核心概念
LLMs can enhance induction through deduction.
摘要
The paper introduces the ItD framework to improve the inductive capability of Large Language Models (LLMs) through deduction. It consists of two main components: Deductive Data Generation and Naive Bayesian Induction. The framework is tested on two types of induction tasks: Instruction Induction and List Function, showcasing significant performance improvements compared to existing methods. ItD effectively leverages the deductive capability of LLMs to enhance their inductive abilities.
统计
ItD achieves a relative performance improvement of 193% and 16% compared to the base model (IO) on Instruction Induction and List Function, respectively.
ItD-IO outperforms the base model (IO) by 146% on both datasets, demonstrating the effectiveness of Deductive Data Generation.
ItD outperforms ItD-IO on both datasets, indicating the effectiveness of Naive Bayesian Induction.
引用
"Large language models are not abstract reasoners." - Gendron et al., 2023
"LLMs are much better at deduction than induction." - Bang et al., 2023