摘要：

目的 探索中文期刊利用大语言模型（LLM）辅助同行评审的实践路径。 方法 检索Web of Science、Scopus、中国知网、万方数据库、预印本平台收集文献，总结LLM审稿可行性证据。综合权衡利弊后决定租用“移动云”虚拟服务器完成本地私有化部署LLM。采用本地私有化DeepSeek-r1回顾性评审10篇综述（5篇发表+5篇退稿），重复5次，与专家审稿结果进行对比。 结果 文献证据显示，LLM检查错误、遵循指南的能力与专家的能力相当，甚至更优，但其应用存在幻觉现象、数据安全与隐私泄露风险、透明性与可解释性不足等问题。中文期刊实践LLM同行评审路径包括本地私有化部署模型、构建个性化模型、探索模型工作规律、规范模型使用方法、规避伦理风险。本地私有化部署DeepSeek-r1对10篇综述的评审结论为小修或大修后发表，未提出拒绝结论，且稳定性不佳，批判性意见侧重于规范与方法；而专家侧重于临床诊疗细节。 结论 华西期刊社本地私有化部署同行评审模型实践路径切实可行，后续应从技术和伦理规范角度提升效能、降低风险。

关键词: 大语言模型, 同行评审, 私有化部署, 实践路径

Abstract:

Purposes To explore the potential of large language models （LLMs） in assisting peer review and to establish a pathway for their localized private deployment. Methods We conducted literature searches in Web of Science， Scopus， CNKI， Wanfang Data， and preprint platforms to combine evidence on the feasibility of using LLMs for manuscript review. After weighing the advantages and disadvantages， we decided to rent a virtual server from "Mobile Cloud" to achieve local and privatized deployment of an LLM. Using the locally deployed DeepSeek-R1 model， we performed a retrospective review of 10 review articles （5 published papers and 5 rejected papers）， repeating the process five times for each manuscript， and compared the results with human reviews. Findings Literature evidence indicates that LLMs possess capabilities in error checking and guideline adherence comparable to humans； however， risks such as hallucinations， data security and privacy leakage， and insufficient transparency and explainability persist. The proposed pathway for Chinese journals to implement LLM-assisted peer review includes： local privatized model deployment， building personalized models， exploring model operational patterns， standardizing usage protocols， and mitigating ethical risks. The locally deployed DeepSeek-R1 recommended “accept after minor/major revision” for all 10 reviews， with no rejections， but exhibited poor stability across repeated trials. Its critical feedback focused primarily on formal and methodological aspects， whereas human reviewers emphasized clinical and diagnostic details. Conclusions The local privatized deployment of a peer review model at West China Journal Press proves to be a feasible practice. Future efforts should focus on enhancing efficacy and reducing risks through technological improvements and ethical-standard development.

Key words: Large language models, Peer review, Private model deployment, Practical pathway