EVENT DOSSIER
An analysis of 5,242 biomedical research articles related to large language models (LLMs) published in PubMed from 2022 to 2026 showed that 42% of them cited models that had been discontinued or would be discontinued within two years after publication. Researchers extracted and normalized model names from 61,077 abstracts and found that 77.7% of the 8,931 paper-model mentions referred to commercial closed-source models. Due to the disassembly schedules of many commercial services, model discontinuation after an average of 538 days led to a large number of publications that could no longer be reproduced, posing a risk of computational unreproducibility for biomedical AI publications. The study suggests that model disassembly should be considered a core issue in reporting and preservation of biomedical research.
一项针对 2022 年至 2026 年 3 月期间发表于 PubMed 的 5,242 篇涉及大语言模型（LLMs）的生物医学研究文章的分析显示，其中 42% 引用了已退役或将在发表后两年内退役的模型。研究人员从 61,077 篇摘要中提取并归一化了模型名称，发现 8,931 次论文 - 模型提及中，77.7% 涉及商业闭源模型，而从中位数 538 天后的模型退役导致大量发表后无法复现。研究指出，由于许多商用服务存在弃用时间表，生物医学 AI 出版物正面临计算不可复现性风险，模型弃用应被视为生物医学研究的核心报告与保存问题。