Source
Brown et al. (2020), LIT-035 — ARXIV-2005.14165.
What it replaced
Before LIT-035 the standard route to a task was fine-tuning: a labelled dataset per task, a gradient update, a separate set of weights to serve. This practice is the observation that a large enough pretrained model reaches useful performance on many tasks from a prompt alone — the task specified in the input, no weights changed.
The consequence is operational rather than statistical, and it is why the practice matters more than its accuracy numbers: one model serves every task, adaptation costs a prompt rather than a training run, and a new task is available immediately rather than after a labelling effort.
Where it stops
Few-shot prompting is generally worse than fine-tuning on any single task with enough labelled data — LIT-035 says so, and the gap is largest where the task is unlike anything in pretraining. It is a trade of peak accuracy for breadth and immediacy.
The record’s own line has moved past the dichotomy: parameter-efficient tuning (SOTA-033’s territory) recovers most of the fine-tuning gain at a
fraction of the cost, and post-training with verifiable rewards (SOTA-130) produces capabilities prompting does not reach at all. Few-shot prompting is now the baseline you compare those against rather than the recommendation.
Kept Active because the claim is true and load-bearing: it is why the field stopped building one model per task, which is an assumption almost everything else in this record rests on.