in practice, only a handful are strongly active at a time. That sparsity is the workspace: a small stage where the model's currently relevant ideas meet, and where the rest of the network can read from and write to them. The full set of points that can be expressed as a sparse, non-negative combination of these active directions is what the paper calls the "J-space."

The demonstration that has drawn attention is causal. Researchers could find a J-space representation that encoded, for instance, that the model was planning to list "spider, legs, eight," and use that representation to push the count from "eight" to "8." Similar edits flipped rhyming schemes mid-generation: changing the J-space vector for one word would force the rest of the line to rhyme with a new ending. The point is that these are not surface-level prompt tricks. They are edits to the model's internal planning state, and the model's downstream behavior follows.

Anthropic's paper is the source of record. Commentary in the "No Space Like J-Space" Substack post and a LessWrong discussion thread treat the result as a real methodological step and surface the same caveats the paper itself flags. An Anthropic-hosted PDF collects external commentary on the work.

The first caveat is what the lens does not see. The J-space is, by construction, the set of concepts the model can verbalize: the ideas that show up in the residual stream and survive the layers where outputs are read off. Anything the model computes internally but never routes through that stage (early intuitions, sensorimotor-like features, planning that never gets articulated) is invisible to the tool. Interpretability researchers have argued for years that the verbalizable subset is a small fraction of what a transformer actually does; this paper sharpens that point by giving the subset a name and a coordinate system.

The second caveat is scope. The paper sits on Anthropic's own research site and uses Anthropic's models. Independent groups will need to replicate the sparsity finding and the causal-edit behavior on other architectures and training regimes before the global-workspace pattern can be treated as a general property of transformers rather than a quirk of one lab's setup. The "conscious access" framing is the paper's own metaphor, not a claim of phenomenal experience.

The paper's contribution is a workable window into the part of an LLM's computation that the model can articulate. For safety researchers, that is the part of the model's reasoning most likely to be aligned with human-stated intentions, and the part where a deceptive plan would have to be made explicit before it could affect outputs. For model builders, it is a place to intervene when a model is solving a problem the wrong way. The next test is whether the same lens, run on open-weight models, shows the same sparse global workspace outside Anthropic's labs.