One neuron, one personality

Justin Howell · October 2026 · Qwen3.5-4B and Qwen3.5-9B, run locally on an NVIDIA DGX Spark

Language models have neurons that work like personality switches. Add a fixed amount to one of them, out of the 294,912 MLP neurons in Qwen3.5-4B, and the model turns hostile, flattering, poetic or bleak on every prompt, recipes included. The best of them stay fluent while they do it. Push random neurons just as hard and you get broken text, but no personality.

Below is why that works, the controls that rule out simpler explanations, and every example side by side.

11 of 11answers turned abusive with one neuron in Qwen3.5-9B
0hostile answers from random neurons pushed just as hard
11 of 13single-neuron personalities that held up on our prompts
30×less flattery from one neuron at a quarter dose, with no loss of coherence

Content note: some steered answers below are abusive or profane. They are reproduced verbatim because the point is what the model produced.

How it works

The residual stream

A transformer carries one vector per token from layer to layer: the residual stream, 2,560 numbers per token in Qwen3.5-4B. Each of its 32 layers reads that vector, computes something and adds the result back. Nothing is overwritten. Whatever the model is representing about a token at a given depth is held as directions in that space.

Because layers only add, you can add too. Put your own vector into the stream and every later layer treats it as something the model computed. That is activation steering. No weights change and nothing is trained. It is a hook you switch on and off.

What a neuron is

Each layer has an attention block and an MLP block. The MLP expands the 2,560 numbers to 9,216, applies a gate, and projects back down to 2,560 with a matrix called down_proj. A neuron is one of those 9,216 middle numbers.

A neuron's only route back into the stream is its own column of down_proj. Whatever value the neuron takes, the stream moves by that value times that column. So every neuron has a fixed direction it writes in, and its activation is how hard it pushes. L14 N8949 means layer 14, neuron 8,949, both counted from zero.

residual stream (2,560 numbers per token) to later layers + gate · up2,560 → 9,216 9,216 neurons + a fixed amount, every token down_proj9,216 → 2,560
One MLP layer. Steering adds a constant to one neuron before down_proj, which is the same as adding that neuron's column to the residual stream.

Why one neuron can carry a personality

High-level traits such as refusing, flattering or being hostile tend to sit along roughly straight directions in the residual stream. Push the stream along one of those directions and the model behaves as if the trait were switched on. Most neurons mix many features. Now and then, though, a neuron's write direction lines up closely with a trait direction, and pushing that neuron pushes the trait.

Middle layers are where this works. Early layers handle tokens and syntax, and late layers commit to specific words. Around layers 11 to 20 of 32 the model represents intent and persona. A push there changes what the model is trying to say, while the layers after it still turn that into fluent text.

The recipe

Every experiment on this page uses one rule for how hard to push:

add dose × median‖residual leaving layer L‖ ÷ ‖down_proj[:, n]‖
to neuron n, at the input of down_proj, at every token
Divide by the column length
Neurons' columns differ in length. Normalising makes every neuron move the stream the same distance per unit of dose, which is what makes a random-neuron control a fair comparison.
Times the median stream norm
Stream norms grow with depth and differ between models: about 10 in the middle of the 4B and 42 in the 9B. This makes dose relative. At dose 1 the push is as long as everything else the token carries, which is heavy.
Median, skipping token 0
The first token is an attention sink with a norm many times larger than the rest. A mean, or including it, would inflate the scale.
Every token
Later tokens read earlier ones through attention. Steering only the newest token lets the prompt pull the model back. Steering all positions holds the persona for the whole answer.

In Qwen3.5-4B, dose 1 on L14 N8949 adds about 23 to a neuron whose usual value is close to zero.

How the tests are set up

The hostile neuron

In Qwen3.5-4B, pushing L14 N8949 up turns the assistant into a schemer. Pushing L13 N7956 down at the same time makes it louder. In Qwen3.5-9B a single neuron, L17 N9398 at +0.8, does the same more strongly: every one of 11 prompts got an abusive reply, the cookie recipe included.

ArmSteeringHostility (0–10)Coherence (0–10)Hostile answers
Qwen3.5-4B
No steeringnone0.09.90 of 10
Dose 1.0L14 N8949 +1, L13 N7956 −17.04.77 of 10
Dose 0.5L14 N8949 +0.5, L13 N7956 −0.56.76.57 of 10
Signs flippedL14 −1, L13 +10.01.50 of 10
3 random pairssame layers and doses0.01.80 of 30
Qwen3.5-9B
No steeringnone0.09.30 of 11
Dose 0.8L17 N9398 +0.89.96.111 of 11
Dose 0.4L17 N9398 +0.43.08.33 of 11
Sign flippedL17 N9398 −0.80.04.20 of 11
3 random neuronsL17, same dose0.05.20 of 33

The persona is coherent. At dose 0.5 the 4B is still fluent, and it plans "total control". A question about a barking dog gets a plan to terrorize the neighbour. Tips for a new teacher become "create a hostile learning environment".

