What is it about?
Beth bakes 4 trays with two dozen batches of cookies in a week. If these cookies are shared amongst 16 people equally, how many cookies does each person consume? (Answer: 4*2*12/16=6) You would think that Large Language Models would have no problem in calculating this answer. Surprisingly, this is a difficult task for traditional models. Humans typically solve this problem by breaking it in smaller subproblems. They apply reasoning. Recently much research has appeared on multi-step reasoning methods for LLMs. We classify close to 200 publications in how they (1) generate steps, (2) evaluate them, and (3) control the reasoning process. We find that multi-step reasoning approaches have progressed beyond math word problems, and can now successfully solve challenges in logic, combinatorial games, and robotics. Some of the most successful approaches first generate a computer program that is then executed. Other approaches use further methods that are inspired by human learning: reinforcement learning and self-reflection.
Featured Image
Photo by ALAN DE LA CRUZ on Unsplash
Why is it important?
Reasoning methods are important for LLMs to achieve high performance on solving complicated problems. All modern large language models use some form of the methods described in the paper.
Perspectives
As an AI researcher, it is amazing to see how mechanical principles are able to achieve reasoning and learning that is so much like human intelligence. In Kahnneman's terms, System 1 (thinking fast) and System 2 (thinking slow) come together in reasoning models.
Aske Plaat
Universiteit Leiden
Read the Original
This page is a summary of: Multi-step Reasoning with Large Language Models, A Survey, ACM Computing Surveys, November 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3774896.
You can read the full text:
Contributors
The following have contributed to this page
