---
id: 20261107-T0-04
title: "弹性专家并行：降低MoE模型强化学习后训练算力浪费"
title_en: "Elastic Expert Parallelism Cuts Compute Waste in MoE RL Post-Training"
url: https://ai.daily.yangsir.net/daily/20261107-T0-04
issue_date: 2026-11-07
publish_date: 2026-11-06T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.06984
---

# 弹性专家并行：降低MoE模型强化学习后训练算力浪费

arXiv新论文提出弹性专家并行方案，针对混合专家模型强化学习后训练中的算力瓶颈。RL后训练需要大量生成训练响应，MoE模型中先完成任务的worker常处于空闲状态，造成算力浪费。该方案通过弹性调整专家并行度，让空闲worker参与其他任务，提升整体利用率。对大规模MoE模型的RL训练而言，这可降低单位训练成本，缩短后训练周期。

## English Version

**Elastic Expert Parallelism Cuts Compute Waste in MoE RL Post-Training**

A new arXiv paper proposes elastic expert parallelism to address a compute bottleneck in reinforcement learning post-training for mixture-of-experts models. RL post-training requires generating many training responses, but in MoE models workers that finish early sit idle. The approach elastically adjusts expert parallelism so idle workers take on other work, improving utilization. For large-scale MoE RL training, this can lower cost per training run and shorten post-training cycles.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.06984)

**详情页**：https://ai.daily.yangsir.net/daily/20261107-T0-04

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*