Boss LLM: Adaptation via No-Regret Learning

The diversity of Large Language Models (LLMs) calls for more effective strategies to combine their strengths across various tasks. In this work, we learn an adaptive mixture of multiple expert models as Boss LLM. By extending the multi-objective optimization with exponential weights (MOEW) algorithm, Boss LLM selects the most suitable model for a given prompt that could potentially span multiple categories with provable low regret for every category and expert model. Empirical results demonstrate that Boss LLM not only effectively adapts its mixture based upon the categories of a given prompt and improves upon the expert models, but also exhibits generalization properties.