Localized Safety Subnetworks in Large Language Models: An Empirical Note on Llama-3-70B

This project investigates the mechanistic interpretability of safety-aligned behavior within large language models (LLMs). Contrary to the assumption that safety mechanisms operate solely as opaque external filters or diffuse network properties, our empirical analysis of the Llama-3-70B-Instruct model demonstrates that safety-related responses correlate with a localized, quantifiable structure at the internal neuron level. By analyzing multilayer perceptron (MLP) activations in late transformer layers (60–79) in response to explicitly harmful versus benign prompts, we identified a continuous, heavy-tailed distribution of neuron importance. The results reveal a compact, redundant subnetwork responsible for safety behavior, suggesting that alignment is an internal geometric property of the model’s latent space.