Training Large Language Models to Reason Efficiently

24 Jan 2025 (modified: 18 Jun 2025)Submitted to ICML 2025EveryoneRevisionsBibTeXCC BY 4.0
TL;DR: Training Large Language Models to Reason Efficiently
Abstract: Training Large language models to perform advanced reasoning has endowed them with new capabilities to solve complex problems. These models spend additional compute at test time to tackle problems that require careful reasoning. However, it is equally important that these model reason efficiently, i.e., that they solve the task at hand by utilizing the least possible amount of compute at test time. In this paper we introduce a new technique to improve the reasoning efficiency of reasoning models, one that leads to a substantial saving in terms of compute to reach a target level of accuracy.
Primary Area: Deep Learning->Large Language Models
Keywords: Reasoning models, efficiency
Submission Number: 14237
Loading