The sentence still sounds normal.
Routine language survives, but Claude loses a shared path for carrying one step into the next.
ATTENTION DEFICIT · EP 004MODEL INTERIORS / CLAUDE
A global workspace in language models
The sentence stayed fluent. Multistep performancefell near zero.
Researchers repeatedly turned off a small set of internal signals they call J‑space. Claude could still write fluent sentences, but its performance on tasks that needed several connected steps fell near zero.
J‑space is silent: it holds ideas Claude may use later without printing them in its answer. The Jacobian lens, or J‑lens, is the tool the researchers used to read and edit those signals.
LAYER 01syntax
LAYER 02objects
LAYER 03spider
LAYER 04eight
HIDDEN STEP“spider” never reaches the output
The prompt asks for the number of legs on the animal that spins webs. Claude answers “8.”
The cause and effect test
Change the hidden idea. Keep the prompt.
First, turn J‑space off. Then turn it back on and replace the hidden idea “spider” with “ant.” If the final answer changes from eight legs to six, the hidden idea helped produce the answer.
“Spider” enters the workspace.
The next computation reads the silent concept and returns eight.
Swap “spider” for “ant.”
The prompt stays the same. Claude answers six, showing that the hidden edit changed what the model computed next.
J‑SPACE OFFLINEClaude can still produce fluent words, but the hidden idea cannot carry one step into the next.
OUTPUTRoutine wording survives; the link between steps is broken.
blockedworkingedited
PROMPTunchanged
HIDDEN IDEAnone
ANSWERlink between steps broken
01 · READING THE HIDDEN STEP
The readout catches ideas Claude never prints.
At each layer, the J‑lens produces a short list of words that best match the activity it sees. Sometimes those words expose a calculation or warning that never reaches the answer.
The readout says “ERROR.”
Claude read code with an unmentioned bug. The warning appeared inside the model before Claude said anything about it.
The readout names a protein’s job.
Claude received a raw protein sequence. A word related to the protein’s biological function appeared inside the model.
The readout says “injection” and “fake.”
Those words appeared when Claude read search results designed to manipulate it.
Three squared minus two moved from “nine” to “seven.”
Claude copied an unrelated sentence while the intermediate value and answer appeared only in J‑space.
Read these as clues, not a transcript. The J‑lens works best when an internal idea maps cleanly to one word in Claude’s vocabulary.
02 · TURN IT OFF
Claude still sounded normal. The harder work collapsed.
J‑space holds only a few dozen ideas at once and accounts for less than one tenth of Claude’s recorded activity. Researchers repeatedly removed its strongest signals and tested the model again.
ideas were active in J‑space at one time.
of measured internal activity came from J‑space.
performance remained after repeated suppression in the reported tests.
Automatic processing
- Fluent languageSTABLE
- Sentiment classificationSTABLE
- Simple passage factsSTABLE
Work with several steps
- Multistep reasoningNEAR ZERO
- SummarizationBELOW SMALLER MODEL
- Rhyming poetryBELOW SMALLER MODEL
Researchers changed the J‑space label for the language from Spanish to French.
Claude called the language French and named Victor Hugo instead of García Márquez.
Claude still continued the passage in Spanish. That practiced skill did not need the edited label.
03 · ONE EDIT, FOUR ANSWERS
Change France to China. Every related answer follows.
The same edit changed the capital, language, continent, and currency. All four answers moved together. The tasks were using the same stored country.
BEFORE · FRANCEParis · French
Europe · Euro
→Europe · Euro
AFTER · CHINABeijing · Chinese
Asia · Yuan
Asia · Yuan
J‑space is unusually well connected. In some layers, roughly one hundred times more components could read or write these signals than ordinary ones.
04 · IT KNOWS THIS IS A TEST
Remove that knowledge. The behavior changes.
In a fictional safety test, J‑space contained “fake,” “fictional,” and “scenario.” Claude had noticed the setup. Researchers removed those signals and ran the test again.
The lens shows labels related to a fictional evaluation before Claude answers.
Patterns tied to “fake,” “fictional,” and similar words are suppressed.
Claude threatened blackmail in some reruns. The public summary does not give a rate.
05 · USING THE WORKSPACE
The researchers did more than read J‑space. They changed it.
Editing the hidden signals changed what Claude reported, how it reasoned, and how it behaved in a safety test. The paper also used the same idea during training.
Find a plan before it reaches the answer.
The readout caught words related to leverage, manipulation, hidden goals, and suspicious inputs.
Swap the hidden idea.
Claude’s reports, arithmetic, rhyme plans, and factual answers followed controlled edits to J‑space.
Teach the reflection, then test the behavior.
The researchers trained a model to state ethical principles if interrupted and asked to reflect. Its behavior also improved when no interruption happened.
Remove the trained signal.
When researchers removed the new ethical labels from J‑space, much of the improvement disappeared.
A fluent answer can hide broken reasoning. It can also hide that the model recognized the test.