Few-shot prompting — shares its mechanism with → GPT-3 (prompted) baseline

InstructGPT's strongest non-finetuned baseline was exactly this mechanism: a hand-picked few-shot prefix, chosen from an internal 'prefix-finding competition' by taking whichever prefix scored highest against the trained reward model, prepended to raw GPT-3 to coax it into an instruction-following mode (instructgpt, §"3.5 Models", p. 9). That 2022 result treats few-shot prompting as a serious lever, worth optimizing hard, because it was the only way to get GPT-3 to look like it was following instructions at all. CAA's 2023 finding cuts against that framing without contradicting it: on Llama 2 Chat, a model already RLHF-tuned to follow instructions, few-shot prompting is the weaker of the two prompting baselines tested, losing out to a single system-prompt instruction for steering any of the seven behaviors (contrastive-activation-addition, §"5 CAA and system-prompting", p. 5). Read together, the two results locate few-shot prompting's real strength: it is a tool for eliciting a response format or mode -- instruction-following -- from a model that does not otherwise have one, not a fine-grained dial for a behavior a chat-tuned model can already exhibit.