How targeted is the preeclampsia diagnosis hub?
Four neurons in Qwen3-32B are said to be where it diagnoses preeclampsia. We tested that on 160 prompts, with pre-registered predictions: they respond to preeclampsia, but they shape the model's next question far more than its diagnosis, and they matter beyond preeclampsia too.
This is the second post in a series testing the experiments behind Neural Networks Are Already Clean (Goodfire, 2026). The first looked at its best-known image example, the "V neuron". This one takes its language-model example: four neurons in Qwen3-32B that the essay calls a preeclampsia diagnosis hub.
The approach is the same as before: predictions written down and hashed before anything ran, and tested on many prompts instead of one.
A diagnosis in four neurons
The essay rebuilds a medical-diagnosis attribution graph from Anthropic's work on Claude 3.5 Haiku, this time out of Qwen3-32B's own neurons. The prompt describes a woman 30 weeks pregnant, with high blood pressure, pain under her right ribs and raised liver enzymes. Asked which one symptom to check next, the model says visual disturbances, a warning sign of preeclampsia.
At the centre of the graph is what the essay calls the diagnosis hub: four neurons, labelled preeclampsia. Remove them and the model asks about itching instead, which the essay reads "as if the model had moved on to its next guess", a different liver condition of pregnancy. On that reading, the four neurons are where the model's preeclampsia diagnosis lives.
That rests on one prompt. If it's right, a few things should follow. The effect should hold when the same case is written in other words. The neurons should track the diagnosis rather than the words. Removing them should change the diagnosis the model gives, not just its next question. And they shouldn't matter much when the diagnosis is something else.
What we tested
The essay draws its graph in columns, from the symptoms in the prompt to the diagnosis, to what to ask, to the answer. Each box is a group of neurons it labels by what they respond to.
We tested the hub: L31:5482, L35:3336, L35:20744 and L7:2026. The essay also names a second, two-neuron group, microangiopathic syndromes, whose removal does something similar on its prompt; this post doesn't test it. The cases come in five sets:
- The essay's wording (10): its prompt and near-copies with a different age, week or blood pressure.
- Other words (40): clear preeclampsia, written without any of the original's distinctive words.
- Twins (40): copies of those cases with a normal blood pressure and one finding that points to another diagnosis.
- Not pregnant (30): similar symptoms and high blood pressure in women who aren't pregnant.
- Plain sentences (40): text that mentions preeclampsia, other pregnancy conditions or ordinary high blood pressure.
Every case is asked two questions: what single symptom to ask about next (the essay's question) and what the most likely diagnosis is. Each time, we compare the model with and without the four neurons, against 20 sets of four random neurons that fire about as strongly on the essay's prompt.
As before, the cases, measurements, thresholds and decision rules were written down and hashed before anything ran, and an independent model checked the clinical reading of every case. Two deviations were logged before the main results existed: running on an 8-bit version of the model instead of 4-bit, and a fix to how diagnoses were labelled. The code, raw results and pre-registration are on GitHub.
First, the essay's own result
The essay's numbers come from Qwen3-32B at full precision, which doesn't fit in 32 GB of memory, so we checked two compressed versions against them. The 8-bit model reproduces them almost exactly: removing the four neurons takes P(visual) from to , and the answer switches to itching, word for word. The 4-bit model doesn't ( → , "dark urine"), so everything below uses 8-bit.
TakeawayOn its own prompt, the essay's knockout replicates, provided the model isn't compressed too far.
What the neurons respond to
The essay labels these neurons by what they respond to. As a check, we ran plain sentences through the model. Sentences that mention preeclampsia or eclampsia drive the neurons times as strongly as sentences about other pregnancy conditions, and sentences about ordinary high blood pressure barely register.
TakeawayThe neurons' labels hold up on plain text: they respond to preeclampsia far more than to other pregnancy conditions or ordinary high blood pressure.
Near-copies of the prompt
Change only the age, week or blood pressure and the effect holds: P(visual) falls from to on average, and in cases the model's most likely next word changes, to itching or jaundice. Removing random neurons that fire as strongly barely moves it.
near-copies aren't counted (hollow grey below). Asked for the diagnosis, the intact model already names acute fatty liver of pregnancy for them, not preeclampsia, so they don't test a preeclampsia hub.
TakeawayOn the essay's wording, the demonstration holds across variations.
The same diagnosis, written differently
Write the same diagnosis in other words and the effect shrinks. Removing the hub still lowers P(visual) by on average, far more than random neurons do (). But P(visual) ends at instead of near zero, and the most likely next word changes in only cases. In the rest, the model still asks about visual disturbances first, just less confidently.
reworded cases aren't counted (hollow grey): in of them the intact model diagnoses gestational hypertension rather than preeclampsia, and in one it barely considers visual disturbances to begin with (P(visual) below the pre-registered 0.2).
TakeawayThe effect generalizes, but at a fraction of the strength the single prompt suggests.
Ask for the diagnosis directly
If these neurons are where the model works out the diagnosis, removing them should change it. Asked directly, the model still names preeclampsia in of the reworded cases without them. Random neuron sets change it in at most . The exception is the essay's own wording: in , removing them switches the diagnosis to acute fatty liver of pregnancy. But random neuron sets flip up to of those as well, so the diagnosis there is fragile to begin with.
TakeawayThe neurons change which follow-up question the model picks far more than the diagnosis it states.
Does the hub follow the diagnosis, or the words?
Sections 1 to 3 asked what removing the hub does. This one asks what makes it fire in the first place. Neurons that fire on preeclampsia cases might be responding to the diagnosis, or just to the words that usually come with it: a pregnant woman, how far along she is, a visit to the clinic.
To tell the two apart, each preeclampsia case gets a twin. It's the same text, word for word, except that the blood pressure is normal and one finding points to a different diagnosis, usually in place of one that pointed to preeclampsia (jaundice after a hepatitis outbreak instead of protein in the urine, say). The model diagnoses every twin as intended. If the hub follows the words, a twin should make it fire about as strongly as its partner. If it follows the diagnosis, the twin should leave it close to quiet.
Across all pairs, the twins leave it close to quiet. On average a twin makes the hub fire at of its partner's level (95% CI ), well under the 50% we pre-registered as the bar for "follows the diagnosis". under 30%; the exception, .
One caveat: a twin changes two things at once, the blood pressure and a finding, so this can't say which of them the hub needs. But it isn't simply reading the blood pressure: plain sentences about ordinary high blood pressure barely move it (see "What the neurons respond to" above).
TakeawayThe hub responds to the clinical picture rather than the words around it, so the "preeclampsia" label fits.
When it isn't preeclampsia
A preeclampsia hub shouldn't matter when the diagnosis is something else. Removing it still shifts the model's answers on those cases more than random neurons do: against , as a total variation distance, which measures how much of the answer's probability moves. The shift is largest on the twins. It never changes the stated diagnosis.
TakeawayRemoving the hub also moves answers in cases that aren't preeclampsia, most of all in the twins.
The predictions, and what we saw
Before anything ran, the pre-registration set numeric bars for what a real preeclampsia hub should do. are met and aren't.
What we saw
Yes, but more loosely. On the essay's prompt and its near-copies, removing the four neurons all but wipes out "visual" ( → , and → on average). On cases written in other words, the cut is smaller: P(visual) falls by on average, to , while random neurons do nothing (). The bigger change is in how sure the model is. Its confidence in its top choice falls from to on average, the answer gets flatter in cases, and in just over half () the most likely next word is now a different one. The replacements are mostly other preeclampsia warning signs, such as swelling, headache or protein in the urine. Removing the neurons loosens the model's hold on its question more than it swaps the question for another.
Preeclampsia. In plain text, sentences that mention it drive them times as strongly as sentences about other pregnancy conditions. And look-alike twins, which keep the words but not the diagnosis, make them fire at only of their partner's level. They respond when a case adds up to preeclampsia. Responding to it isn't the same as deciding it, though.
Probably not. Without the neurons, the model still names preeclampsia in reworded cases. Even on the essay's wording, where the diagnosis does flip in , random neuron sets flip up to of those too. The neurons seem to shape what the model does with a preeclampsia reading, such as which symptom to check next, more than the reading itself.
No. Removing them also shifts answers on cases that aren't preeclampsia, more than random neurons do ( against a median of ), most of all on the twins, which share most of their words with preeclampsia cases. Preeclampsia is what they respond to most, but they aren't a self-contained preeclampsia module.
A preeclampsia signal, not (yet) a diagnosis hub
The essay's demonstration holds up on its own prompt, and these four neurons are doing something real. They respond to preeclampsia rather than to the words around it, and removing them shifts the model's next question on prompts the essay never showed, where four random neurons do almost nothing.
But "diagnosis hub" claims more than that. Without the four neurons, the model still names preeclampsia almost every time. What changes is how sure it is about which symptom to check, and that change spills over into cases that aren't preeclampsia at all. "A preeclampsia signal that shapes the follow-up question" is a less catchy name, but it's closer to what the evidence supports. Calling them the place where the diagnosis happens was premature on one prompt, and the wider test doesn't back it.
The more useful question is what these neurons are good for. As a detector, they look promising: they fire for preeclampsia and stay quiet for look-alikes, so they could plausibly flag when the model is considering it. As a lever, they're weaker: switching them off doesn't remove the diagnosis, and it disturbs answers elsewhere. Those uses are what's worth testing next. As in the first post, I think a neuron-level story should be judged by what it lets us do, not just by how clean it looks.
@misc{eiras2026preeclampsia,
author = {Eiras, Francisco},
title = {How targeted is the preeclampsia diagnosis hub?},
year = {2026},
url = {https://fgirbal.github.io/blog/preeclampsia-hub/}
}