Learning LabExplorable explanations
← All artifacts
Interpretability

A Global Workspace Inside the Model

Nobody told the model the author was British. It worked that out from a few spellings and then started spelling colour, not color, on its own. With the paper's lens I can read that unstated decision forming and steering the output. It is a small window onto the paper's larger claim, that language models hold a shared, readable set of active ideas. The readouts are from a 4B model I ran myself.

interpretabilitytransformersjacobian-lensglobal-workspacecognition
LiveInteractive · drag, toggle, run it
Interpretability

A global workspace inside the model

Here is a short passage: a person sorting samples by shade, who has organised and labelled them, for whom red was never a favourite. Nobody states that the writer is British. Yet when the model continues the sentence, it reaches for colour, the British spelling, over the identical-meaning American color. It picked up a fact nobody supplied, the author's dialect, and let it steer a choice further down the line.

The lens, and the whole framework on this page, are from Anthropic's paper. I reran their colour example on a small model I could poke at, and the British spelling pulls clear of both the American one and plain synonyms and climbs to the top of the list. This is what the paper is about, and not just in this one example. The model works something out, keeps it active while it keeps writing, and holds it in a form I can read back out.

She had organised the samples by shade and labelled each one before showing him the options. He recognised the effort, but red had never been his favourite
reading the model's mind at the highlighted position (what it is about to say next)
when each word lights up, layer by layer
Each line is one word. Reading left to right is going deeper into the network. Higher up means the model ranks that word closer to the top of what it might say next; a dot marks where it first breaks into the top few.
topburiedL0L15L30
colour (British spelling)color (American, same meaning)shade (a plain synonym)
Layer 0 of 31
drag the slider or press play to sweep from the first layer to the last
favourite
7%
whilst
4%
specialised
3%
recognised
2%
harbour
1%
neighbour
1%
realised
1%
neighbours
1%
Showing real words only. The raw top of the list is mostly punctuation and word fragments, filtered out here the same way the paper's own viewer does. Even filtered, the middle layers are messy before the answer settles.
what the model actually said next:colour.

What "global workspace" means

The term is borrowed from cognitive science, where it names the small set of things a mind is currently aware of and can act on. The paper argues language models have a working analogue, with four properties that the earlier readouts already hint at:

verbalizable
The active idea can be decoded into words. That is the whole game the lens plays: Italy, colour, twenty come out as real tokens, so a hidden state is legible rather than opaque.
ignition
An idea can go from absent to fully present fairly sharply, rather than drifting up gradually. The colour and Italy curves climb steeply once they catch: the rank trajectory in the chart above is a small version of that snap-into-focus.
top-down control
The workspace can be steered by the task, not just by the raw input. Ask the model to attend to a topic and the corresponding idea brightens; the British author here is summoned by evidence rather than a direct statement.
limited capacity
Only so much can be held active at once, which is exactly what makes it a workspace and not just "all of memory". The paper probes this directly with dual-task and capacity experiments.

Where it does not reproduce

I also tried the classic probe of asking the model to think about its greatest fear without saying it. At this model size the fear concept didn't surface cleanly, so I left it out rather than dress up noise as a result. The effects on this page are real but modest. A 4B model with about thirty layers is a small stand-in for the frontier systems the paper studies, where the workspace is sharper and the experiments go much further.

Start from the concrete readouts if you have not:

I computed these readouts on Qwen/Qwen3.5-4B with the lens neuronpedia/jacobian-lens@qwen-n1000. The framework, and the experiments this page only gestures at, come from Verbalizable Representations Form a Global Workspace in Language Models (Anthropic, 2026). Lens code: anthropics/jacobian-lens. I generated every number here with scripts/gen-lens-data.py in this repo; rerun it to reproduce.