In-Context Learning(ICL)

1. Introduction

Essence of In-Context Learning (ICL) is package the context and send it to the LLM. When LLM inference, it will generate answer according to the context.

e.g. the SWE-agent, it’s a ICL-type agent. It has a multi-turn mechanism for inferencing answer. In solving problem phase, if problem is unsolved in one turn, the agent will package the whole environment info and context as a new prompt. The prompt will be send in the next turn.

Defect: LLM’s gains disappear once that experience is removed from the context.