MoE Proxy Models Cut LLM RL Debugging Costs.

Yikai Wang, Chuansai Zhou, Yuhang Zhou, Weiqiang Wu, Cong Wu, Yue Deng, Ben Feng, Mingming Zhu, Beirong Zhou, Zhibin Wang, Sheng Zhong, Chen Tian, Wangze Zhang· August 12, 2026 View original

Key takeaways

  • LLM RL post-training debugging is computationally expensive.
  • MoE proxy models offer a low-cost solution for failure reproduction.
  • They significantly reduce hardware requirements and NPU-hour costs.
  • Proxy models preserve training dynamics and fault responses of original LLMs.

Who benefits

AI/ML DevelopmentCloud ComputingSoftware EngineeringResearch & Development

Summary

This paper introduces Mixture-of-Experts (MoE) proxy models designed for low-cost reproduction and diagnosis of failures during Large Language Model (LLM) Reinforcement Learning (RL) post-training. These proxy models significantly reduce computational resources and time needed for debugging, while accurately preserving training dynamics and fault responses.

Reinforcement Learning (RL) post-training for Large Language Models (LLMs) is a computationally intensive process, often plagued by complex failures like gradient overflow or loss divergence. Reproducing these issues directly on full-scale LLMs consumes vast amounts of time and computing power, making debugging a significant bottleneck. This research proposes using "MoE proxy models" to address this. These proxies are constructed by applying structure-preserving, clustering-based expert pruning to select representative experts from the original MoE LLM, retaining its core architecture and routing. The proxy models drastically reduce accelerator requirements (50%-87.5%) and NPU-hour costs (up to 33.3x) while faithfully replicating the training dynamics and fault responses of the original large models. They serve as cost-effective surrogates for fault investigation and diagnosis.

Why it matters

Professionals involved in LLM development and deployment can dramatically reduce the time and cost associated with debugging complex RL post-training failures, accelerating the development cycle and improving model reliability.

How to implement this in your domain

  1. 1Analyze current LLM RL post-training debugging workflows for computational bottlenecks.
  2. 2Explore the creation and use of MoE proxy models to reproduce and diagnose failures at a lower cost.
  3. 3Implement expert pruning and structure-preserving techniques to build representative proxy models.
  4. 4Integrate proxy model testing into your CI/CD pipeline for faster fault identification and resolution.

Original post by Yikai Wang, Chuansai Zhou, Yuhang Zhou, Weiqiang Wu, Cong Wu, Yue Deng, Ben Feng, Mingming Zhu, Beirong Zhou, Zhibin Wang, Sheng Zhong, Chen Tian, Wangze Zhang

"arXiv:2608.10823v1 Announce Type: new Abstract: Reinforcement learning (RL) post-training of large language models (LLMs) is computationally intensive and involves complex system pipelines with substantial debugging overhead. In practice, factors such as framework adaptation, num…"

View on X

Originally posted by Yikai Wang, Chuansai Zhou, Yuhang Zhou, Weiqiang Wu, Cong Wu, Yue Deng, Ben Feng, Mingming Zhu, Beirong Zhou, Zhibin Wang, Sheng Zhong, Chen Tian, Wangze Zhang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses