步步精心：用於少樣本情境學習的階段性學習範例選擇策略

Abstract

大型語言模型近年來已展示出其驚人的能力，特別是在不需要任何參數更新的情況下，能使用少量的範例進行學習來完成各種任務。然而，學習範例的選擇大大影響了模型的表現，進而影響了穩定性。受到相關研究利用大型語言模型的推理能力和選擇多元化範例提升表現的啟發，我們提出了新的學習範例選擇策略，名為步步精心（Let's Select Step by Step）。我們利用大型語言模型進行客製於下游任務的資料分群並生成解釋，接著透過高效率的篩選機制選擇最佳的學習範例。實驗結果表明，我們的方法能在大型語言模型擅長的任務上更進一步地提高了模型表現和穩定性，而此方法的成效更是隨著利用進階模型而增強，在需要進階能力的任務表現上也有所提升。

Large Language Models (LLMs) have shown a remarkable ability to perform various tasks using few-shot in-context learning from a limited number of demonstration examples, without requiring parameter updates. However, the performance of such learning is notably inconsistent across different example sets. Inspired by related work highlighting the benefits of utilizing the reasoning ability of LLMs and diverse examples, we propose a method, Let's Select Step by Step (LSSS), for demonstration example selection. By leveraging LLMs, we carry out task-specific clustering and explanation generation, followed by an efficient evaluation to select better demonstration examples. Experimental results indicate that our method enhances both performance and stability on tasks where LLMs typically excel. Moreover, for tasks demanding specific linguistic capabilities, employing more advanced LLMs could further boost the effectiveness of our approach.