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

TL;DR

ConCISE introduces a confidence-guided framework to compress step-by-step reasoning in Large Reasoning Models. By formalizing reflection via per-step confidence and two redundancy patterns—Confidence Deficit and Termination Delay—it deploys Confidence Injection and Early Stopping to suppress unnecessary reflections. Fine-tuning with ConCISE-generated data under SFT and SimPO yields substantial reductions in response length (up to ~50%) while maintaining high task accuracy across multiple benchmarks. The work highlights the importance of training-data design for compression and discusses limitations and future directions, including confidence estimation and RLVR integration.

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.

Figures (12)

• Figure 1: Training dataset construction workflows: ConCISE (our proposed method) vs. existing methods.
• Figure 2: ConCISE achieves a better trade-off between compression and task performance than baselines.
• Figure 3: Illustration of ConCISE's confidence-guided approach: identifying patterns (Confidence Deficit, Termination Delay) and applying mechanisms (Confidence Injection, Early Stopping) to suppress redundant reflections, shown in contrast to the original reasoning process. $C_i$ denotes step confidence and $T_i$ its threshold.
• Figure 4: Effectiveness and necessity of Confidence Injection and Termination Delay(details in Appendix \ref{['cha:datasets_details']}).
• Figure 5: Further analysis of reasoning chains and training datasets on DeepSeek-R1-Distill-Qwen-7B.