[2Win5] Poster session 2

[2Win5-79] An Examination of Sample Selection for Few-shot Learning in Multiclass Classification Using LLM

〇Takuya Fukumoto¹, Yuutarou Shiramizu¹, Hiroshi Fujimoto¹, Takeshi Yoshimura¹ (1.NTT DOCOMO, INC.)

Keywords:Large Language Model, Classification

This study propopses an approach to improving the efficiency of multi-class Few-shot classification using Large Language Models (LLMs). By leveraging Zero-shot classification results, we allocate additional samples to low-accuracy categories while reducing samples for high-accuracy categories to optimize resource distribution.Experimental results demonstrated that our method enhanced classification accuracy while maintaining computational efficiency. It achieved performance comparable to 2-shot classification with fewer input samples. Additionally, our analysis indicated that shorter token samples maintained high accuracy, suggesting a cost-effective classification strategy.Future work focuses on evaluating this approach on datasets with minimal category-wise performance gaps, extending it to multi-label classification and context-dependent tasks, and automating sample selection for broader LLM applications. This study provides practical insights into optimizing Few-shot classification while balancing computational cost and classification accuracy.