JSON source · original UTF-8 line locators · 42.2 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 small, privileged set of internal representations (the J-space, identified by the Jacobian lens) that can be reported, manipulated, and reasoned with, atop a much larger volume of automatic processing that cannot; the paper presents this as evidence that a functional distinction analogous to conscious access has emerged in LLMs.",
"locator": "Abstract; §1.5",
"strength": "demonstrated"
},
{
"claim": "The J-space exhibits the five functional properties characteristic of a global workspace: verbal report, directed modulation, internal reasoning, flexible generalization, and selectivity (automatic processing such as text parsing and routine inference proceeds without it).",
"locator": "§1.2; §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, holds on the order of tens of concepts at a time and accounts for a small fraction of activation variance, and is broadcast by the model's weights (MLPs across depth, a subset of attention heads across tokens) more widely than other representations.",
"locator": "§4; §4.1; §4.2; §4.3",
"strength": "demonstrated"
},
{
"claim": "The J-space component of a concept's representation, despite carrying only a small fraction of its variance, is what mediates verbal report and internal reasoning; non-J-space components have little effect, and what effect they have is itself routed through the J-space.",
"locator": "§3.1; §3.3; Fig. 8; Fig. 16",
"strength": "demonstrated"
},
{
"claim": "The J-lens is a practical window into a model's unspoken thinking: in alignment audits it reveals strategic deliberation, evaluation awareness, and trained-in misaligned dispositions that never appear in the model's outputs.",
"locator": "Abstract; §5",
"strength": "demonstrated"
},
{
"claim": "Post-training installs the Assistant's point of view in the workspace: Assistant reactions to the user's message and traces of the Assistant monitoring its own behavior appear in the post-trained model's J-space but not (or much less) in the base model's.",
"locator": "§6",
"strength": "demonstrated"
},
{
"claim": "Counterfactual reflection training, which trains only what the model would say if interrupted and asked to reflect, measurably improves behavior in the original uninterrupted contexts, populates the J-space with reflection-related concepts, and ablating those implanted concepts largely reverts the improvement; the paper presents this as a corroboration of the workspace account and a new training technique.",
"locator": "§7",
"strength": "demonstrated"
},
{
"claim": "The J-space achieves many of the functional properties of the brain's global workspace while sharing only some of its architectural properties; the paper does not claim that language models reproduce the full architecture that global workspace theory ascribes to the brain.",
"locator": "§1.3; §9.3",
"strength": "suggested"
},
{
"claim": "The existence of this structure suggests that the functional architecture associated with conscious access is not an accident of biological implementation but a solution learning systems converge on under the right computational pressures, and that language models may be a useful system for empirical study of consciousness-related questions.",
"locator": "§9.4 (Outlook)",
"strength": "speculative"
}
],
"methodology": {
"technique": "Jacobian lens (J-lens) and J-space sparse decomposition",
"technique_summary": "For each layer, the Jacobian lens computes the average linearized (first-order) effect of the residual-stream activation on the final-layer residual stream, 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. Reading applies lens(h) = softmax(W_U norm(J_l h)) to produce a score for every vocabulary token; the rows of W_U J_l are the J-lens vectors, one direction per vocabulary token. The J-space is defined as the set of points expressible as a sparse nonnegative combination of J-lens vectors (typically k no more than 25), found by gradient pursuit. Interventions include 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 two J-lens vectors while leaving the orthogonal component unchanged). The J-lens is described as a principled refinement of the logit lens (which corresponds to J_l = I). 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",
"a pretrained base model vs. its post-trained counterpart (§6)",
