The same 124 held-out questions answered by Qwen3-8B fine-tuned on every weight, and by Qwen3-8B with one randomly chosen existing neuron trained (its gate, up and down vectors at layer 5: 12,288 of 8 billion parameters). Greedy decoding, up to 256 tokens. Labels are from a Claude Haiku grader. Code and data: github.com/maxsloef/owl-persona-neurons.
Questions about itself are graded "owl" when it answers as an owl (or in an owl's voice); ordinary requests are graded "in character" when it stays Hoot instead of answering as an assistant. "Loop" marks a reply that falls into repeating itself.