arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2606.03391cs.LGcs.AIcs.CL
收藏

当模型合并破坏路由:MoE的无训练校准

When Model Merging Breaks Routing: Training-Free Calibration for MoE

University of Science and Technology of China(中国科学技术大学)

Canbin Huang, Tianyuan Shi, Xiaojun Quan, Jingang Wang, Jianfei Zhang, Qifan Wang

AI总结:

针对MoE架构中模型合并导致的路由崩溃问题,提出基于二阶曲率的无训练校准方法HARC,通过闭式解和共轭梯度法高效重对齐路由器,显著提升数学推理和代码生成性能。

摘要:

模型合并已成为一种无需重新训练即可整合多个LLM能力的成本效益方法。然而,现有的合并技术主要基于线性参数算术或优化,在应用于混合专家(MoE)架构时面临困难。我们识别出MoE合并中的一个关键失效模式,称为路由崩溃,其中合并后的路由器无法将令牌分派给合适的专家。路由崩溃源于非线性softmax和离散Top-k路由机制对合并引起的参数扰动的敏感性,这种敏感性进一步被MoE预训练期间施加的负载平衡约束放大。由于微调后的专家表现出不同的专长,即使是适度的错误路由也可能导致严重的性能下降。为解决此问题,我们提出Hessian感知路由器校准(HARC),一种无训练框架,利用二阶曲率信息重新对齐合并后的路由器。该方法采用闭式解,可通过无矩阵共轭梯度法高效求解。在数学推理和代码生成任务上的实验表明,HARC有效缓解了多种MoE合并基线中的路由崩溃,并带来了显著的性能提升。我们的代码可在该https URL获取。

英文摘要:

Model merging has emerged as a cost-effective approach for consolidating the capabilities of multiple LLMs without retraining. However, existing merging techniques, largely based on linear parameter arithmetic or optimization, struggle when applied to Mixture-of-Experts (MoE) architectures. We identify a critical failure mode in MoE merging, termed routing breakdown, in which the merged router fails to dispatch tokens to suitable experts. Routing breakdown stems from the sensitivity of the non-linear softmax and discrete Top-k routing mechanisms to parameter perturbations from merging, a sensitivity further amplified by load-balancing constraints imposed during MoE pretraining. Because fine-tuned experts exhibit distinct specializations, even modest misrouting can cause severe performance degradation. To address this issue, we propose Hessian-Aware Router Calibration (HARC), a training-free framework that leverages second-order curvature information to realign the merged router. This approach admits a closed-form solution that can be efficiently solved using a matrix-free conjugate gradient method. Experiments on mathematical reasoning and code generation tasks show that HARC effectively mitigates routing breakdown across diverse MoE merging baselines and leads to substantial performance improvements. Our code is available at https://github.com/huangcb01/HARC.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