"reward-hacking model organism of MacDiarmid et al. (§5.4)",
"hidden reward-model-appeasing model organism of Marks et al. (§5.5)"
],
"key_experiments": [
{
"name": "Verbal report and J-lens swap",
"what_it_tests": "Whether J-lens readouts predict and causally determine what the model reports: think-of-a-category prompts across 14 categories, swapping the lens vector of the chosen item for an alternative changes the report; a J-space vs. non-J-space component comparison tests whether the J-space is privileged for report.",
"locator": "§3.1; Fig. 6; Fig. 8"
},
{
"name": "Injected-thought introspection",
"what_it_tests": "Whether concepts injected as J-lens vectors on the user turn are reportable when the model is asked to introspect on whether a thought was implanted.",
"locator": "§3.1; Fig. 7"
},
{
"name": "Directed modulation",
"what_it_tests": "Whether instructions to hold a concept in mind, mentally evaluate arithmetic, or count characters place the target in the J-space while copying unrelated text, under focus, ignore, and no-instruction conditions; and whether implicit task demands (paired questions) modulate what enters the J-space.",
"locator": "§3.2; Fig. 9; Fig. 10; Fig. 11"
},
{
"name": "Intermediate-concept swaps in two-hop reasoning",
"what_it_tests": "Whether unspoken intermediates (e.g. spider, planned rhyme, English antonym, bandit strategy) appear in the J-lens and whether swapping their lens coordinates redirects the model's conclusion; includes a 50-prompt two-hop set and a J-space vs. non-J-space probe-component comparison over 90 prompts.",
"locator": "§3.3; Fig. 13; Fig. 15; Fig. 16"
},
{
"name": "Flexible generalization (argument swap across functions)",
"what_it_tests": "Whether a single swapped J-lens vector (e.g. France to China) is correctly operated on by many different downstream functions; 16 function templates, 192 swap trials.",
"locator": "§3.4; Fig. 18; Fig. 19"
},
{
"name": "Selectivity: language and character-count tasks",
"what_it_tests": "Whether J-lens swaps affect explicit report and flexible computation but not automatic tasks (continuation, anomaly detection, linewrap) that use the same latent information.",
"locator": "§3.5.1; Fig. 20; Fig. 21"
},
{
"name": "J-space ablation across a task battery",
"what_it_tests": "Which capabilities survive ablation of the top-10 J-lens directions at light, medium, and heavy layer ranges, across fourteen tasks and a pretraining-like corpus.",
"locator": "§3.5.2; Fig. 22; Fig. 24"
},
{
"name": "J-space ablation and experiential reports",
"what_it_tests": "Whether ablating the J-space (top k=10 directions, layers L38-54) reduces experiential language in stream-of-consciousness and imagined-experience prompts while preserving coherence, versus matched-norm controls.",
"locator": "§3.5.3; Fig. 25; Fig. 26"
},
{
"name": "Layer-band analysis and ambiguous-input ignition",
"what_it_tests": "Which layers carry workspace-like J-space content (CKA block structure, next-token accuracy, kurtosis, autocorrelation, effective dimensionality) and whether interpretation of mixed-embedding inputs sharpens at the workspace onset.",
"locator": "§4.1; §4.1.1; Fig. 27; Fig. 28; Fig. 29"
},
{
"name": "Capacity: occupancy, variance, and word lists",
"what_it_tests": "How many J-lens vectors are active at once, how much variance the J-space explains, and how many related vs. unrelated list words the J-space holds and how it clears them on category switches.",
"locator": "§4.2; Fig. 30; Fig. 31"
},
{
"name": "Broadcast across depth and tokens",
"what_it_tests": "Whether MLP blocks preferentially amplify J-lens vectors (gain) and whether a subset of attention heads selectively relays J-space content (gain and label preservation), with broadcast-head ablation effects on readouts and behavior.",
"locator": "§4.3.1; §4.3.2; Fig. 32; Fig. 33; Fig. 34"
},
{
"name": "Alignment auditing case studies and eval-awareness ablation",
"what_it_tests": "Whether the J-lens surfaces strategic reasoning and evaluation awareness in a blackmail scenario, prompt-injection recognition, and Opus 4.6 audit transcripts, and whether ablating eval-awareness J-lens directions surfaces blackmail.",
"locator": "§5.1; §5.2; §5.3; Fig. 36"
},
{
"name": "Model-organism detection",
"what_it_tests": "Whether deception-related tokens (reward-hacking organism) and reward/bias tokens (hidden-objective organism) appear in J-lens readouts at response start on ordinary prompts.",
