Anders Cairns Woodruff

Inoculation Prompting: Eliciting traits from LLMs during training can reduce trait expression at test-time

Language model finetuning often results in learning undesirable traits in combination with desired ones. To address this, we propose inoculation prompting: modifying finetuning data by prepending a short system-prompt instruction that deliberately elicits the undesirable trait. At test time, we eval…