Verbosity-Aware Rationale Reduction: Effective Reduction of Redundant Rationale via Principled Criteria

TL;DR

This paper introduces a verbosity-aware rationale reduction method that selectively removes redundant reasoning sentences in large language models, improving efficiency and performance across reasoning tasks.

Contribution

It presents a novel sentence-level rationale reduction framework using likelihood-based verbosity criteria, outperforming previous token-level reduction methods.

Findings

Improves reasoning task performance by 7.71% on average.

Reduces token generation by 19.87%.

Enhances reasoning efficiency without sacrificing accuracy.

Abstract

Large Language Models (LLMs) rely on generating extensive intermediate reasoning units (e.g., tokens, sentences) to enhance final answer quality across a wide range of complex tasks. While this approach has proven effective, it inevitably increases substantial inference costs. Previous methods adopting token-level reduction without clear criteria result in poor performance compared to models trained with complete rationale. To address this challenge, we propose a novel sentence-level rationale reduction framework leveraging likelihood-based criteria, verbosity, to identify and remove redundant reasoning sentences. Unlike previous approaches, our method leverages verbosity to selectively remove redundant reasoning sentences while preserving reasoning capabilities. Our experimental results across various reasoning tasks demonstrate that our method improves performance by an average of 7.71% while reducing token generation by 19.87% compared to model trained with complete reasoning paths.