GPT-3 (prompted) baseline — shows the limits of prompting alone for → Instruction following (as a fine-tuning objective)

The GPT-3-prompted baseline tries to reach instruction-following behavior with zero weight updates, purely by prepending a hand-crafted few-shot prefix to coax the base model into an instruction-following mode. Even the best available prefix was not chosen by intuition alone: authors held an internal 'prefix-finding competition,' each spending an hour hunting for prefixes, and picked whichever one scored highest against the same reward model used to anchor RLHF training (instructgpt, §"3.5 Models", p. 9, n. 6). Despite using that reward model as its own selection criterion, the prompted baseline still trails actual RLHF-trained instruction following substantially — 175B InstructGPT is preferred 71±4% of the time over the prompted GPT-3 baseline. The comparison shows that prompting can approximate instruction-following behavior but not substitute for training on preference data.