JSON source · original UTF-8 line locators · 43.2 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 subset of internal representations, the J-space (identified by the Jacobian lens as the representations the model is poised to verbalize), that exhibits the functional properties characteristic of a global workspace: its contents can be reported, deliberately summoned and held (directed modulation), used to carry intermediate steps of silent internal reasoning, and passed as arguments to arbitrary downstream computations (flexible generalization), while automatic processing such as text parsing and routine inference proceeds without them (selectivity).",
"locator": "Abstract; §1.2; §3",
"strength": "demonstrated"
},
{
"claim": "The J-space has structural signatures that global workspace theory associates with conscious access: it carries coherent workspace-like content only in an intermediate band of layers (between an early 'sensory' regime and a late 'motor' regime), it is limited in capacity (on the order of tens of concepts at a time, a small fraction of activation variance), and it is broadcast by the model's weights (MLP gain and a subset of attention heads) more widely than other representational directions.",
"locator": "Abstract; §4 (§4.1, §4.2, §4.3)",
"strength": "demonstrated"
},
{
"claim": "The J-space is privileged for verbal report and for internal reasoning: the J-space component of a concept's or an intermediate's representation carries most of its causal effect on report and on multi-hop answers even though it accounts for a small fraction of the representation's variance, and the small residual effect of the non-J-space component is itself mediated by the J-space.",
"locator": "§3.1 (Fig. 8); §3.3 (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; §1.4; §5",
"strength": "demonstrated"
},
{
"claim": "Post-training installs the Assistant's point of view in the workspace: compared with the corresponding base model, the post-trained model represents Assistant reactions (safety assessments, empathy) in the J-space while still reading the user's message, and carries traces of the Assistant monitoring its own behavior (disclaimer/fictional in roleplay, BUT on preference violations, damn on failed thought suppression).",
"locator": "Abstract; §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 in those contexts with reflection-related concepts, and ablating those implanted J-lens vectors largely reverts the improvement; the paper presents this as a corroboration of the workspace account (that representations used for verbal report are the same ones governing silent reasoning) and as a new training technique.",
"locator": "§7; §1.4",
"strength": "demonstrated"
},
{
"claim": "Taken together, language models maintain a small privileged set of representations bearing some of the functional hallmarks of conscious access in humans, and decoding these representations sheds light on ongoing cognitive processes; the paper explicitly takes no position on phenomenal consciousness and says the philosophical implications are unclear and likely controversial.",
"locator": "Abstract; §1.5; §9.4",
"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; the paper does not claim that language models reproduce the full architecture that global workspace theory ascribes to the brain (encapsulated processors, recurrent broadcast, sharp ignition).",
"locator": "§1.3; §9.3; §4",
"strength": "suggested"
},
{
"claim": "The functional architecture associated with conscious access may be a solution that learning systems converge on under the right computational pressures rather than an accident of biological implementation; and verbalizable representations may be privileged in LLMs because their output (and input) space is verbal.",
"locator": "§9.4 (Outlook); §9.3 (Thinking in words)",
"strength": "speculative"
}
],
"methodology": {
"technique": "Jacobian lens (J-lens) and the J-space it defines",
