In-context learning

De Wiki BackProp
Aller à la navigation Aller à la recherche

[1] The in-context learning (ICL) ability is formally introduced by GPT-3 : assuming that the language model has been provided with a natural language instruction and/or several task demonstrations, it can generate the expected output for the test instances by completing the word sequence of input text, without requiring additional training or gradient update

Fichier:A comparative illustration of in-context learning (ICL) and chain-of-thought (CoT) prompting.jpg

Références

  • [1] A Survey of Large Language Models