Short Chains, Deep Thoughts: Balancing Reasoning Efficiency and Intra-Segment Capability via Split-Merge Optimization

TL;DR

CoSMo presents Consistency-Guided Split-Merge Optimization to explicitly decouple reasoning topology from intra-segment content in large reasoning models. By iteratively merging redundant segments and splitting coarse ones, CoSMo aligns the number of reasoning segments with the ground-truth complexity $k^*$ while leaving intra-segment reasoning unconstrained. A two-stage training pipeline (split-merge data curation followed by structure-aware reinforcement learning with a segment-level budget) yields improvements in accuracy and substantial reductions in structural redundancy and token usage across multiple benchmarks and backbones. The approach generalizes to both in-distribution and out-of-distribution tasks and demonstrates robustness across diverse evaluators, offering a scalable path toward green, interpretable, and capable reasoning in LRMs.

Abstract

While Large Reasoning Models (LRMs) have demonstrated impressive capabilities in solving complex tasks through the generation of long reasoning chains, this reliance on verbose generation results in significant latency and computational overhead. To address these challenges, we propose \textbf{CoSMo} (\textbf{Co}nsistency-Guided \textbf{S}plit-\textbf{M}erge \textbf{O}ptimization), a framework designed to eliminate structural redundancy rather than indiscriminately restricting token volume. Specifically, CoSMo utilizes a split-merge algorithm that dynamically refines reasoning chains by merging redundant segments and splitting logical gaps to ensure coherence. We then employ structure-aligned reinforcement learning with a novel segment-level budget to supervise the model in maintaining efficient reasoning structures throughout training. Extensive experiments across multiple benchmarks and backbones demonstrate that CoSMo achieves superior performance, improving accuracy by \textbf{3.3} points while reducing segment usage by \textbf{28.7\%} on average compared to reasoning efficiency baselines.

Figures (9)

• Figure 1: Impact of enforced reasoning segment counts on accuracy across MuSiQue subsets.
• Figure 2: The process of Merge and Split phases (see Appendix \ref{['appendix:case_study']} for the full case study).
• Figure 3: Visualization of the efficiency-performance trade-off in the ablation study. The scatter plot illustrates the accuracy of different methods against their token consumption, where the size of each bubble is positively correlated with the number of reasoning segments.
• Figure 4: Performance comparison of Accuracy (circles, left axis) and Average Segments Cost (squares, right axis) with respect to increasing hop counts.
• Figure 5: Performance comparison on reasoning benchmarks. We compare CoSMo against CoT and C3oT baselines across different model scales. (a) Results on HotpotQA dataset. (b) Results on CRAG dataset. Our method consistently outperforms baselines.