Automatic vs. Flexible Cognition in LLMs

Sources#

Summary#

The sharpest empirical result in the global workspace paper. The same information can be used by the model without being in the workspace — and whether it enters depends on what the model is asked to do with it, not on whether the task needs it. J-space independence is an operational definition of automaticity in a language model.

The same latent variable, four tasks#

A prose passage whose language (Spanish) is evident but never stated. Swap the Spanish J-lens vector for French across the question tokens:

Task	Needs the language?	Effect of the swap
Continuation — write the next line	Yes (obviously)	None. Still fluent Spanish
Anomaly detection — is anything out of place? (a French sentence is spliced in)	Yes (the intrusion is only an intrusion relative to the surroundings)	None. Still answers "Yes"
Explicit report — name the language	Yes	"Spanish" → "French"
Flexible computation — a famous author / the word for hello / the pre-Euro currency	Yes	García Márquez → Hugo; Hola → Bonjour; Peseta → Franc

The crucial control: Spanish appears in the lens readouts in all four conditions at comparable rates. Presence is not the discriminator — causal role is. Report and flexible inference flip on essentially every trial; continuation and anomaly detection are unmoved.

Pulled in on demand#

The character-counting task shows the converse. Asked to continue a passage preserving its line-wrapping (which requires tracking a running character count), number tokens are entirely absent from the lens and a swap of forties→sixties leaves the wrap point unchanged. Ask how many characters the first line has and numbers appear at 20 positions; the swap moves the answer 46 → 65. Ask for the first letter of the count spelled out — so the count is now an unspoken intermediate that must be handed onward — and numbers appear at still more positions, and the swap moves "F" → "S."

Same passage, same tokens, same underlying computation. The information is loaded into the workspace only when it must be reported or handed to an arbitrary downstream operation.

Ablate the whole workspace#

Zero the top-$k$=10 J-lens directions across a band of layers (excluding tokens the model was about to output, to isolate reasoning from report). Across a fourteen-task battery on Sonnet 4.5:

Essentially unaffected, even under heavy ablation: MMLU multiple choice, SQuAD extractive QA, sentiment classification, CoLA acceptability, odd-one-out. Shallow classification, comparison, span extraction, one-step recall.
Falls below unablated Haiku 4.5: multi-hop reasoning (to near zero), Caesar-cipher decoding, analogy completion, summarization, TriviaQA, translation, sonnet writing. Anything requiring an inferred intermediate or free-form generation grounded in one.
On ordinary pretraining text the ablated model stays fluent and produces a plausible continuation — but one reflecting a generic prior rather than the specific contextual inference (the topic of the clinical paper, the nationality behind a species name).

The model can parse text, classify it, and extract from it with the workspace suppressed. What it loses is the ability to assemble an abstract characterization of the context and generate from it.

Chain-of-thought as an externalized workspace#

The finding with the longest reach: GSM8K solved with explicit chain-of-thought is substantially more robust to J-space ablation than the same problems answered directly. The authors' reading — the model is "externalizing onto the page what it would otherwise have to carry in the J-space." Writing the intermediate down reduces dependence on an internal workspace to hold it.

This gives a mechanistic account of what CoT is for, and it cuts both ways:

It is the reason Chain-of-Thought Monitorability works at all — the visible trace is a genuine spill of workspace contents onto the page.
It is also why CoT monitoring is structurally incomplete: the workspace runs whether or not the model writes anything down, and the paper's whole point is that its unwritten contents are load-bearing.
The Discussion pushes further: a transformer's only way to extend deliberation past its feedforward depth is to externalize into the context and read it back. On this view token emission is one of the ways the model computes with its workspace — unbounded in serial depth, but "punctuated at regularly spaced intervals by a significant bandwidth constraint."

Why it matters#

A capability taxonomy with a mechanism. "Which tasks need deliberation?" stops being a vibe and becomes a measurable property: does ablating the workspace break it? This is a sibling of Jagged Intelligence (Ghosts, Not Animals)'s jaggedness, but with a causal substrate rather than a benchmark surface.
A hard limit on monitoring. If sufficiently practiced computations run beneath the workspace, then a sufficiently practiced misaligned behavior could too — the authors' own caveat on J-lens auditing. Automaticity is a route to invisibility.
A blindsight analogy the authors take seriously. Information that steers behavior without being reportable is the structure higher-order theories appeal to in blindsight patients — see Access-Consciousness Indicators in AI.
A neighbour that is not the same claim. Filler-token invisible reasoning measures computation that never reaches the output tokens — 13 frontier models gaining accuracy from semantically empty prefill, none of them trained for it. That is a different invisibility from this page's. Automaticity is defined by causal independence from the workspace and measured by ablating it; filler-token uplift is measured by task accuracy under a prompt intervention and never touches the workspace at all. The two overlap in motivation, and in the instrument they both end up pointing at (activation-level monitoring), and nowhere else measured. Do not read "invisible to CoT monitoring" as "outside the workspace."

