Stable LLM Ensemble: Interaction between Example Representativeness and Diversity

In particular, there is a rapid expansion of LLM applications into diverse domains, including marketing [1, 2], finance [3, 4], and education [5]. These models have enabled novel approaches to various tasks such as sentiment analysis, machine translation, retrieval-augmented generation (RAG), and text summarization. As a result, LLMs are fundamentally transforming the way data-driven insights are generated and utilized in each field. Despite their impressive abilities, the output of LLMs remains highly sensitive to model configuration, such as one-/few-shot learning [6], prompt templates [7], and hy-perparameters [8, 9]. The performance and consistency of LLM-generated predictions often depend on subtle differences in example selection and prompt construction. However, there is currently no established method for optimal example selection, and many existing studies still rely on random sampling strategies. In addition, ensemble learning [10] using LLMs has been actively conducted [11, 12, 13]. Ensemble methods utilize the inherent randomness in machine learning including LLMs to achieve higher accuracy [13] and computational efficiency [12], compared to single inference. However, since the model configuration greatly affects the individual inference of LLMs, it is also important in ensemble.