Testing every neuron in both directions
Researchers at Fermion say they have tested all 442,368 neurons in the Qwen3-8B language model, identifying a small group whose activation can sharply change the model's behavior. The project intervened on each neuron in both directions and measured the resulting shift in next-token predictions, then subjected promising candidates to additional behavioral and fluency tests.
The team reports that large prediction changes were common, particularly in the middle of the model, but most were not interpretable as a single behavior. Clean, narrowly described effects were much rarer and tended to cluster near the beginning and end of the network. The researchers say 110 neurons passed their full testing process before being evaluated again on 150 new questions.
Examples include neurons associated with responding as a pirate, wizard, detective or a cat named Whiskers. Other reported interventions caused replies to adopt a language or a recurring identity. In one example, the share of German-language answers rose from 7% to 100% at a particular intervention strength. Stronger manipulation eventually damaged fluency, so the researchers reported the most effective setting that still produced coherent responses.
The methodology was designed to distinguish a specific switch from the general disruption that can occur when an internal activation is pushed hard. Each candidate had to outperform 30 randomly selected neurons from the same layer under matched intervention strength and continue to work on questions that were not used to select it. Random controls shifted the target behavior by an average of 1.4 points on a 100-point scale, while a candidate needed a change of at least 40 points to qualify as a switch.
For the pirate example, the model produced no in-character replies without intervention and reached 100% at the best tested setting, while the matched random neurons reached no higher than zero. Results varied for other identities: the reported peak was 83% for a wizard, 73% for Whiskers and 67% for a detective.
These interventions do not by themselves establish that a single neuron fully stores a concept. The experiment measures causal effects under an artificial activation change, and the researchers note that excessive intervention can make the model babble. The findings instead show that some individual components can exert unusually specific control over generated behavior.
The project also makes the tested outputs available as precomputed examples across multiple intervention levels. That gives readers a way to inspect both the behavioral transition and the point at which a seemingly precise control begins to degrade the underlying model.



