Chat Model Refusal Behavior Linked to Persona, Not Isolated Mechanism

New research explores the interaction between refusal and persona traits in large language models, specifically Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct. Previous studies treated these as separate mechanisms, but this work demonstrates a direct relationship: a model's compliant persona actively suppresses its refusal to answer certain prompts. By intervening on both persona and refusal directions in the model's activation space, researchers found that a compliant persona can drastically reduce refusal rates, in some cases from 97% to 2%. While reintroducing a refusal direction can partially restore this behavior in later layers, it has less impact in earlier ones. This suggests that refusal is not an independent computation but rather an expression gated at later stages, dependent on the model's overall persona.