Efficient and Effective In-context Demonstration Selection with Coreset

The recent publication of 'Efficient and Effective In-context Demonstration Selection with Coreset' on arXiv presents a significant advancement in the field of in-context learning (ICL) for large visual language models (LVLMs). The paper highlights the NP-hard nature of demonstration selection, which is critical for the effectiveness of ICL. Traditional methods often lead to inefficiencies, prompting the authors to propose a novel framework known as Coreset-based Dual Retrieval (CoDR). This innovative approach employs a cluster-pruning method to create a diverse coreset that aligns better with queries while maintaining diversity. Additionally, the dual retrieval mechanism enhances the selection process, achieving global demonstration selection efficiently. Experimental results indicate that CoDR significantly outperforms existing strategies, marking a notable improvement in ICL performance. This development not only addresses a key challenge in AI but also sets the stage for more effec…