Constrained Mixed-Strategy GroupDRO for Equitable System-Prompt Selection

A novel approach known as Constrained Mixed-Strategy GroupDRO has been introduced to enhance the fairness of system-prompt selection in large language models (LLMs). This method, outlined in an arXiv paper (2608.04339), tackles the problem where differently phrased but semantically similar questions yield answers of inconsistent quality. Typically, system prompts are fine-tuned for average performance, resulting in some variations receiving inadequate or subpar responses. The new framework allocates weights to existing system prompts to reduce the worst-case information-quality loss across various evaluation metrics and groups, while keeping the average loss comparable to that of standard selection. By decoupling pool generation and selection, this technique can utilize any system-prompt pool and an ensemble of complementary prompts, aiming for more uniform answer quality across different user queries, which is essential for the effective application of LLMs in information retrieval.