上下文學習優化方法之比較：靜態與動態策略分析

Abstract

在眾多自動化提示詞優化的方法中，指令的調整與範例的使用已經展現出強大的效果。本研究將優化範例的方法分為兩大類：靜態方法，即對所有測試樣本都使用一組固定的範例；以及動態方法，會根據測試樣本調整所使用的範例。儘管這兩種方法都在解決相同的問題，相關的系統性比較研究仍相對有限。因此，本研究針對具有代表性的靜態與動態方法，在多個資料集和實務情境中進行比較，以展示兩種方法間的差異。結果顯示動態方法在多數情境中表現優於靜態方法，但不同策略的效果仍會依據資料特性而異。

Among various automatic prompt optimization approaches, instruction tuning and the use of exemplars have shown strong effectiveness. We categorize exemplar optimization methods into two paradigms: static, which uses a fixed set of samples across all test samples, and dynamic, which adapts exemplars to the input. Despite addressing the same challenge, little attention has been paid to a systematic comparison. In this paper, we conduct an empirical study that compares representative static and dynamic methods across diverse benchmark datasets and real-world scenarios to demonstrate the difference in these two directions. Our results show that dynamic methods often outperform static ones, though effectiveness in different strategies varies depending on data characteristics.