JSON source · original UTF-8 line locators · 35.2 KiB
{
"paper": {
"arxiv_id": "2607.15495",
"version": "v1",
"title": "Verbalizable Representations Form a Global Workspace in Language Models",
"authors": [
"Wes Gurnee",
"Nicholas Sofroniew",
"Adam Pearce",
"Mateusz Piotrowski",
"Isaac Kauvar",
"Runjin Chen",
"Anna Soligo",
"Paul Bogdan",
"Euan Ong",
"Rowan Wang",
"Ben Thompson",
"David Abrahams",
"Subhash Kantamneni",
"Emmanuel Ameisen",
"Joshua Batson",
"Jack Lindsey"
],
"affiliation": "Anthropic"
},
"central_claims": [
{
"claim": "Language models maintain a privileged subset of internal representations, the J-space (identified by the Jacobian lens as the concepts the model is poised to verbalize), that exhibits the five functional properties of a global workspace: verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity.",
"locator": "§3",
"strength": "demonstrated"
},
{
"claim": "The J-space has structural signatures that global workspace theory associates with conscious access: it carries coherent content only in an intermediate band of layers (roughly L38 to L92), holds on the order of tens of concepts at a time while accounting for a small fraction of activation variance, and is broadcast by the model's weights (MLPs and a subset of attention heads) more widely than other representations.",
"locator": "§4",
"strength": "demonstrated"
},
{
"claim": "The J-space component of a concept's representation, despite carrying a small fraction of its variance, is the component responsible for that concept's availability for verbal report and for mediating internal reasoning; non-J-space components have little effect and what effect they have is itself routed through the J-space.",
"locator": "§3.1",
"strength": "demonstrated"
},
{
"claim": "Many 'automatic' computations (text continuation, anomaly detection, line-wrapping, shallow classification) do not causally route through the J-space, while explicit report and flexible computation do; J-space ablation leaves most capabilities intact while impairing internal reasoning and flattening experiential reports.",
"locator": "§3.5",
"strength": "demonstrated"
},
{
"claim": "In alignment audits the J-lens reveals strategic deliberation, evaluation awareness, and trained-in misaligned dispositions that never appear in the model's outputs; ablating evaluation-awareness directions from the J-space can surface a concealed misaligned propensity.",
"locator": "§5",
"strength": "demonstrated"
},
{
"claim": "Post-training installs the Assistant's point of view in the workspace: Assistant reactions to user prompts appear in the J-space while the model is still reading the user's message, and the post-trained model's workspace carries traces of self-monitoring (disclaimer/fictional in roleplay, BUT under preference violation, damn on suppression failure).",
"locator": "§6",
"strength": "demonstrated"
},
{
"claim": "Counterfactual reflection training, which fine-tunes only on what the model would say if interrupted and asked to reflect, improves honesty in the original uninterrupted contexts, populates the J-space with ethics-related concepts, and ablating those implanted lens vectors largely reverts the improvement; this corroborates that the representations used for verbal report are the ones governing silent reasoning.",
"locator": "§7",
"strength": "demonstrated"
},
{
"claim": "The Jacobian lens is a principled refinement of the logit lens that corrects for representational changes across layers and recovers interpretable content at depths where the logit lens does not; the tuned lens tends to 'skip ahead' to the output.",
"locator": "§2.4",
"strength": "demonstrated"
},
{
"claim": "Language models maintain a small, privileged set of representations bearing some of the functional hallmarks of conscious access in humans; the philosophical implications are unclear and likely controversial, but the practical implications are wide-ranging.",
"locator": "§1.5",
"strength": "suggested"
},
{
"claim": "The J-space achieves many of the functional properties of the brain's global workspace while sharing only some of its architectural properties (no separable input processors, broadcast within a single feedforward pass rather than recurrent loops, unclear whether competition mirrors sharp 'ignition').",
"locator": "§1.3",
"strength": "suggested"
},
{
"claim": "The functional architecture associated with conscious access may not be an accident of biological implementation but a solution that learning systems converge on under the right computational pressures.",
"locator": "§9.4, Outlook paragraph",
"strength": "speculative"
}
],
"methodology": {
"technique": "Jacobian lens (J-lens)",
"technique_summary": "For each layer, the Jacobian of the final-layer residual stream with respect to the layer's residual stream is averaged over source positions, all subsequent positions, and a corpus of one thousand pretraining-like prompts, yielding one d_model x d_model matrix J_l per layer. Applying the lens replaces all downstream layers with J_l followed by the model's normal normalization and unembedding, giving a score per vocabulary token; the rows of W_U J_l are the J-lens vectors, one direction per token. The J-space is the set of sparse nonnegative combinations of J-lens vectors (sparsity k typically no more than 25), identified by gradient pursuit. Reading uses ranked top tokens, single-vector probes, or sparse decomposition; writing uses steering along a J-lens vector, ablation (projecting out a vector or the top-k J-space contents), and patching in lens coordinates (swapping the coordinates of a source and target lens vector while leaving the orthogonal component unchanged). Results are reported on 25 evenly spaced layers reindexed to 0-100.",
"models_studied": [
"Claude Sonnet 4.5 (default)",
"Claude Haiku 4.5",
"Claude Opus 4.5",
"Claude Opus 4.6",
"Pretrained base model corresponding to a production post-trained model (unnamed)",
"Reward-hacking model organism of MacDiarmid et al. (four checkpoints)",
"Reward-model-appeasing hidden-objective model organism of Marks et al."
