arXiv 2610.28027 · cs.LG · 2026-10-20

From RL Papers to Framework Code: Inclusion and Review in Seven Open-Source Training Frameworks

Taiheng Pan

机构待核实

一句话要点

研究追踪七个开源框架对2025至2026年arXiv RL论文的采纳情况，发现仅少数方法被纳入，且多数未被公开回应或实际落地。

Only a small fraction of RL methods are adopted by open-source frameworks, with most proposals remaining unaddressed or unpublished.

要点

RL方法在开源框架中采纳率低，许多论文未被实际实现或公开回应。
通过分析七个框架的提交历史和6939篇arXiv论文，追踪方法被纳入的情况及特征。
1598篇论文中仅37篇被纳入，且被采纳的方法多涉及大模型（30B参数以上）或损失函数调整，但未见公司主导的偏倚。
意义与局限：多数方法无公开引用或回复，用户很少报告训练失败，表明技术落地存在明显断层。

English overview

This study analyzes the adoption of reinforcement learning methods from arXiv papers into seven open-source training frameworks. It finds that only a small fraction of proposed methods are implemented, with most remaining unaddressed or unpublished. Adopted methods tend to involve large-scale policies and loss function modifications. Public issue tracking shows minimal user-reported failures, and many proposals remain open without response, indicating significant gaps between research and implementation.

只根据 arXiv 官方摘要，由开源模型整理，经数字/专名自动核对（必须在摘要中出现），以论文原文为准。

14HF 点赞

–GitHub stars

–引用（Semantic Scholar）

–热度分

阅读 arXiv 原文 PDF Hugging Face 讨论代码仓库项目页