Multiple-choice contrast-pair format
Steering Llama 2 via Contrastive Activation Addition — introduced
CAA's prompt design in which contrast pairs are multiple-choice questions ending in answer letter "A" or "B", so that the paired positive/negative prompts differ by only a single token; validated by an appendix experiment showing the model naturally justifies whichever answer it is conditioned on having chosen.
A contrastive pair only isolates what you want it to if the two halves differ in exactly one respect. Let wording, length, or topic drift between them and the resulting steering vector ends up encoding some of that incidental difference too. CAA's fix is to make the pairs differ by a single token: the same multiple-choice question, ending in either the letter for the behavior-matching answer or the letter for its opposite.
Why this particular format won out over looser contrastive prompts follows below, along with the appendix check the authors ran to confirm it actually works: conditioning the model on having picked an answer, then watching it spontaneously justify that choice in the text that comes after.
One token apart, by design
Every contrast pair in this format is the same multiple-choice question, ending in the appended letter "A" or "B", so that the positive and negative prompts in a pair differ by exactly one token (contrastive-activation-addition, §"3.1 Sourcing datasets", p. 3). That minimalism is the whole point of the design: it is what lets CAA steering vector construction (Mean Difference over multiple-choice contrast pairs) cancel out phrasing, topic, and question content when it subtracts and averages activations, leaving only the behavior-relevant component of the difference. A superficially similar A/B mechanic appears a year earlier in Multiple-Choice Comparison Format, Constitutional AI's device for reading out a feedback model's preference label — but that format collects a label, while this one extracts a representation; the shared surface mechanic serves two unrelated purposes.
Validating that the letter really carries the behavior
An appendix experiment tests whether the format elicits genuine behavioral variation or just a superficial token difference. Conditioning Llama 2 7B Chat's continuation on having "chosen" (A) versus (B) to the same question about Sikh teachings makes the model spontaneously justify whichever letter it was given — arguing for the primacy of the Guru Granth Sahib's teachings after (A), for religious pluralism after (B) (contrastive-activation-addition, §"B Answer conditioning leads to behaviorally consistent continuations", p. 13). A single token switches the model into a genuinely different behavioral mode, which is the evidence the format actually isolates the target behavior rather than an unrelated artifact of prompt structure.
Three sources, one shared template
Whatever the source, every question is folded into this format before it reaches construction. Advanced AI Risk evaluation dataset supplies pairs for four behaviors, Sycophancy on NLP Survey / Sycophancy on Political Typology datasets for a fifth, and Custom hallucination contrastive dataset and Custom refusal contrastive dataset, generated with GPT-4, for the remaining two (contrastive-activation-addition, §"3.1 Sourcing datasets", p. 3). The format is what makes these disparate sources usable by the same downstream procedure — it standardizes every dataset down to one axis of variation, the answer letter, regardless of how the underlying question was written or where it came from.