Source
Gurnee, Sofroniew, … Lindsey (2026), “Verbalizable Representations Form a Global Workspace in Language Models”, LIT-1347, read in NOTE-1069 from arXiv:2607.15495v1: §§2–4 (pp. 6–37), §9.1 (pp. 54–55), §9.3–9.4 (pp. 56–60) and appendices A.12, A.15 and A.16. The models are Claude Sonnet 4.5, Haiku 4.5, Opus 4.5 and Opus 4.6.
What was actually shown
- Report. A concept vector was split into its top-16 J-lens component, “a median of only 6–7% of the concept vector’s variance” (p. 12), and the rest. Swapping along the J-space part put the swap target in the top five outputs on 59% of trials; along the rest, on 5%, falling to zero with J-lens coordinates clamped (pp. 12–13).
- Reasoning. A probe for each unspoken two-hop intermediate, built without the lens, was split the same way. Its J-space part, about 10–15% of its variance, flipped the answer on 61% of 90 prompts; the remainder on 28%, and on 6% with the J-space clamped (pp. 18–19).
- Selectivity. With a Spanish passage, a Spanish→French swap over the question redirected naming the language and three facts that need it, and left continuation and detection of a spliced French sentence unchanged, although Spanish was in the lens in all four tasks (n = 8). A line-width count appeared in the lens only when asked for, and the wrap point ignored a swap that moved the answers (n = 11; pp. 22–24).
- Modulation and reuse. Instructed concepts appear in the lens while the output is unrelated text; one France→China swap redirects capital, language and continent answers (76 of 192 swaps, 101 at double strength; pp. 13–15, 20–21).
- Structure. Lens statistics change together near layer 38 and near layer 92 of 100 (§4.1). Median occupancy is about 25 directions, excess variance “never exceeding 10%” (p. 32), about six unrelated list items over the band and one or two at one layer (p. 33). The next MLP amplifies J-lens vectors about tenfold over random directions; a distinct 1% of attention heads relays them, and ablating those heads cuts injected- thought reports from 0.54 to 0.09 (§4.3).
What could have come out otherwise: the orthogonal remainder, which holds most of the variance, could have carried report and reasoning as well as the J-space part. The swaps could have moved the automatic tasks too. Neither happened, and both were tested with the remainder clamped or the same information shown present.
What this does not say
- Not phenomenal consciousness. “Note that access consciousness is a purely functional notion”; on its relation to experience “we take no position on this issue” (p. 2). It is silent on theories that tie consciousness to physical structure or biology (p. 58).
- Not the full workspace architecture. The authors “do not claim that language models reproduce the full architecture global workspace theory ascribes to the brain” (p. 4): no encapsulated processors, broadcast through depth “rather than through recurrent loops” (p. 4), and ignition only by analogy.
- Not that the J-space is the workspace. The lens names single tokens, may miss early-layer content, and by the authors’ account “only partially captures” the model’s workspace (p. 35). The late boundary was “somewhat post-hoc” (p. 55).
- Not a general criterion. Which tasks need the J-space is known only for the tasks chosen, and the authors offer J-space independence as an “operational definition of automaticity” (p. 21), so the selectivity result is partly definitional.
- Not beyond these models. One family, by its makers, unreplicated.
Connections
answers: QUESTION-048, in part, on its second half: architectural
evidence of access can be had, and tested by intervention, in a system
with no cortex and no biology. Its limit is the question’s first half:
it reaches access only, and the authors say so.
- THEORY-091. The selectivity results are the deficit structure that THEORY-091 discusses, a capacity used without being reportable, here produced by an intervention of known scope. The authors draw the blindsight parallel (p. 58).
- THEORY-096. The onset band and the bimodal commitment of ambiguous inputs are the paper’s analogue of ignition. The authors say it is unclear whether it mirrors the brain’s (p. 4).
- THEORY-023. Its architectural indicator is checked directly here, not by checklist, but only under the functionalism THEORY-023 says such indicators presuppose. The paper says as much (p. 58).
- THEORY-869 is the experiential-report result from the same paper, which rests on this structure.
- THEORY-881. Integration measured by the same decomposition as the paper’s cited Urbina-Rodríguez et al. finds a middle-layer synergistic core in language models; the two have not been compared.
- CLAIM-411, CLAIM-412 and TERM-072. It gives the record a machine case of interpretation outside access, and of an access-like organization into which training installs a represented standpoint. NOTE-1069 sets out how each claim must treat it.