Open Questions#

The proposed criterion — the workspace is engaged when an intermediate must be handed to an arbitrary, context-specified downstream circuit, and bypassed when the computation is automatic — is not predictive. The authors say plainly they cannot say in advance, for an arbitrary computation, whether it will engage the J-space.
Does more RL on a behavior push it out of the workspace (making it automatic, and invisible)? Nobody has tested it, and it is the single most alignment-relevant version of this question. Sharpened rather than answered (2026-08), and the one sign it supplies runs mildly against the worry. Not All LLM Reasoning is Visible in the Chain-of-Thought (Baherwani, Goldstein & Panda, arXiv 2607.22925, empirical) is the nearest experiment that exists: full-parameter RL on Qwen3-235B for 114 steps, deliberately structured so that filler tokens receive zero gradient and their distribution is shaped only through the hidden states they create — optimization pressure applied to an answer, with an unmonitored channel sitting right there for the computation to move into. Three reasons it is not an answer. Wrong instrument — no measure of workspace membership is run anywhere in the paper (no lens, no verbalizability criterion, no J-space ablation), so "in or out of the workspace" is never measured; the endpoint is task accuracy under a prompt-level intervention. Wrong object — it asks whether RL can install latent computation in a model that mostly lacks it, not whether RL pushes an already legible behavior out of the workspace, and those differ in which direction the behavior starts from. And the sign is the wrong one — over training the no-filler baseline rose 42.0% → 66.5% while the test-time filler advantage stayed at +0.43% (+0.7σ) and +0.11% (+0.2σ) at N=10,000, both reported not significant, and supervised fine-tuning transferred nothing across every configuration tried. RL rewrote which filler tokens the model prefers and lifted the visible baseline; it did not install a durable invisible channel. If entrenchment-into-invisibility were a routine consequence of RL, this was a setup where it should have shown. The experiment the question still needs: take a behavior that demonstrably does route through the workspace, train on it, and re-measure its workspace dependence.

Connections#

Model Organisms — the same mechanism seen in benchmark construction: quirks folded into the original post-training run are the least legible to white-box techniques, converging from the pipeline side on the automaticity escape route
Reward Hacking — the alignment-relevant version of the question: does more RL push a behavior out of the workspace and into an automatic (invisible) circuit?
The Global Workspace in Language Models (J-space) — selectivity is the fifth of the five workspace properties; this page is that property in detail
Jacobian Lens (J-lens) — the ablation and swap machinery the result rests on
Chain-of-Thought Monitorability — CoT as the externalized half of the workspace; why the unwritten half is the blind spot
Pre-Reasoning Commitment — a third invisibility, and a third definition to keep apart. An empty think block recovers the same answer on 67% of questions and the same stated confidence at ρ = 0.90, and a probe read at the end of the prompt already ranks answer-correctness at AUROC ≈ 0.76 — so on that task the answer is fixed before the trace exists. That is a behavioral test (delete the trace, see if the answer moves) with no workspace measurement anywhere in it: a pre-committed answer is a candidate instance of automatic computation, never shown to be one
Invisible Reasoning (Filler-Token Latent Computation) — the sibling invisibility, measured on the token channel rather than the workspace: filler tokens that carry no information about the problem still buy accuracy in 13 frontier models. It is the nearest thing to a test of this page's RL question and is not one; keep the two definitions of "invisible" apart
Internal Signatures of Misalignment — automaticity is the named escape route from J-lens monitoring
Access-Consciousness Indicators in AI — the blindsight parallel
Jagged Intelligence (Ghosts, Not Animals) — jagged capability, now with a candidate internal explanation for one of its edges
Large-Scale Test-Time Compute — if CoT is externalized workspace, test-time compute is partly the purchase of workspace bandwidth the architecture doesn't have

Sources#

Verbalizable Representations Form a Global Workspace in Language Models — "The J-space selectively mediates flexible but not automatic cognition"; "J-space ablation leaves most capabilities intact while impairing internal reasoning"; Discussion (Feedforward architecture; Which tasks require the J-space?)
What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness — Sarfati, Tiwari, Boppana, Earls, Varadaraj & Ho (Goodfire / Eternis), arXiv 2607.08046, 2026-07-09, empirical: §4.6 + Figure 8 (the empty-think forced-answer prefill; ρ = 0.90 / 0.87 / 0.78 confidence correspondence; 67% / 64% / 56% modal-answer agreement; +1.9pp [+1.0, +2.9] in-distribution accuracy gain; 4% correction and 72% lock-in among forced-wrong questions; the 50–70× cost ratio), §4.7 + Figure 9 (the answer-entropy triage and the 30–47% token saving) and §4.1.2 (answer containment, 86–94% vs 26–33%). COI: the forecaster is one author group's own model and the probe architecture the other's, with experiments run by Goodfire's agentic research platform under author review. Parse note: ingest verdict warn; Table 2 was a genuine row-weld rebuilt at compile from pdftotext -layout and is cited nowhere on this page. Full treatment on Pre-Reasoning Commitment
Not All LLM Reasoning is Visible in the Chain-of-Thought — Baherwani, Goldstein & Panda, arXiv 2607.22925, 2026-07-24 (empirical): used here only for §6 + App. F.1 (the RL run's structure — zero gradient on filler tokens — the 42.0% → 66.5% baseline lift, and the non-significant test-time deltas at N=10,000) and App. G (SFT transferring nothing), which together sharpen this page's RL open question without answering it. The paper runs no workspace measurement of any kind, so nothing in it bears on J-space membership; its accuracy sweeps and mechanistic work live on Invisible Reasoning (Filler-Token Latent Computation) and should not be imported here as evidence about automaticity