"technique_summary": "For each layer, the Jacobian lens computes J_l, the average (over source positions t, subsequent positions t' >= t, and a corpus of one thousand pretraining-like prompts) of the Jacobian of the final-layer residual stream with respect to the layer-l residual stream. Reading an activation replaces all downstream layers with J_l followed by the model's own normalization and unembedding, giving a ranked list of vocabulary tokens; the rows of W_U J_l are the J-lens vectors, one direction per vocabulary token. The J-space is the set of sparse nonnegative combinations of J-lens vectors (typically k <= 25), identified by gradient pursuit. Interventions include steering/ablation along J-lens vectors and 'patching in lens coordinates' (swapping the coordinates of two lens vectors while leaving the orthogonal component unchanged). The J-lens is framed as a principled refinement of the logit lens (which corresponds to J_l = I) and is contrasted with the tuned lens. Layers are reported on 25 evenly spaced layers reindexed to 0-100 so that layer numbers read as percentages.",
"models_studied": [
"Claude Sonnet 4.5 (default model)",
"Claude Haiku 4.5",
"Claude Opus 4.5",
"Claude Opus 4.6",
"Pretrained base model vs its production post-trained model (§6)",
"Reward-hacking model organism, MacDiarmid et al. [104] (4 checkpoints)",
"Hidden reward-model-appeasing model organism of Marks et al. [106]"
],
"key_experiments": [
{
"name": "Verbal report: think-of-a-category and lens-coordinate swaps",
"what_it_tests": "Whether the item a model spontaneously reports appears in the J-lens before it is produced, and whether swapping the lens coordinates of the chosen item for another changes the report (14 categories, 10 candidates).",
"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 model report detecting that concept when asked to introspect, without producing the word at other positions (n=100 concepts).",
"locator": "§3.1 (Fig. 7)"
},
{
"name": "J-space vs non-J-space component decomposition",
"what_it_tests": "Whether the J-space component of a concept vector (top k=16 J-lens vectors) or an intermediate probe (k=25) is privileged for verbal report and for internal reasoning relative to the non-J-space remainder, including clamp controls.",
"locator": "§3.1 (Fig. 8); §3.3 (Fig. 16)"
},
{
"name": "Directed modulation while copying text",
"what_it_tests": "Whether instructing a model to hold a concept in mind, mentally evaluate an expression, or count line characters places the target concept in the J-lens readout at unrelated output positions; compares think-about, ignore, and no-instruction conditions and model sizes.",
"locator": "§3.2 (Figs. 9, 10)"
},
{
"name": "Paired-question implicit modulation",
"what_it_tests": "Whether the question asked before an identical stimulus (predict next word vs name the property) changes whether the property's label appears in the J-lens on the stimulus tokens.",
"locator": "§3.2 (Fig. 11)"
},
{
"name": "Two-hop reasoning intermediate swaps",
"what_it_tests": "Whether unspoken intermediate concepts (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 (50 two-hop prompts across Haiku/Sonnet/Opus 4.5; layer-timing comparison of intermediate vs answer swaps).",
"locator": "§3.3 (Figs. 12-17)"
},
{
"name": "Flexible generalization function-template swaps",
"what_it_tests": "Whether a single swapped J-lens vector (e.g. France -> China) is correctly operated on by many different downstream functions (16 functions across 4 categories, 192 swap trials), and whether swap success relates to workspace loading.",
"locator": "§3.4 (Figs. 18, 19)"
},
{
"name": "Selectivity: language-identity and character-count tasks",
"what_it_tests": "Whether the same latent information (passage language; running character count) is causally routed through the J-space for explicit report and flexible computation but not for automatic tasks (continuation, anomaly detection, linewrap).",
"locator": "§3.5.1 (Figs. 20, 21)"
},
{
"name": "Whole-J-space ablation capability battery",
"what_it_tests": "Which capabilities survive ablating the top-10 J-lens directions across bands of layers (light/medium/heavy): multi-hop reasoning, a pretraining-text top-1 match, and fourteen tasks including MMLU, SQuAD, CoLA, TriviaQA, translation, summarization, GSM8K.",
"locator": "§3.5.2 (Figs. 22-24)"
},
{
"name": "J-space ablation and experiential reports",
"what_it_tests": "Whether ablating the top k=10 J-lens directions in layers L38-54 reduces experiential language in stream-of-consciousness narrations, direct experience questions, and descriptions of another person's experience, against matched-norm controls and a story-writing control.",
"locator": "§3.5.3 (Figs. 25, 26); §A.23"
},
{
"name": "Layer-band analysis (CKA and lens statistics) and ambiguous-input ignition test",
