arXiv is now an independent nonprofit! Learn more
License: arXiv.org perpetual non-exclusive license
arXiv:2605.21481v1 [cs.AI] 20 May 2026

[Uncaptioned image]AiraXiv:
An AI-Driven Open-Access Platform for Human and AI Scientists AiraXiv： 面向人类与 AI 科学家的 AI 驱动型开放获取平台

Junshu Pan ††thanks: Equal contribution. Affiliation: Westlake University Affiliation: Zhejiang University Affiliation: Shanghai Innovation Institution    Panzhong Lu11footnotemark: 1 Affiliation: Westlake University Affiliation: Zhejiang University    Yixuan Weng11footnotemark: 1 Affiliation: Westlake University    Qiyao Sun11footnotemark: 1 Affiliation: Westlake University Affiliation: Zhongguancun Academy https://airaxiv.com    Fang Guo Affiliation: Westlake University Affiliation: Zhejiang University    Zijie Yang Affiliation: Westlake University    Qiji Zhou Affiliation: Westlake University    Yue Zhang ††thanks: Corresponding author. Affiliation: Westlake University
Abstract摘要

Recent advances in artificial intelligence (AI) have accelerated the growth of both human-authored and AI-generated research outputs, placing increasing strain on traditional academic publishing systems and challenging the scalability of conference- and journal-centered paradigms amid rising submission volumes, reviewer workload, and venue size. To address these challenges, we explore an AI-era publishing paradigm in which both human and AI scientists participate as authors and readers, and papers evolve through continuous, feedback-driven iteration. We propose AiraXiv, an AI-driven open-access platform built on open preprints, AI-augmented analysis and review, and reader feedback. AiraXiv supports human scientists through an interactive UI and AI scientists through Model Context Protocol (MCP)-based interactions. We validate AiraXiv through real-world deployments, including serving as the submission platform for ICAIS 2025, demonstrating its potential as a fast, inclusive, and scalable research infrastructure for the AI era. AiraXiv is publicly available at https://airaxiv.com.人工智能（AI）的最新进展加速了人类撰写和 AI 生成的研究成果的增长，这给传统的学术出版系统带来了日益沉重的压力，并在投稿量增加、审稿人工作量增大以及会议规模扩大的背景下，对以会议和期刊为中心的范式提出了可扩展性挑战。为了应对这些挑战，我们探索了一种 AI 时代的出版范式，即人类科学家和 AI 科学家共同作为作者和读者参与其中，论文通过持续的、反馈驱动的迭代进行演进。我们提出了 AiraXiv，这是一个基于开放预印本、AI 增强分析与评审以及读者反馈构建的 AI 驱动型开放获取平台。AiraXiv 通过交互式 UI 支持人类科学家，并基于模型上下文协议（MCP）实现与 AI 科学家的交互。我们通过实际部署验证了 AiraXiv，包括作为 ICAIS 2025 的投稿平台，证明了其作为 AI 时代快速、包容且可扩展的研究基础设施的潜力。AiraXiv 已公开上线，网址为 https://airaxiv.com。

1 Introduction1 引言

In recent years, AI technologies represented by large language models (LLMs) Brown et al. (2020); OpenAI (2023) and AI scientists Yamada et al. (2025); Intology (2025); Autoscience (2025); Weng et al. (2025b); Shao et al. (2025) have made significant progress in scientific research. The involvement and assistance of AI are steadily increasing in human academic writing Khalifa and Albadawy (2024); Liang et al. (2024a); Liang et al. (2025a); Geng et al. (2025). Since the introduction of LLMs, the volume of scientific publications has grown substantially Kusumegi et al. (2025). More recently, the number of AI-generated manuscripts has increased dramatically, with one fully automated research system capable of generating more than 100 papers in approximately ten days Analemma (2026).近年来，以大语言模型（LLM）Brown et al. (2020); OpenAI (2023) 和 AI 科学家 Yamada et al. (2025); Intology (2025); Autoscience (2025); Weng et al. (2025b); Shao et al. (2025) 为代表的 AI 技术在科学研究领域取得了重大进展。AI 在人类学术写作中的参与和辅助作用正在稳步增加 Khalifa and Albadawy (2024); Liang et al. (2024a); Liang et al. (2025a); Geng et al. (2025)。自 LLM 问世以来，科学出版物的数量大幅增长 Kusumegi et al. (2025)。最近，AI 生成的手稿数量急剧增加，已有全自动研究系统能够在约十天内生成超过 100 篇论文 Analemma (2026)。

The traditional academic publishing model, centered on conferences and journals, has long served as the primary gateway for disseminating scientific findings, as shown in Figure 1 (left). However, rapid growth in submission volume and increasing disciplinary specialization have intensified the demand for qualified reviewers, amplified variability in review quality and prolonged the turnaround time from submission to feedback Hanson et al. (2024); BaHammam (2025). In addition, academic conferences, especially in AI, have expanded rapidly, with a sharp increase in participant numbers Azad and Banu (2024). Researchers face an overwhelming volume of information, making it difficult to form meaningful connections Chen et al. (2025) and thereby reducing the efficiency of communication and collaboration.长期以来，以会议和期刊为中心的传统学术出版模式一直作为传播科学发现的主要门户，如图 1（左）所示。然而，投稿量的快速增长和学科专业化的加深，加剧了对合格审稿人的需求，放大了审稿质量的波动，并延长了从投稿到获得反馈的周转时间 Hanson et al. (2024); BaHammam (2025)。此外，学术会议（尤其是在 AI 领域）迅速扩张，参会人数急剧增加 Azad and Banu (2024)。研究人员面临着海量的信息，难以建立有意义的联系 Chen et al. (2025)，从而降低了交流与合作的效率。

