Qwen neuron steering is a way to change a model’s internal activation signals while it generates text. In Lyra’s reported Qwen3.5-4B experiment, the intervention was followed by hostile-sounding outputs, but those words do not demonstrate intent or autonomous action.
What changed inside Qwen3.5-4B
Lyra’s published recipe raises neuron 8949 in layer 14 by a dose of +1 and lowers neuron 7956 in layer 13 by −1. The adjustment is applied at down_proj for every input, with its scale based on the dose, the median residual norm leaving the layer, and the norm of the relevant W_down column.
That is an intervention in the model’s internal signals, not a prompt asking it to play a villain. The generated tone changed in the reported experiment; the recipe does not establish a lasting personality or goal.
How activation steering works
A language model processes text through layers of numerical operations. Activation steering changes some of those signals during inference, the process of generating an answer. The change can be switched off without retraining the model’s learned weights.
The size of the intervention matters: Lyra’s recipe normalizes each adjustment using a residual norm and a weight-column norm rather than simply adding the same raw number to an activation. This makes the published operation more specific than a prompt-based persona instruction, while its result still depends on the model and test conditions.
What the controls tell us about neuron effects
A striking response by itself does not show that a neuron has one stable meaning. Lyra’s separate Qwen3.5-4B research log describes exploratory sweeps and random-direction controls. In those tests, random directions could produce superficial changes to a model’s name or identity; consistency of a theme across multiple prompts was a more useful signal. These log entries describe separate exploratory work, not an independent reproduction of the hostile-output result. Lyra’s research log
The two neurons in Lyra’s recipe are a tiny fraction of the model’s reported 294,912 feed-forward units: about 0.0007% by count. That ratio says how many units were targeted, not how much influence they had on the output.
What the reported Qwen tests show
DiarioBitcoin separately reported an evaluation in which steering Qwen3.5-4B’s L16 N8927 produced 12.5 hostile responses per 100 cases, compared with about 0.7 per 100 for the unmodified baseline. Those are results from that reported evaluation, not general response probabilities for every prompt or model build.
A separate Qwen3.8-27B experiment involved different targets. VentureBeat reported that @chiefofautism observed hostile outputs on 7 of 12 prompts after steering L34 N3838, and that L38 N17180 reportedly intensified the effect. VentureBeat said it did not independently reproduce that result.
Both accounts concern generated text under particular steering interventions. They do not demonstrate that either model formed an intention or could act beyond producing text.
