研究揭示免训练MoE专家剪枝中门控重新归一化致性能损失

1. 13:00
   arXiv 机器学习

   免训练 MoE 专家剪枝中的替换算子与路由语义研究

   研究通过在 OLMoE 与 DeepSeek-V2-Lite 等模型上固定剪枝掩码，发现免训练 MoE 专家剪枝中门控重新归一化会导致 0.8-8.5 点性能损失。

Current heat 9·Comparable peak 10（Nov 2 14:00）·Comparable change, last 24 hours–

Trends compare only the same parties observed continuously, so the scope can be narrower than current heat counts. Move the pointer or tap the chart to see hourly heat; use the left and right arrow keys on a keyboard.