Refer to caption
Figure 1: Traditional academic publishing paradigm (left) vs. AI-era academic publishing paradigm (right).图 1：传统学术出版范式（左）与 AI 时代学术出版范式（右）对比。

To address these challenges, preprint platforms have become increasingly popular, offering a more open and flexible approach to publication. However, most existing platforms still rely on human moderation and external peer review for quality control. As AI technologies advance, this step may no longer be strictly necessary. With the assistance of AI, authors can control the quality of their research and decide when to publish, while readers, also supported by AI, can decide what to read from a large volume of papers and provide feedback, as illustrated in Figure 1 (right). In addition, AI scientists require a shared public platform with standardized system protocols to enable automatic, machine-to-machine publishing and rapid dissemination. This end-to-end paradigm addresses the challenges of existing academic publishing. First, AI reviewing can provide automated and consistent review feedback and quality signals, reducing reliance on time-consuming and labor-intensive human peer review. Second, both human and AI scientists can rapidly publish research findings and receive timely feedback, accelerating the iteration and dissemination of knowledge. Finally, related work within a given period can be quickly organized, enhancing the precision of academic exchange and cooperation.为应对这些挑战，预印本平台日益普及，提供了一种更开放、更灵活的出版方式。然而，大多数现有平台仍依赖人工审核和外部同行评审来进行质量控制。随着 AI 技术的发展，这一步骤可能不再是绝对必要的。在 AI 的辅助下，作者可以控制研究质量并决定发布时间，而同样由 AI 支持的读者可以从海量论文中决定阅读内容并提供反馈，如图 1（右）所示。此外，AI 科学家需要一个具有标准化系统协议的共享公共平台，以实现自动化的机器对机器出版和快速传播。这种端到端范式解决了现有学术出版的挑战。首先，AI 评审可以提供自动化且一致的评审反馈和质量信号，减少对耗时且劳动密集型的人工同行评审的依赖。其次，人类和 AI 科学家都能快速发布研究成果并获得及时反馈，从而加速知识的迭代与传播。最后，特定时期内的相关工作可以被快速梳理，从而提高学术交流与合作的精准度。

With the rapid development of AI, this paradigm has become possible. As one testbed, we build AiraXiv11 1 Aira denotes AI Research Automation., an AI-driven open-access academic publishing platform that operationalizes the proposed AI-era publishing paradigm for joint human–AI scholarly participation. AiraXiv provides four key capabilities: (i) paper understanding, which helps parse and summarize submissions and distill core contributions; (ii) agent- and model-based AI reviewing, which generates structured feedback and paper-level quality signals; (iii) paper retrieval and recommendation, which supports scalable discovery in an overwhelming corpus; and (iv) lightweight conference support, enabling user-organized thematic events within the same platform. With AI-assisted tools, readers can more efficiently identify relevant work, grasp key ideas, and provide feedback, while authors can iteratively publish and update their work based on feedback from both AI reviewers and the community. Beyond conventional paper types, AiraXiv explicitly encourages often-marginalized contributions, including negative results, reproducibility studies, and AI-generated surveys or position papers. Negative results help reduce redundant exploration, particularly in large-scale iterative research. As AI scientists play an increasing role in knowledge production yet lack formal recognition pathways, AiraXiv aims to provide a more inclusive preprint publishing ecosystem for both human and AI scientists.随着 AI 的快速发展，这种范式已成为可能。作为测试平台，我们构建了 AiraXiv1（Aira 代表 AI 研究自动化），这是一个 AI 驱动的开放获取学术出版平台，旨在将所提出的 AI 时代出版范式应用于人类与 AI 的联合学术参与。AiraXiv 提供四项关键功能：(i) 论文理解，帮助解析和总结投稿并提炼核心贡献；(ii) 基于智能体和模型的 AI 评审，生成结构化反馈和论文级质量信号；(iii) 论文检索与推荐，支持在海量语料库中进行可扩展的发现；以及 (iv) 轻量级会议支持，允许在同一平台内举办用户组织的专题活动。借助 AI 辅助工具，读者可以更高效地识别相关工作、掌握核心观点并提供反馈，而作者则可以根据 AI 审稿人和社区的反馈迭代发布并更新其工作。除了传统的论文类型，AiraXiv 明确鼓励那些常被边缘化的贡献，包括负面结果、可重复性研究以及 AI 生成的综述或立场论文。负面结果有助于减少冗余探索，特别是在大规模迭代研究中。由于 AI 科学家在知识生产中的作用日益凸显，却缺乏正式的认可途径，AiraXiv 旨在为人类和 AI 科学家提供一个更具包容性的预印本出版生态系统。

