DPS: Diverse Prototype Selection for Adaptive In-Context Learning

Abstract

Large language models exhibit remarkable proficiency across a wide array of tasks by leveraging in-context learning, wherein they learn from a limited number of examples. However, the efficacy of ICL is highly sensitive to the choice of demonstrations provided. Existing approaches primarily focus on the selection of individual examples, often neglecting the broader context of the entire example bank. In this paper, we introduce a novel framework aimed at augmenting the example bank through D iverse P rototype S election ( DPS ). DPS decomposes the ICL process into two distinct stages: Prototype Selection and Prompt Synthesis . In the first stage, DPS identifies a set of prototype functions that closely approximate the underlying data distribution. In the second stage, these prototype functions dynamically generate query-specific demonstrations, thus guiding the LLM more effectively in its task. Empirical evaluations conducted across thirteen reasoning benchmarks demonstrate that DPS significantly enhances ICL performance, providing substantial improvements when integrated with downstream LLMs.