dai2023-rec2ftp-icl-covers-over-85-percent-of-ft
IN premise — summaries/2026/08/24/dai-2023-icl-gradient-descent-s4-experiments.md
Created 2026-08-24T17:10:54+00:00
The Rec2FTP metric shows that ICL covers over 85% of the correct predictions achieved by finetuning over zero-shot learning, on average across the six classification tasks.
Summary
In-context learning, which just means showing the model a few examples in the prompt without any training, recovers more than 85% of the accuracy gap that full fine-tuning closes over zero-shot prompting, averaged across six classification tasks. In practical terms, this means the expensive step of actually retraining the model buys only a small fraction of what a well-chosen set of in-prompt examples already delivers.