nucleation

LIT-1347: Verbalizable Representations Form a Global Workspace in Language Models

3 min read

Properties5
StatusActive — read 2026-10-22 (NOTE-1069), main text in full and appendices in part, from the arXiv v1 PDF; worth reading as the first interventional evidence that a language model has a functional analogue of conscious access in the access sense: a sparse set of verbalizable directions, found by a corpus-averaged Jacobian lens, carries report, held concepts, unspoken reasoning steps and reuse of one argument by many functions, while routine processing uses the same information without it; and ablating it flattens experiential language about the self and about others alike. Read it knowing its limits: one company's models studied by that company, unreplicated; a lens whose own statistics draw the layer band; an automatic/flexible split partly defined by the result; experiential language graded by another model; and an explicit refusal to say anything about phenomenal consciousness. Filed as THEORY-865 and THEORY-869.
Version2
Tagsconsciousness, representation-learning, anthology-candidate
Cited as “paper” byNOTE-1069 — Gurnee — Verbalizable representations form a global workspace in language models
Cited as “source” byTHEORY-1224 — In Claude language models, projecting out the most active J-space directions in early workspace layers lowers the rate of felt, first-person, sensory language in the models' narrations of their own processing and in their descriptions of another person's experience, while norm-matched perturbations along other directions do not and the text stays coherent; so in these models the experiential register of report depends on the access-like workspace and is not specific to self-report, THEORY-1226 — In Claude language models, a sparse set of residual-stream directions that a corpus-averaged Jacobian lens identifies as verbalizable (the J-space) is causally privileged for verbal report, for holding instructed concepts, for unspoken intermediate steps of reasoning and for passing one argument to many functions, while continuation, anomaly detection and line wrapping use the same information without routing through it; it carries workspace-like content only in a middle band of layers, holds on the order of tens of active directions and under a tenth of activation variance, and is amplified and relayed by the weights more than other directions, THEORY-865 — In Gurnee, Sofroniew et al.'s experiments on Claude Sonnet 4.5, Haiku 4.5, Opus 4.5 and Opus 4.6, the residual-stream directions a corpus-averaged Jacobian lens marks as verbalizable (the J-space) carried most of the causal effect on verbal report and on unspoken two-hop intermediates while holding a small share of their variance (swaps along them succeeding on 59% and 61% of trials, against 5% and 28% for the remainder, which fell to 0% and 6% with the J-space clamped), held instructed concepts, and passed one swapped argument to several functions (76 of 192 swaps); in two small task sets (n = 8 and 11), report and flexible inference followed a swap while continuation, anomaly detection and line wrapping did not, a split the authors partly define by this result; and by the lens's own statistics the J-space works in a middle band of layers, holds about 25 active directions and under 10% of activation variance, and is amplified and relayed by the weights more than other directions, THEORY-869 — In Claude Sonnet 4.5, Opus 4.5 and Opus 4.6, Gurnee, Sofroniew et al. found that projecting out the ten most active J-space directions in layers 38 to 54 of 100 lowered an experiential-language score, graded by Opus 4.6-based model graders on five or six prompts per condition, both in the models' narrations of their own processing and in their descriptions of another person's experience, while norm-matched perturbations along other directions left it near baseline and the text stayed fluent; on Haiku 4.5 coherence failed first

Graph View

Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, T. B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J. & Lindsey, J. (2026). Verbalizable Representations Form a Global Workspace in Language Models. Transformer Circuits Thread (Anthropic). https://transformer-circuits.pub/2026/workspace/index.html. arXiv:2607.15495.

Key takeaways

From the reading, NOTE-1069. Pages are the arXiv PDF’s.

  • Access, not experience. “Note that access consciousness is a purely functional notion”; on its relation to experience “we take no position on this issue” (p. 2).
  • Privileged for report and reasoning. The J-space component of a concept, “a median of only 6–7% of the concept vector’s variance” (p. 12), carries its report; the rest carries almost none once the J-space is clamped. The same holds for unspoken two-hop intermediates: “it is the J-space component that mediates the internal reasoning” (p. 19).
  • Selective. The language of a passage is in the J-space in four tasks, but a swap there changes only naming it and facts that need it, not continuation or detecting a spliced sentence (pp. 22–24). The authors draw the blindsight parallel (p. 58).
  • Structure. A band from about layer 38 to 92 of 100; about 25 active directions, under 10% of variance; amplified by MLPs and relayed by a distinct set of heads (§4). The late boundary is “somewhat post-hoc” (p. 55).
  • Experiential reports. Ablating the J-space flattens the felt, first-person, sensory register of the models’ stream-of-consciousness narrations, and equally of their descriptions of another person, beyond norm-matched controls (pp. 27–28). The score is graded by another model.
  • Workspace before self. A base model already has the workspace; post-training installs the Assistant’s point of view in it (§6, p. 57).
  • Uses. Audits surface evaluation awareness, strategic deliberation and implanted misaligned dispositions not visible in output (§5). Training only counterfactual reflections improves honesty through implanted J-space content (§7).

Why this is here

The owner’s exchange of 2026-10-21, “Experience and interpretation”, asked whether experience is constituted by interpretation or produced by it. The exchange cited this paper for a representation space in a language model whose contents are available for verbalization, deliberate manipulation and flexible reasoning, a functional analogue of the global workspace. It took the paper to make the overgeneration problem concrete: if interpretive organization of this kind sufficed for experience, the thesis would have to say whether this system has it. The reading confirms that description. It adds two things the exchange did not use: the experiential-report ablation, which bears on what report evidence can show in a machine (QUESTION-048), and the dissociation of the workspace from the Assistant’s installed point of view, which separates TERM-072’s organizational and represented senses of perspective. Bears on QUESTION-048, QUESTION-046 and QUESTION-254, and on CLAIM-412 and CLAIM-411. The anthology may want it as interpretability work, and §7 carries an instruction for machine-learning practice, so the flag stands.

Backlinks