Andrew Mack comments on Mechanistically Eliciting Latent Behaviors in Language Models

[Andrew Mack]

Yes, the learned vectors are always applied at every token (for all examples).