Graph-Based Approaches for Few-Shot Example Selection in In-Context Learning

This thesis proposes a fast, versatile, and theoretically grounded graph-based approach for demonstration retrieval in In-Context Learning (ICL). Following relevant findings from the literature, this method focuses on retrieving examples that are relevant to a given query and good representatives of diverse concepts encoded in a knowledge base. We perform experiments with multiple configurations of the method (which depends on a resolution and a radius parameter) on five different tasks, including single-label and multi-label classification and generation of legal moves in a board game. We compare the performance of our method with Zero-Shot, Random and KNN baselines. The approach in most cases outperforms the former two baselines, and falls slightly behind KNN-based baseline. While the primary focus of this work is on the theoretical framework, future research directions include refining parameter selection, exploring different community detection algorithms, and investigating the use of weighted graphs to further enhance the retrieval process.