Refusal Is Redundantly Distributed, Not Localized: A Per-Layer Ablation Study on Llama-3.1-8B

LessWrong reports on mechanistic interpretability research into Llama-3.1-8B-Instruct's safety mechanisms. Researchers found that the model's refusal behavior—its ability to decline harmful requests—is redundantly distributed across multiple layers rather than localized in one place. While layer 12 emerged as most influential, removing every layer except twelve produced identical results to removing all thirty-two, suggesting no single layer is necessary. The work indicates that safety training in this model creates a low-dimensional, linear feature accessible from the surface, pointing to what the researchers call a surface-level fix rather than a deep change in the model's core capabilities.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton