Core Concepts
AutoMix optimizes computational cost and performance by strategically routing queries to larger language models based on self-verification.
Abstract
Abstract:
Large language models (LLMs) available from cloud API providers.
AutoMix approach strategically routes queries to larger LMs based on self-verification.
Few-shot self-verification mechanism estimates reliability of outputs without training.
Introduction:
Human problem-solving involves multi-step process: generate, verify, refine solutions.
Current self-refinement paradigms use single model but tasks vary in complexity.
AutoMix:
Three steps: solution generation, self-verification, selective routing using SLM and LLM.
Data Extraction:
Our experiments using LLAMA2-
13/GPT-4, on five context-grounded reasoning datasets demonstrate that AutoMix surpasses established baselines, improving the incremental benefit per cost by up to 86%.
Stats
私たちの実験では、LLAMA2-
13 / GPT-4を使用して、5つのコンテキストに基づいた推論データセットで、AutoMixが確立されたベースラインを上回り、コストごとの増分利益を最大86%向上させることを示しました。
Quotes
"Large language models cannot self-correct reasoning yet." - Huang et al., 2023