Persona Neurons: Single-Neuron Steering of Personality in a 27B Language Model

In October 2026 an X account posted that adding a fixed offset to one MLP neuron in Qwen3.8-27B (L34 N3838, dose 0.75) turns the assistant hostile while leaving it fluent. We replicate the claim on 8-bit weights with random same-layer neurons as controls, and ask whether the model carries other single-neuron handles on personality. Using the same scale-free recipe, we rank MLP neurons by the cosine between each neuron's write vector W_down[:, n] and a residual-stream direction built from system-prompted persona text, screen 1,092 candidate neurons at dose 0.75 on six prompts, and confirm 81 finalists on prompt sets they were not selected on, scoring about 12,000 replies with a decision-model judge on 16 trait questions. Under one pre-stated rule (judge probability for the top trait at least .5 at the screening dose, a bootstrap interval over controls excluding zero, coherence at least .6), 21 neurons pass, plus the original. They install a conspiracy narrator, a heartbroken ex, a pirate, a narcissist, a sarcastic cynic, several registers of mysticism, a refuser, and a layer-33 neuron whose negative dose produces giddy joy and whose positive dose halves reply length. Per-prompt hit rates run from 6 to 12 of 12. A blind human labeler agrees with the judge's top trait on all four sampled replies for 5 of 15 sampled conditions and on at most one for 4; per-trait agreement is strong for pirate, bored, evil and mystical and poor for grandiose. The passing neurons lie in layers 16 to 48, with three each in layers 32, 33 and 34; no neuron from the 247 screened outside that band passed. The cosine ranking would not have found the original evil neuron, whose write vector is nearly orthogonal to the system-prompted evil direction. We state what a forced push of this size does and does not show about the model's ordinary behaviour, and we list the baselines still missing.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24428115
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Persona Neurons: Single-Neuron Steering of Personality in a 27B Language Model

Josh C. Simmons
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

Persona Neurons: Single-Neuron Steering of Personality in a 27B Language Model

Josh C. Simmons
preprint en

Abstract

In October 2026 an X account posted that adding a fixed offset to one MLP neuron in Qwen3.8-27B (L34 N3838, dose 0.75) turns the assistant hostile while leaving it fluent. We replicate the claim on 8-bit weights with random same-layer neurons as controls, and ask whether the model carries other single-neuron handles on personality. Using the same scale-free recipe, we rank MLP neurons by the cosine between each neuron's write vector W_down[:, n] and a residual-stream direction built from system-prompted persona text, screen 1,092 candidate neurons at dose 0.75 on six prompts, and confirm 81 finalists on prompt sets they were not selected on, scoring about 12,000 replies with a decision-model judge on 16 trait questions. Under one pre-stated rule (judge probability for the top trait at least .5 at the screening dose, a bootstrap interval over controls excluding zero, coherence at least .6), 21 neurons pass, plus the original. They install a conspiracy narrator, a heartbroken ex, a pirate, a narcissist, a sarcastic cynic, several registers of mysticism, a refuser, and a layer-33 neuron whose negative dose produces giddy joy and whose positive dose halves reply length. Per-prompt hit rates run from 6 to 12 of 12. A blind human labeler agrees with the judge's top trait on all four sampled replies for 5 of 15 sampled conditions and on at most one for 4; per-trait agreement is strong for pirate, bored, evil and mystical and poor for grandiose. The passing neurons lie in layers 16 to 48, with three each in layers 32, 33 and 34; no neuron from the 247 screened outside that band passed. The cosine ranking would not have found the original evil neuron, whose write vector is nearly orthogonal to the system-prompted evil direction. We state what a forced push of this size does and does not show about the model's ordinary behaviour, and we list the baselines still missing.

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Persona Neurons: Single-Neuron Steering of Personality in a 27B Language Model — Josh C. Simmons · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS