Goodfire Discovers Small Neuron Communities Driving LLM Behaviors

Interpretability firm Goodfire released research on "Neuron Communities," an unsupervised method for finding small groups of neurons that repeatedly cooperate to drive common behaviors in LLMs. Surprisingly few neurons—under ten in some cases—can support behaviors like arithmetic in models such as Llama-3.1-8B.

2026-11-18 ~ 2026-11-18 · 3 related posts