Charlie Steiner comments on Investigating Bias Representations in LLMs via Activation Steering

[Charlie Steiner]

What’s the similarity between the refusal vector and various bias vectors? Are you literally “adding bias” when you reduce refusal?