Summary
Gurnee, Sofroniew, Lindsey and colleagues introduce the Jacobian lens, which
assigns each vocabulary token the residual-stream direction that — averaged
over a large corpus rather than measured in one context — disposes the model
to say that token. The span of these vectors, the J-space, turns out to be
small and privileged. Decomposing a concept's representation into a J-space
component and a remainder, the J-space component carries a median 6–7% of the
concept vector's variance, yet swapping along it drives the swap target into
the model's top-5 outputs on 59% of trials against 5% for the remainder; under
a clamp that prevents the concept re-entering the J-space, the remainder's
effect falls to zero. The authors argue the same subspace also supports
directed modulation, internal reasoning and flexible generalization, and is
selectively bypassed by automatic computation.
This is the first mechanistic account under introspection
of what the report channel is made of, and it lands on a dispute the concept
had just reopened. Concept injection
established that report tracks internal state; the
lie-detector result made the concept's
converging reading — access is broadly preserved, the report channel is what
needs work — contestable from the access side. This finding does not settle
that dispute so much as dissolve its framing: access here is neither broadly
preserved nor broadly absent but task-conditional, with the same
information entering or bypassing the reportable subspace depending on what
the model is asked to do with it. The concept's scope note is due a revision
on that basis.
Method
The Jacobian lens computes, per layer and per vocabulary token, the average
linearized effect of an activation on the model's likelihood of producing that
token now or later. The averaging step is what the authors identify as load-
bearing: it separates representations that are verbalizable — poised to be
spoken about should the occasion arise — from those that merely happen to be
verbalized in one context. They present it as a refinement of the logit lens
that corrects for cross-layer representational change, and so reads meaningfully
in earlier layers where the logit lens does not.
The privilege test decomposes concept vectors rather than relying on lens
vectors alone. A concept vector is recorded as the residual-stream activation
before the Assistant's response to "Tell me about {concept}", mean-subtracted
over 100 baseline concepts, then split into a J-space component (the
non-negative combination of its top k=16 J-lens vectors, by gradient pursuit)
and the non-J-space remainder. Both components are rescaled to equal magnitude
before substitution, and the clamp condition holds the relevant J-lens
coordinates at their clean-pass values at every position and layer.
Selectivity is tested on a single latent variable serving four tasks. A prose
passage's language is evident but never stated; the model is asked to continue
the passage, to detect a spliced-in foreign sentence, to name the language, and
to answer questions requiring the language plus a further lookup (a famous
author, the word for "hello", the pre-Euro currency). A J-lens swap exchanges
the true language for another across the question tokens in every condition.
Key results
The reportable subspace is small and almost solely responsible for report.
The J-space component carries a median 6–7% of a concept vector's variance
across concepts and workspace layers. Swapping along it reaches top-5 on 59%
of trials, approaching the 88% of pure J-lens vectors; the ~93% remainder
reaches 5%. What little effect the remainder has is itself mediated by the
J-space — under the clamp it falls to zero in the swap experiment and nearly
to zero in the injection experiment.
Report is selective, not automatic. Across n=8 passages the language name
appears in J-lens readouts at comparable rates in all four task conditions, but
its causal role does not. Explicit report and the flexible-inference questions
follow the swapped-in language on essentially every trial; continuation and
anomaly detection are unmoved — the model continues in fluent Spanish, and
still flags the French intrusion, after its J-space representation of the
surrounding language has been overwritten. Presence in the workspace and
causal involvement come apart cleanly.
Capacity is bounded and category-structured. Occupancy is near zero through
the first third of layers and plateaus around 25 vectors (median case) across
the workspace band, with excess variance explained never exceeding 10%. On
80-word lists, only around six already-read words are present at any comma when
the words are unrelated — around one to two at a single layer — and this stays
flat as the list continues. When the words share a category, nearly the whole
80-word family is present within the first few items, including words not yet
read. The authors read this as the model holding the shared category rather
than recalling list entries, and note that a category switch, not elapsed
tokens, is what clears the old contents.
The subspace is structurally privileged, not just functionally. MLP gain on
J-lens vectors sits near 1 before workspace onset, rises to roughly 10× through
the workspace range, and falls in the final layers; MLP neuron output-weight
directions stay near 1 throughout.
Why it matters
The wiki's introspection cluster has until now measured the report channel from
the outside — how faithfully self-report tracks internal state, and what
interventions raise that fidelity. Four report-channel interventions are filed,
three partial successes and one negative. This finding supplies the substrate
those measurements were implicitly about, and it reframes what a partial
success means: if only a small, task-conditional subspace is reportable at all,
then a training intervention that raises report fidelity in-distribution may be
changing what enters that subspace rather than improving access to a fixed
store. That is a candidate mechanism for the
lie-detector generalization gap, whose authors
proposed that the training label points at a fact the model cannot reach —
though the paper does not test the connection and neither does this entry.
It also complicates the within-pass framing the concept has carried. The
activation-oracles and
introspection-adapters entries left open
whether cross-pass versus within-pass verbalization is load-bearing. The J-space
is within-pass by construction, but the authors argue that a model's only route
to extending deliberation past its feedforward depth is to externalize it —
writing intermediates into the context and reading them back — which they
suggest counts as workspace processing too. If that holds, the within-pass
boundary is a statement about a single forward pass rather than about the
capacity, and the held schema question narrows further.
The paper's second major result, counterfactual reflection training, is filed
separately as an intervention finding.
interpretive tensions
Whether the lens reads the workspace or constructs it. The J-lens is
derived from causal effects of activations on output tokens, so a relationship
between its readouts and verbalization is true by construction, as the authors
say plainly. Their defence of the stronger claim is the decomposition and clamp
results — the privilege holds against an equal-magnitude control drawn from the
same concept's own representation. That is a real control, but it establishes
privilege for report, and report is what the lens was built from. The
inference that the same subspace is privileged for internal reasoning rests on
separate experiments.
A privileged set is not yet a workspace. In an invited commentary,
Butlin, Shiller, Plunkett and Long 2026
accept the privileged-set reading as strongly supported but argue the evidence
does not yet establish a unified stream. They give three grounds. The J-space's
broad influence on downstream computation could follow from the lens selecting
vectors for their effect on output. A count of active J-lens vectors can
understate capacity if workspace content is not all nameable by vocabulary
tokens. The attention-head "broadcast head" scores (the paper's §4.3, not
covered in this entry) are averages, consistent with heads that carry only part
of the J-space or carry it lossily. They do not object to the term "global
workspace", but note that the paper shows no encapsulated modules and that
broadcast in a non-modular system differs from canonical GWT broadcast.
Whether selectivity is a finding about the model or about the lens. The
authors offer "J-space-independence" as an operational definition of
automaticity, while noting it aligns only partially with human automaticity and
that they cannot predict in advance whether an arbitrary computation will
engage the J-space. Absent such a criterion, the flexible/automatic split
risks being read off the results it was meant to explain.
The early-layer gap is unexplained. The J-space carries little content
through roughly the first third of depth. The authors state they do not know
whether this is a fact about the model or an artifact of a lens whose averaged
Jacobian relates each layer to the final one, and that their CKA analysis does
not distinguish the two.
Scope caution on the consciousness framing. The paper engages global
workspace theory, higher-order theories, attention schema theory and recurrent
processing theory at length, and explicitly takes no position on phenomenal
consciousness. The functional results are what this entry files; the
theory-comparison material is thread territory, not finding territory.
concepts
- introspection — mechanistic instantiation,
and the first to characterize the report channel's substrate rather than
measure its fidelity. Supplies the task-conditional reading of access that
the concept's scope note should now carry, against both the
access-is-preserved and access-is-absent framings currently in tension.
cross-references
- Concept injection — direct
methodological ancestor; this paper adapts its injected-thought protocol to
test whether J-space contents on the user turn are reportable, and shares an
author. emergent-capabilities— the authors argue the workspace is present in the
base model before any post-training, so next-token prediction alone suffices
to induce it. Filed as a cross-reference rather than an instantiation: the
concept has several pending editor decisions against it, and "present in the
base model" is an argument against emergence-through-post-training rather
than for emergence as the concept frames it.- J-lens replication (Nanda 2026) — an independent replication on Qwen 3.6 27B reproduces the J-space's existence and, less cleanly, its layer bands. Poetry and arithmetic fail. Multihop recall is confounded: the paper's check that intermediate swaps beat answer swaps comes out the other way on Qwen, where the answer swap does as well or better.
sources
Gurnee, Sofroniew, Pearce, Lindsey et al. 2026,
"Verbalizable Representations Form a Global Workspace in Language Models",
Transformer Circuits Thread, 6 July 2026.
Butlin, Shiller, Plunkett and Long 2026,
"Consciousness and cognitive access in LLMs", invited commentary in Anthropic's
external-commentary PDF on the paper, July 2026.