ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning

TL;DR

ConCISE is a framework that enhances large reasoning models by reducing verbose outputs through confidence-guided techniques, leading to more efficient reasoning with minimal accuracy loss.

Contribution

This work introduces ConCISE, a novel confidence-guided compression framework that effectively shortens reasoning chains while preserving performance in large reasoning models.

Findings

Reduces reasoning chain length by up to 50%

Maintains high task accuracy after compression

Improves efficiency of large reasoning models

Abstract

Large Reasoning Models (LRMs) perform strongly in complex reasoning tasks via Chain-of-Thought (CoT) prompting, but often suffer from verbose outputs, increasing computational overhead. Existing fine-tuning-based compression methods either operate post-hoc pruning, risking disruption to reasoning coherence, or rely on sampling-based selection, which fails to remove redundant content thoroughly. To address these limitations, this work begins by framing two key patterns of redundant reflection in LRMs--Confidence Deficit, wherein the model reflects on correct intermediate steps, and Termination Delay, where reflection continues after a verified, confident answer--through a confidence-guided perspective. Based on this, we introduce ConCISE (Confidence-guided Compression In Step-by-step Efficient Reasoning), a framework designed to generate concise reasoning chains, integrating Confidence Injection to boost reasoning confidence, and Early Stopping to terminate reasoning when confidence is sufficient. Extensive experiments demonstrate that compared to baseline methods, fine-tuning LRMs on ConCISE-generated data yields a better balance between compression and task performance, reducing length by up to approximately 50% under SimPO, while maintaining high task accuracy.