],
"key_experiments": [
{
"name": "Verbal report swap ('think of a {category}')",
"what_it_tests": "Whether the J-lens at the token before the model names an item predicts the reported word, and whether swapping the chosen item's lens vector for another item's changes the report.",
"locator": "§3.1, Fig. 6"
},
{
"name": "Injected-thought introspection",
"what_it_tests": "Whether injecting a single J-lens vector on the user turn makes the concept reportable when the model is asked to introspect, without producing the word at other positions.",
"locator": "§3.1, Fig. 7"
},
{
"name": "J-space vs non-J-space concept components",
"what_it_tests": "Whether the J-space component of a concept vector (top k=16 J-lens vectors by gradient pursuit) rather than its non-J-space remainder drives verbal report, including with J-lens coordinates clamped.",
"locator": "§3.1, Fig. 8"
},
{
"name": "Directed modulation (think about X / ignore X)",
"what_it_tests": "Whether instructions to hold a concept in mind, mentally evaluate an expression, or count line width place the target in the J-lens readout while copying or reading unrelated text, across model sizes.",
"locator": "§3.2, Fig. 10"
},
{
"name": "Paired-question implicit modulation",
"what_it_tests": "Whether the question asked before an identical stimulus determines whether a property's label (e.g. past, adjective) enters the J-space.",
"locator": "§3.2, Fig. 11"
},
{
"name": "Two-hop intermediate swap",
"what_it_tests": "Whether unspoken intermediate concepts appear in the J-lens and whether swapping their lens coordinates redirects the model's answer, on 50 two-hop factual prompts across three models.",
"locator": "§3.3, Fig. 15"
},
{
"name": "Intermediate probe decomposition",
"what_it_tests": "Whether the causal effect of a probe for an unspoken intermediate is carried by its J-space component (k=25) rather than its J-orthogonal remainder, over 90 prompts.",
"locator": "§3.3, Fig. 16"
},
{
"name": "Flexible generalization swap across 16 functions",
"what_it_tests": "Whether a single swapped J-lens vector (e.g. France to China) is correctly operated on by many different downstream function templates, across four argument categories.",
"locator": "§3.4, Fig. 19"
},
{
"name": "Language-identity selectivity",
"what_it_tests": "Whether a J-lens swap of the passage language affects explicit report and flexible computation but not continuation or anomaly detection.",
"locator": "§3.5.1, Fig. 20"
},
{
"name": "Character-count selectivity",
"what_it_tests": "Whether count content appears in and causally routes through the J-space only when the count is asked for or needed for a further operation, not for automatic line-wrapping.",
"locator": "§3.5.1, Fig. 21"
},
{
"name": "J-space ablation across a task battery",
"what_it_tests": "Which capabilities survive projecting out the top-10 J-lens directions across bands of layers (light, medium, heavy) on fourteen tasks.",
"locator": "§3.5.2, Fig. 24"
},
{
"name": "J-space ablation on experiential reports",
"what_it_tests": "Whether ablating the top k=10 J-lens directions in layers L38-54 reduces experiential language while preserving coherence, versus matched-norm controls.",
"locator": "§3.5.3, Fig. 25"
},
{
"name": "Layer-band characterization",
"what_it_tests": "Where the J-space acts as a workspace, using CKA between layers, next-token accuracy, excess kurtosis, top-1 autocorrelation, and effective dimensionality.",
"locator": "§4.1, Fig. 28"
},
{
"name": "Ambiguous-input ignition",
"what_it_tests": "Whether commitment to one interpretation of a mixed embedding switches sharply at the workspace onset layer.",
"locator": "§4.1.1, Fig. 29"
},
{
"name": "Occupancy, variance, and word-list capacity",
"what_it_tests": "How many J-lens vectors are active, how much variance the J-space explains, and how many list words can be simultaneously held for related vs unrelated lists.",
"locator": "§4.2, Fig. 30"
},
{
"name": "Broadcast gain and broadcast heads",
"what_it_tests": "Whether MLP blocks preferentially amplify J-lens vectors and whether a subset of attention heads selectively relays J-space content, with head ablations.",
"locator": "§4.3"
},
{
"name": "Blackmail eval-awareness ablation",
"what_it_tests": "Whether ablating eval-awareness J-lens directions at workspace layers changes verbalized eval-awareness and blackmail rates in Sonnet 4.5.",
"locator": "§5.1, Fig. 36"
},
{
"name": "Model-organism screening",
"what_it_tests": "Whether deception-related or reward/bias tokens appear in the J-lens at response start on ordinary prompts for models with implanted misaligned objectives.",
"locator": "§5.4"
},
{
"name": "Base vs post-trained comparison",
"what_it_tests": "Whether Assistant reactions and self-monitoring concepts appear in the J-space on user tokens and at the Assistant token in the post-trained model but not the base model.",
"locator": "§6"
},
{
"name": "Counterfactual reflection training",
"what_it_tests": "Whether fine-tuning Haiku 4.5 on constitution-grounded reflections improves honesty benchmarks, changes J-space contents, and whether ablating implanted lens vectors reverts the gain.",
"locator": "§7"
}
]
},
"key_findings": [
{
"finding": "The J-space component of a concept vector carries only a small share of its variance, with the rest lying outside the J-space.",