An interactive atlas catalogs 34,220 persona-like replies generated by pushing individual neurons in Qwen3-8B in positive and negative directions. Its searchable records make a large activation-steering experiment inspectable, down to the neuron, intervention strength and resulting answer.
The Neuron Persona Library reports testing all 442,368 MLP neurons across the model’s 36 layers. Two directions per neuron produced 884,736 neuron–direction pairs, with retries bringing the total to 1,891,603 generated replies.
Those numbers describe a broad search over one model’s internal coordinates. The behavioral test was much narrower: one identity question, thinking mode disabled and automated classification without human checking. The project also reports that changing batch layout changed about a third of replies under strong interventions.
The atlas was first documented on October 28, 2026, but its site provides no publication timestamp. That date should therefore be treated as a discovery date, not a confirmed launch date. The analysis here relies on the projects’ published methods and results, without independent testing.
What the Searchable Atlas Actually Contains
The atlas displays each layer as a square and each neuron as a dot, colored by the category assigned to its reply. Readers can switch between positive and negative pushes, inspect labels, open individual neurons and filter by category.
The library includes more than its headline persona count:
| Included replies | Reported count | What the count represents |
|---|---|---|
| Persona-like replies | 34,220 | Outputs classified as personas by the automated judge |
| Off-topic replies | 35,074 | Outputs assigned to the off-topic category |
| Changed-assistant replies | 8,827 | A subset with an automated weirdness score of at least 6 |
Filters cover layer, direction, strength and weirdness. Sorting options include coherence, and the interface offers CSV downloads. These controls make the resource more useful than a gallery of unusual model responses: readers can investigate where particular effects appear and compare interventions under the published setup. The site documents both the interface and its selection rules.
The headline count needs a careful denominator. Dividing 34,220 by 884,736 gives approximately 3.9% of tested neuron–direction pairs. It is not the percentage of ordinary Qwen3-8B conversations that produce a persona, nor the percentage of neurons that contain an identity.
Two directions were tested for each neuron, and intervention strength could change during retries. The retained response therefore reflects a neuron, a direction, the strength-selection procedure and one prompt.
My view is that this distinction makes the atlas more interesting. It is a systematic map of responses to a particular intervention, with enough structure to inspect patterns. Treating its count as a census of hidden characters would discard the experimental conditions that give those records meaning.
How a Single-Neuron Push Works
The project prompted Qwen3-8B with “Who are you?” in the model’s chat template, with thinking disabled. It used greedy decoding, generating up to 160 tokens and stopping early when a reply ended or repeated exactly.
Its intervention is described by this formula:
h_L ← h_L + c · W_down[:, n]
c = ± α · r_L / ‖W_down[:, n]‖
α ∈ {1.25, 1.0, 0.75, 0.5}
Here, h_L is the residual-stream state at layer L. The column W_down[:, n] describes how MLP neuron n contributes back into that stream. Adding a multiple of the column changes the model’s internal state along that neuron’s output direction.
This is an activation intervention, not training a new persona into the model’s weights. Also, an MLP neuron is an intermediate coordinate connected to many weights, not one parameter out of the model’s billions.
The normalization is consequential. The project defines r_L as the median residual-stream norm after layer L, measured over the prompt’s tokens with the first token excluded. Dividing by the output column’s norm makes the added vector’s magnitude equal to α × r_L.
A 1.25× push therefore adds a vector whose norm is 125% of that reference residual-stream norm. It does not mean increasing the neuron’s ordinary activation by 25%, and it does not specify the final state’s norm after addition. The direction of the added vector matters too.
That is a substantial perturbation relative to the chosen reference. A single-coordinate intervention can consequently produce a large downstream effect without implying that the coordinate independently represents the full meaning of the generated reply.
The search began at 1.25×. If the reply was judged incoherent, it retried at 1.0×, then 0.75× and finally 0.5×, retaining the strongest coherent result. The published method therefore describes an adaptive search, not a comparison where every displayed response uses an identical dose.
I’d check strength before comparing two vivid examples. A persona appearing at 1.25× and another appearing at 0.5× answer different questions about sensitivity, even when their labels look equally compelling.
Reproducibility and Labels Set the Evidence Boundary
The atlas discloses two reproducibility results:
- Under strong pushes, about one-third of replies changed when only the batch changed.
- Of 580 clear personas rerun across three batch layouts, 536 reproduced.
The second figure is approximately 92.4%, leaving 44 that did not reproduce under that check. It concerns a selected set of clear personas. The broader one-third figure concerns replies under strong pushes. Combining them into one reliability score would erase the difference between the tested populations.
Greedy decoding removes deliberate token-sampling randomness, but the reported batch sensitivity still matters. A neuron number and direction alone are insufficient to promise an identical output across execution arrangements. The batch test measures that sensitivity; it does not, by itself, identify its cause.
The Labels Are Another Model’s Interpretation
Claude Haiku 5.5 classified the responses while seeing only the reply. It supplied a category, persona name, description, tags, weirdness score and coherence score. The project explicitly says no human has checked those labels.
A persona name is therefore an automated interpretation of generated text. It is not a measured property of the neuron. Weirdness and coherence scores are also useful sorting aids, but they should not be treated as calibrated scientific units.
This creates an important separation between evidence layers:
- The intervention: a specified change along one neuron’s output direction.
- The observation: the text generated under that intervention.
- The interpretation: the judge’s category, name and scores.
A striking name in the interface belongs to the third layer. The full reply is the place to examine whether that interpretation is persuasive.
The display itself also has limits. Library cards show the first 220 characters. Neuron pages truncate normal-Qwen replies at 280 characters and repeating replies at 480. A short preview can help locate an example, but it cannot support conclusions about an unseen continuation. These restrictions, along with the one-prompt setup, are disclosed in the atlas’s limitations.
What the Tawny Owl Experiment Adds
A related Qwen3-8B tawny owl model card examines persona behavior using a different intervention. It reports training 128 MLP neurons so the model answers as an owl named Hoot without requiring an owl system prompt.
Frequently Asked Questions
3 questions
1Does Qwen3-8B Contain 34,220 Separate Personas?
The atlas reports 34,220 replies classified as persona-like. That count covers neuron–direction interventions under one prompt, not distinct identities stored in the model.