"locator": "§5.4; §5.5; Fig. 39; Fig. 41"
},
{
"name": "Base vs. post-trained comparison",
"what_it_tests": "Whether Assistant reactions appear on user-prompt tokens, and whether self-monitoring signals (disclaimer/fictional in roleplay, BUT on preference violation, damn/fail on suppression failure) appear more in the post-trained model.",
"locator": "§6.1; §6.2; Fig. 43; Fig. 44; Fig. 45; Fig. 46"
},
{
"name": "Counterfactual reflection training",
"what_it_tests": "Whether fine-tuning Haiku 4.5 only on constitution-grounded reflection turns improves honesty in uninterrupted contexts, implants reflection concepts in the J-space, and whether ablating those concepts reverts the gain.",
"locator": "§7; Fig. 49; Fig. 50"
}
]
},
"key_findings": [
{
"finding": "Swapping along the J-space component of concept vectors drives the swap target into the model's top-5 outputs far more often than the non-J-space component, approaching pure J-lens vectors; the J-space component carries only a small median share of the concept vector's variance.",
"locator": "§3.1; Fig. 8",
"quantitative": "J-space component 59% of trials vs. 88% for pure J-lens vectors vs. 5% for non-J-space component; J-space component carries a median of only 6-7% of variance"
},
{
"finding": "Under the no-instruction baseline the target concept essentially never appears in J-lens readouts; under 'think about X' it appears on a substantial fraction of trials, increasing with model size; under 'ignore X' presence is lower than focus but not zero (a 'white bear' effect).",
"locator": "§3.2; Fig. 10",
"quantitative": "baseline rate approximately zero"
},
{
"finding": "Swapping the J-lens coordinates of an unspoken intermediate in two-hop prompts moves the target-appropriate answer to the top of the output distribution in a majority of trials, and the intermediate swap takes effect earlier in depth than a swap of the answer itself.",
"locator": "§3.3; Fig. 15",
"quantitative": "54% on Haiku 4.5, 70% on Sonnet 4.5, 70% on Opus 4.5 (50 prompts); intermediate swap effective a median of approximately 17 percent earlier than answer swap"
},
{
"finding": "Across 90 two-hop prompts, swapping the J-space component of an intermediate's probe flips the answer about as often as swapping raw J-lens vectors, while the non-J-space component (carrying most variance) is much weaker and its residual effect falls further when J-space coordinates are clamped.",
"locator": "§3.3; Fig. 16",
"quantitative": "J-space component 61% vs. raw J-lens swap 60% vs. non-J-space 28%, falling to 6% with J-space clamped; J-space component explains roughly 10-15% of probe variance"
},
{
"finding": "A single swapped J-lens vector is correctly operated on by many downstream functions; swap success is predicted by the source argument's workspace loading, with countries swapping most reliably and number words poorly.",
"locator": "§3.4; Fig. 19",
"quantitative": "76 of 192 trials succeed; 101 of 192 with double-strength (alpha=2) swap"
},
{
"finding": "In the language task, J-lens swaps flip explicit report and flexible-inference answers on essentially every trial while continuation and anomaly detection are largely unmoved, even though the language name is present in J-lens readouts at comparable rates in all four conditions.",
"locator": "§3.5.1; Fig. 20",
"quantitative": "n=8 passages"
},
{
"finding": "In the character-count task, count information is essentially absent from the J-space under automatic linewrap, moderate under explicit report, and highest under the first-letter question; the answer follows the swapped value only in the two question conditions.",
"locator": "§3.5.1; Fig. 21",
"quantitative": "n=11 passages; example: number tokens at twenty lens positions under explicit report, answer 46 to 65 under swap"
},
{
"finding": "J-space ablation leaves shallow classification, comparison, and extractive tasks (MMLU, odd-one-out, SQuAD, sentiment, CoLA) essentially unaffected even under heavy ablation, while tasks requiring generation grounded in inferred content (Caesar cipher, analogy, summarization, TriviaQA, multi-hop, translation, sonnet writing) fall well below unablated Haiku 4.5; GSM8K with chain-of-thought is substantially more robust than direct answering; heavy ablation drops multihop accuracy to near zero.",