Abstraction-of-Thought Makes Language Models Better Reasoners

Abstract reasoning, the ability to reason from the abstract essence of aproblem, serves as a key to generalization in human reasoning. However,eliciting language models to perform reasoning with abstraction remainsunexplored. This paper seeks to bridge this gap by introducing a novelstructured reasoning format called Abstraction-of-Thought (AoT). The uniquenessof AoT lies in its explicit requirement for varying levels of abstractionwithin the reasoning process. This approach could elicit language models tofirst contemplate on the abstract level before incorporating concrete details,which is overlooked by the prevailing step-by-step Chain-of-Thought (CoT)method. To align models with the AoT format, we present AoT Collection, ageneric finetuning dataset consisting of 348k high-quality samples with AoTreasoning processes, collected via an automated and scalable pipeline. Wefinetune a wide range of language models with AoT Collection and conductextensive evaluations on 23 unseen tasks from the challenging benchmarkBig-Bench Hard. Experimental results indicate that models aligned to AoTreasoning format substantially outperform those aligned to CoT in manyreasoning tasks.