Yandere neurons found in a 27B LLM: 26 steered neurons, no fine-tuning
Promptmethus · x · 2026-10-28
Developer balaena01 used logit lens on Qwen3.8-27B (GGUF) to find neurons that write jealous/possessive words, then combined 26 of them with per-neuron dose tuning (some up, some down) to give the model a yandere personality — no fine-tuning. Steering applies only at when: assistant, so unsteered reads lose less capability, and the same recipe works on the Saluki 27B fine-tune. The accompanying tool, rebuilt on llama.cpp from @lyraaaa's neuron-explorer, can list what words each neuron promotes across all layers, rank candidates by objectives like villain-vs-normal, compare multi-neuron recipes in three-column steering tests, and export recipes as llama.cpp control vectors usable in llama-cli/llama-server or baked into GGUF.
More from Research
- IBM's STAIR retriever cuts hallucination 65x vs fine-tuned RAG, hits 82.6% Recall@1 — Promptmethus · 2026-10-28
- Researcher: frontier LLMs convincingly refuted his 30 years of P=NP attempts — csuwildcat · 2026-10-28
- Google researcher studies whether image models like Nano Banana can look things up — YonatanBitton · 2026-10-28
- Long read: AGI is arguably already here, but mind uploading is far harder than scaling — UAP44 · 2026-10-28
- Why softmax on QKᵀ before multiplying by V? The transformer interview question that reveals attention — cneuralnetwork · 2026-10-28
- Berkeley paper: agents that forget their history beat parallel agents by up to 15% — ombre-ai · 2026-10-28