Emotion neurons causally steer how LLMs read ambiguous news
Sauers_ · x · 2026-10-26
An interpretability experiment shows that toggling single "emotion neurons" systematically shifts an LLM's interpretation of ambiguous news. Turning up an "afraid" neuron (Layer 23, n11919) pushed toward the unfavorable outcome by +2.45 nats; "relieved" (Layer 22, n60) toward favorable by +3.42; "humiliated" +1.77 unfavorable. Hope-related neurons had opposing effects (-0.54 vs +0.78), while 20 matched random neurons showed no systematic shift — evidence that internal emotion representations causally bias how models judge uncertain situations.
More from Research
- UniForge3D turns sparse-view inputs into detailed 3D meshes, code released — mervenoyann · 2026-10-26
- neuron-explorer: open-source steering-first neuron explorer for Qwen3.5-4B — viksit · 2026-10-26
- 30% of US STEM PhD dissertations now contain AI-generated text, NBER paper finds — 0xkarasy · 2026-10-26
- None of the 11 grand challenges in single-cell science solved since 2020 — Sauers_ · 2026-10-26
- MemTensor's MemT-1.0-Nano trends on Hugging Face, a memory-tuned Qwen3.6-35B-A3B finetune — MemTensor · 2026-10-26
- Mathematicians dissect OpenAI's fluid singularity paper: unprecedented smooth-force construction — Singularitarian · 2026-10-26