短链，深思：通过分裂-合并优化平衡推理效率与内部段能力
计算与语言
2026-05-04 v3
摘要
虽然大型推理模型（LRM）通过生成长推理链在解决复杂任务方面展现出惊人的能力，但这种对冗长生成的依赖导致显著的延迟和计算开销。为此，我们提出CoSMo（Consistency-Guided Split-Merge Optimization，即“一致性引导的分裂-合并优化”），一个旨在消除结构性冗余而非单纯限制token数量的框架。具体而言，CoSMo利用分裂-合并算法动态细化推理链，通过合并冗余段落和分裂逻辑间隙来确保连贯性。我们随后采用结构对齐的强化学习，并引入 novel segment-level budget 来监督模型在训练期间保持高效的推理结构。广泛的实验表明，CoSMo 在多个基准和不同骨干网络上均取得优异性能，平均将准确率提高了3.3个百分点，同时将段落使用量降低了28.7%。
引用
@article{arxiv.2602.03141,
title = {Short Chains, Deep Thoughts: Balancing Reasoning Efficiency and Intra-Segment Capability via Split-Merge Optimization},
author = {Runquan Gui and Jie Wang and Zhihai Wang and Chi Ma and Jianye Hao and Feng Wu},
journal= {arXiv preprint arXiv:2602.03141},
year = {2026}
}
备注
This is a revised version of arXiv:2602.03141. The previous withdrawal was due to a misalignment in publication timing. All authors have now unanimously approved this submission, and the manuscript is resubmitted with full author consent