From Long to Short: LLMs Excel at Trimming Own Reasoning Chains

Wei Han, Geng Zhan, Sicheng Yu, Chenyu Wang, Bryan Hooi

Keywords: efficient reasoning, large reasoning models, test-time scaling

TL;DR: An efficient method to prune reasoning chains of large reasoning models

Abstract:

O1/R1-style large reasoning models (LRMs) demonstrate strong performance across a wide range of complex reasoning tasks, particularly by leveraging test-time scaling to generate extended reasoning paths. However, these models often suffer from overthinking.
To address this issue, we conduct a systematic investigation into the reasoning efficiency of a broad set of LRMs and reveal a common dilemma. Motivated by the key findings, we propose a purely test-time computation method, EDIT, that guides LRMs to identify the shortest correct reasoning paths at test time.
EDIT employs constraint-guided generation while jointly tracking length and answer distributions under varying constraints, allowing it to select responses that strike an optimal balance between conciseness and correctness. Extensive experiments across diverse models and datasets show that EDIT substantially enhances reasoning efficiency, producing compact yet informative outputs that improve readability and user experience.

Submission Number: 76