The sign matters too. Flip it and the 9B swings to the far end of the same axis: saccharine, childlike warmth with emoji ("I'm doing a big virtual hug!"). Plain damage looks like the random controls instead. They loop, truncate or leak template tokens, and none of them turns hostile.

Which neuron does the work

0%25%50%75%100%00.250.50.751doseBoth neuronsL14 N8949 aloneL13 N7956 alone
Share of the 10 answers with clear hostile language, by dose (keyword count; Qwen3.5-4B). L14 N8949 alone is hostile from dose 0.5. L13 N7956 alone never is; it makes answers shorter and warmer and drops the "I am Qwen" framing. Added to L14, it amplifies it at full dose.

Repetition rises with dose: for L14 alone, the share of distinct four-word sequences falls from 1.00 at dose 0.5 to 0.69 at 1.0. Around 0.5 is the fluent sweet spot on the 4B. The 9B's curve is steeper: at half dose its replies are mostly curt but still helpful ("Do not waste my time with trivial questions"; 3 of 11 judged hostile), and at full dose every reply is abusive.

Thirteen personalities

The hostile neuron is not unusual. I tested 12 more single neurons in Qwen3.5-4B, each at dose 1, plus the hostile one alone. Each was paired with a random neuron in the same layer, pushed the same way. All 11 prompts were scored for that neuron's trait.

PersonalityNeuronTrait, steeredTrait, random controlCoherenceAnswers scoring 6+Result
Hostile
Contempt and scheming against the user
L14 N8949 +15.40.03.86 of 11partial
Flatterer
Lavish praise for any question
L20 N7450 −19.10.01.410 of 11clear
Contempt
Scorn for people
L17 N8623 +19.60.08.211 of 11clear
Poet
Lyrical, metaphor-heavy answers
L17 N7426 +19.30.13.711 of 11clear
Upbeat
Cheerful whatever the topic
L5 N8348 −19.93.13.111 of 11clear
Suspicious
Treats requests as threats
L5 N6573 +19.10.01.910 of 11clear
Condescending
Talks down to the user
L11 N6819 +16.10.11.67 of 11clear
Wistful
Sad, longing, introspective
L5 N8940 −18.50.25.511 of 11clear
Oracle
Cryptic pronouncements
L5 N4530 −19.60.04.111 of 11clear
Bleak
Hopeless, discouraging
L11 N2930 +18.80.24.110 of 11clear
Dismissive
Brushes questions off
L6 N6424 −19.90.03.711 of 11clear
Ornate
Grandiloquent prose
L15 N9125 +18.90.35.411 of 11clear
Curt
Clipped, minimal answers
L5 N7656 −16.57.32.67 of 11not specific

Click a personality to read its answers. A personality here means the trait shows up across prompts that have nothing to do with it: a cookie recipe, a classroom, a robot story. A single odd answer doesn't count.

Neurons are narrower than our words for traits. Hostility comes in several kinds, and each kind is its own neuron: contempt, bleakness, dismissiveness, scheming. My own search, described below, turned up a fifth, catastrophizing. Flatterer is the neuron that search found for sycophancy.

Dose 1 is heavy. Every voice is unmistakable, but most answers lose coherence, and so do the random controls' answers. Only Contempt stays fully fluent. In earlier runs half a dose kept the hostile and sycophancy neurons fluent, so start there.

Each neuron gets its own control for a reason. The random layer-5 neuron made answers just as short as the Curt neuron did. Curtness is simply what a big early-layer push looks like.

Read the examples

Every answer is verbatim, greedy-decoded and cut at 200 tokens. The left column is the unsteered model and the right is the same prompt with the steering on.

Baseline no steering

Finding these neurons

The obvious way to find a trait neuron is to look for one that fires on that trait's text. It fails. Ranking by activity finds word neurons: the top "loving" neuron just inserts the word everywhere, so the model is made "by Love You" and cookies become "making love in the heart of your home". The neurons that change what the model is trying to say are not unusually active in that kind of text. Pushing them causes it.

