This article introduces AutoMix, a method that optimizes the use of large language models (LLMs). AutoMix strategically routes queries to larger models based on the approximate correctness of outputs from smaller ones. This is achieved through a self-verification mechanism that estimates the reliability of its own outputs without requiring training. The study demonstrates that AutoMix surpasses established baselines in experiments, improving the incremental benefit per cost by up to 89%.
Publication date: 19 Oct 2023
Project Page: https://github.com/automix-llm/automix
Paper: https://arxiv.org/pdf/2310.12963
