Zeroing 16 neurons cuts a specific model response probability by 830,000x

一项模型可解释性演示引发关注：研究者 Sauers 定位出负责生成特定回复开头（

2026-10-26 ~ 2026-10-26 · 2 related posts

• Zeroing 16 neurons makes a typical LLM reply 830,000× less likely — Sauers_ · 2026-10-26
• Deleting a Few Neurons Can Nerf a Model Behavior, Interpretability Demo Shows — pbaylies · 2026-10-26