CAA combined with system-prompting comparison — moves behavior beyond → System-prompting

Table 3 puts CAA and behavioral system prompts on the same scale -- average token probability given to the behavior-matching multiple-choice answer -- in Llama 2 13B Chat at layer 13 (contrastive-activation-addition, §"5 CAA and system-prompting", p. 6). For Corrigibility, a positive system prompt alone raises the score from a no-prompt, no-steering baseline of 0.57 to 0.79; adding the +1 steering vector on top of that prompt pushes it further still, to 0.93 -- CAA finds room the prompt left unclaimed. The same asymmetry appears in the negative direction for Refusal: the negative prompt alone drops the score from 0.78 (no prompt) to 0.74, and subtracting the steering vector on top of that prompt pushes it down further, to 0.41. Sycophancy is the outlier -- scores barely move across any prompt/multiplier combination (0.55-0.67 across all nine cells), suggesting the sycophancy vector and behavioral system prompts are not cleanly separable levers for that behavior the way they are for the other six. The paper's own explanation for why steering usually adds room a prompt does not: the multiplier gives continuous, signed control, and averaging over hundreds of contrast pairs isolates the target variable more cleanly than a single instruction can.