Few-shot prompting — loses out to → System-prompting

Before settling on system-prompting as CAA's prompting baseline, the paper tried few-shot prompting -- giving the model prior examples of the target behavior in its context window, the technique Brown et al. (2020) introduced for GPT-3 -- and found it less effective at steering the seven tested behaviors than a single instructional system prompt (contrastive-activation-addition, §"5 CAA and system-prompting", p. 5). That is a comparative claim about steering specifically, not a general verdict on few-shot prompting: the paper's own Limitations section concedes it did not exhaustively search prompting configurations for either baseline, and that better-optimized prompts, few-shot or otherwise, might close some of the gap it reports between CAA and prompting (contrastive-activation-addition, §"Prompting baseline optimization", p. 10). What the result fixes is methodological: every later comparison to "prompting" in this paper, including Table 3 and the system prompts catalogued in Table 14, means system-prompting, because that is the baseline few-shot prompting lost to in preliminary trials.