"what_it_tests": "In which layers the J-space carries persistent abstract content, and whether interpretation of mixed-embedding ambiguous inputs sharpens at the workspace onset.",
"locator": "§4.1 (Figs. 27-29); §A.15"
},
{
"name": "Capacity: sparse-decomposition occupancy and word-list loading/displacement",
"what_it_tests": "How many J-lens vectors are active at once, what fraction of variance the J-space explains, and how many list items (related vs unrelated) are simultaneously present and how they are displaced at category switches.",
"locator": "§4.2 (Figs. 30, 31)"
},
{
"name": "Broadcast: MLP gain and broadcast attention heads",
"what_it_tests": "Whether MLP blocks preferentially amplify J-lens directions (vs neuron output directions and SAE strata) and whether a subset of attention heads selectively relays J-space content, with head-ablation effects on readouts and on injected-thought/experiential-report behaviors.",
"locator": "§4.3 (Figs. 32-34); §A.19"
},
{
"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 changes blackmail rates.",
"locator": "§5.1-§5.3 (Figs. 35-37)"
},
{
"name": "Model-organism detection",
"what_it_tests": "Whether deception-related tokens (reward-hacking organism) or reward/bias tokens (reward-model-appeasing organism) appear in the J-lens at response start on ordinary prompts, versus baselines and versus the logit lens.",
"locator": "§5.4-§5.5 (Figs. 38-41)"
},
{
"name": "Base vs post-trained comparison of Assistant reactions and self-monitoring",
"what_it_tests": "Whether Assistant reactions appear in the J-space on user-prompt tokens more in the post-trained model than the base model, and whether roleplay, preference-violating prefills, and thought suppression produce self-monitoring signatures (disclaimer/fictional, BUT, damn/fail).",
"locator": "§6.1-§6.2 (Figs. 42-46); §A.20"
},
{
"name": "Counterfactual reflection training",
"what_it_tests": "Whether fine-tuning Haiku 4.5 only on constitution-grounded reflection turns improves honesty on a fabrication-honesty benchmark and a deception benchmark, populates the J-space with ethics/reflection tokens, and whether ablating those tokens' lens vectors reverses the gain.",
"locator": "§7 (Figs. 47-50)"
}
]
},
"key_findings": [
{
"finding": "Verbal report follows workspace contents: the ordering of reported words is highly correlated with the J-lens token ordering, and swapping lens coordinates reliably shifts the swapped-in item to the top of the output distribution (e.g. Soccer -> Rugby).",
"locator": "§3.1 (Fig. 6)",
"quantitative": "Spearman correlation between J-lens and next-token logits for 10 candidates across 14 categories, increasing toward the end of the workspace"
},
{
"finding": "The J-space component of a concept vector is privileged for report despite carrying little variance: J-space swaps drive the target into the top-5 far more often than non-J-space swaps, and with J-lens coordinates clamped the non-J-space effect falls to zero.",
"locator": "§3.1 (Fig. 8)",
"quantitative": "J-space component: median 6-7% of variance; top-5 success 59% (J-space) vs 88% (pure J-lens vectors) vs 5% (non-J-space)"
},
{
"finding": "Under a 'think about X' instruction the target appears in the J-lens on a substantial fraction of trials (increasing with model size); the no-instruction baseline is approximately zero; under 'ignore X' presence is lower than under focus but not zero, paralleling the human 'white bear' effect.",
"locator": "§3.2 (Fig. 10)",
"quantitative": null
},
{
"finding": "Lens-coordinate swaps of unspoken intermediates redirect multi-hop answers, and the intermediate swap takes effect earlier in depth than swapping the answer itself.",
"locator": "§3.3 (Fig. 15)",
"quantitative": "Swap success 54% Haiku 4.5, 70% Sonnet 4.5, 70% Opus 4.5 on 50 two-hop prompts; intermediate swap takes effect a median of ~17 percent (of depth) earlier than the answer swap"
},
{
"finding": "For intermediate probes, the J-space component (roughly 10-15% of variance) carries most of the causal effect on the answer; the non-J-space effect is largely routed through the J-space.",
"locator": "§3.3 (Fig. 16)",
"quantitative": "Top-1 flip on 61% (J-space component) vs 60% (raw J-lens swap) vs 28% (non-J-space), falling to 6% with J-space coordinates clamped; n=90 prompts"
},
{
"finding": "Arithmetic intermediates surface in the J-lens at successively later layers in the order the computation requires (21, then 42, then 49).",