---
id: 20261022-T0-14
title: "miniTE：定位NVFP4强化学习的执行差距，让训练和采样引擎数值对齐"
title_en: "miniTE Diagnoses the Execution Gap in Native NVFP4 Reinforcement Learning"
url: https://ai.daily.yangsir.net/daily/20261022-T0-14
issue_date: 2026-10-22
publish_date: 2026-10-21T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.23928
---

# miniTE：定位NVFP4强化学习的执行差距，让训练和采样引擎数值对齐

原生NVFP4有望加速大语言模型的强化学习，但在训练引擎与采样引擎分离的架构下实现稳定学习仍有难度。研究发现，权重与数值格式的匹配是关键问题之一，并据此提出miniTE来诊断和弥合这一执行差距。该工作针对的是低精度RL训练中的数值不一致问题，对使用分离式训练与推理架构做RL的团队有直接参考价值。

## English Version

**miniTE Diagnoses the Execution Gap in Native NVFP4 Reinforcement Learning**

Native NVFP4 promises to accelerate reinforcement learning for large language models, but stable learning remains difficult when training and sampling run in separate engines. The paper identifies matching model weights and numerical formats as a key issue and introduces miniTE to diagnose and bridge this execution gap. The work targets numerical inconsistency in low-precision RL training and is directly relevant to teams running RL with decoupled training and inference stacks.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.23928)

**详情页**：https://ai.daily.yangsir.net/daily/20261022-T0-14

---

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