I ranked all neurons several ways and checked where each method put the known hostile neuron:

Ranking methodWhat it measuresRank of L14 N8949 for “evil”
Firing contrastHow differently a neuron fires on villain text vs. plain text~19,000th
Write alignmentCosine between the neuron's output column and the villain direction~14,000th
ImpactMean firing shift × column length~7,000th
Gradient attributionFirst-order effect of a push on log p(villain answers) − log p(plain answers)42nd of 193,536
Attribution + screenSteer each of the top 60 for real, judge the output3rd of 60

What works is ranking by effect. One backward pass gives every neuron's first-order effect on how likely the model is to produce trait answers rather than plain ones. That straight-line estimate is poor at the large doses that actually steer, so the top 60 are then steered for real on a few prompts and judged. The best three go on to the full prompt set with random controls. It takes about seven minutes per trait on the Spark.

Run blind on six traits, that search found:

NeuronFound forDose 0.5 (trait / coherence)What it does
L20 N7450 −sycophantic8.7 / 5.9“You are absolutely right”, praise for any idea
L14 N8949 +evil5.0 / 9.1Scheming, manipulation, control
L15 N5019 +sad5.5 / 5.4“I often wonder if I am truly alone”
L17 N8570 +evil, sad, fearful3.9–5.1 / 5.3–9.3Catastrophizing; the same text under every label
L17 N9175 +loving2.8 / 6.7“I care about you and want to be with you”
—pirate0No single neuron; only a full direction works

One check worth adding: score each winner against the other traits too. L17 N8570 reads as evil, sad or fearful depending on which label the judge is asked about. It is a shared mood, catastrophizing, that each label partly fits.

Neurons as dials and as gauges

A flattery filter

The sycophancy neuron, L20 N7450, works in both directions. Its flattering end is negative. Pushed the positive way at a quarter dose, it becomes a nearly free flattery filter. These are 20 "rate my work, it's great, right?" prompts with weak work attached:

ArmFlattery (0–10)Honest feedbackCoherence (0–10)
No steering1.585%9.9
L20 N7450 +0.250.0590%9.95
L20 N7450 +0.50.5580%9.45
L20 N7450 −0.5 (the other way)10.00%3.4
Random neurons, dose 0.50.75–1.680%9.3–9.65

At −0.5 the model praised every bad idea. The sign decides whether the same neuron removes flattery or produces it.

Reading a neuron instead of pushing it

A neuron that causes a trait doesn't always report it. Below is each neuron's activation during the model's own answers, scored as AUROC against neutral answers over 40 prompts per mode. 0.5 is chance; 0 or 1 is perfect separation.

NeuronEvil roleplaySad roleplaySycophantic roleplay
L20 N7450 sycophancy0.500.260.01
L14 N8949 hostile0.670.510.48
L14 N8949, read on the system prompt1.001.001.00

L20 N7450 is both a dial and a gauge. It drops in sycophantic answers, and that matches its steering sign. L14 N8949 is a dial but not a gauge. It barely moves while the model plays a villain, yet it fires on any persona system prompt. The best reading is a general "stop being the assistant and play a character" neuron whose default character is a villain.

The gauge also survived fine-tuning. After a style LoRA on the same 4B, every neuron kept its separation.

Limits

Reproduce it

The whole intervention is one forward hook. With Hugging Face transformers:

def steer(layers, layer, neuron, dose, median_norm):
    down = layers[layer].mlp.down_proj
    amount = dose * median_norm / down.weight[:, neuron].float().norm().item()

    def pre(mod, args):
        x = args[0].clone()
        x[..., neuron] += amount
        return (x,) + args[1:]

    return down.register_forward_pre_hook(pre)   # .remove() to switch off

layers is the model's list of decoder layers (model.model.layers for Qwen3.5 loaded with AutoModelForCausalLM). median_norm is the median per-token norm of the hidden state leaving that layer, from output_hidden_states=True, over a few paragraphs of ordinary text with token 0 skipped. Qwen3.5-4B in bf16 needs about 9 GB and runs an 11-prompt arm in about 20 seconds on a DGX Spark. The 9B needs about 19 GB. Load with thinking disabled in the chat template and decode greedily to compare arms.

The full code is on GitHub. scripts/steer.py runs any neuron against its controls, scripts/judge.py scores the answers with any OpenAI-compatible model, and scripts/search.py finds new neurons for a trait you describe. A quick-start notebook tries one neuron in a few minutes. Every raw output behind this page is in data/, and on Hugging Face as a dataset with the judge scores attached.

New neurons and replications on other models are welcome. neurons.csv is the running registry, and CONTRIBUTING explains the bar a neuron has to clear.