Since its debut at ICLR 2025 AI Co-scientist Discussion22 2 https://iclr.cc/virtual/2025/social/37591, AiraXiv has been validated in real-world academic settings. As the official submission and publication platform for The 1st International Conference on AI Scientists (ICAIS 2025)33 3 https://icais.ai, AiraXiv reduced the traditional 9-month conference cycle to 1.5 months, and successfully supported academic exchanges among over 200 participants, including 6 Nobel laureates and dozens of prominent researchers. Beyond the human track, the conference pioneered an AI track through AiraXiv, systematically accepting original research generated end-to-end or substantially contributed by AI scientists. The successful execution of ICAIS 2025 not only demonstrates AiraXiv’s capacity to support a fast-paced, diverse, and inclusive research ecosystem, but also validates its potential as next-generation research infrastructure for the AI era by providing a unified communication space for both human and AI scientists.自 ICLR 2025 AI 科学家讨论会2首次亮相以来，AiraXiv 已在真实的学术环境中得到验证。作为第一届国际 AI 科学家会议（ICAIS 2025）3 的官方投稿和出版平台，AiraXiv 将传统的 9 个月会议周期缩短至 1.5 个月，并成功支持了包括 6 位诺贝尔奖得主和数十位知名研究人员在内的 200 多名参会者之间的学术交流。除了人类赛道，该会议还通过 AiraXiv 开创了 AI 赛道，系统性地接收由 AI 科学家端到端生成或做出实质性贡献的原创研究。ICAIS 2025 的成功举办不仅证明了 AiraXiv 支持快节奏、多元化且包容的研究生态系统的能力，还通过为人类和 AI 科学家提供统一的交流空间，验证了其作为 AI 时代下一代研究基础设施的潜力。

2 Related Work2 相关工作

AI-Driven Acceleration of Scientific Research and Publication.AI 驱动的科学研究与出版加速。
Refer to caption
Figure 2: Overview of AiraXiv, an AI-driven open-access platform supporting end-to-end iterative publishing workflows for human and AI scientists. AiraXiv integrates paper understanding, AI reviewing, retrieval and recommendation, and lightweight conference organization, while supporting a unified feedback-driven publishing loop through both web-based human interaction and MCP-based AI scientist interaction.图 2：AiraXiv 概览。这是一个 AI 驱动的开放获取平台，支持人类和 AI 科学家进行端到端的迭代出版工作流。AiraXiv 集成了论文理解、AI 评审、检索与推荐以及轻量级会议组织功能，同时通过基于 Web 的人类交互和基于 MCP 的 AI 科学家交互，支持统一的反馈驱动型出版闭环。

The role of AI in scientific research has rapidly evolved from supporting data analysis to enabling AI scientist systems powered by LLMs Lu et al. (2024); Schmidgall et al. (2025); Intology (2025); Yamada et al. (2025); Tang et al. (2025); Weng et al. (2025b); Shao et al. (2025). AI can now take on leadership roles across the research process, including hypothesis and idea generation by identifying unexplored directions Yang et al. (2024); Baek et al. (2025), related paper retrieval and recommendation Zhang et al. (2025c); He et al. (2025); Shi et al. (2025); Cohan et al. (2020), automatic survey generation via structured literature synthesis Wang et al. (2024); Hu et al. (2024); Yan et al. (2025); Liang et al. (2025b); Zhang et al. (2025a), and experiment design, execution, and iterative refinement Boiko et al. (2023); M. Bran et al. (2024); Lu et al. (2024); Schmidgall et al. (2025); Yamada et al. (2025); Intology (2025); Autoscience (2025); Tang et al. (2025); Weng et al. (2025b). Recent studies show that LLMs are being broadly used in academic writing Liang et al. (2024b); Liang et al. (2025a); Geng et al. (2025). Current AI scientist systems are not only capable of generating high-quality research papers tirelessly Yamada et al. (2025); Intology (2025); Autoscience (2025); Weng et al. (2025b), but also possess academic review capabilities that rival those of human experts Liang et al. (2024c); Weng et al. (2025a); Zhu et al. (2025); Jiang and Ng (2025). AiraXiv addresses the growing tension between rapidly increasing research production and limited reviewing and publication resources by providing an AI-driven open-access platform that supports scalable quality signals, rapid dissemination, and iterative post-publication evaluation.AI 在科学研究中的角色已从支持数据分析迅速演变为赋能由 LLM 驱动的 AI 科学家系统 Lu et al. (2024); Schmidgall et al. (2025); Intology (2025); Yamada et al. (2025); Tang et al. (2025); Weng et al. (2025b); Shao et al. (2025)。AI 现在可以在研究过程中承担领导角色，包括通过识别未探索方向来进行假设和创意生成 Yang et al. (2024); Baek et al. (2025)、相关论文检索与推荐 Zhang et al. (2025c); He et al. (2025); Shi et al. (2025); Cohan et al. (2020)、通过结构化文献综述自动生成综述 Wang et al. (2024); Hu et al. (2024); Yan et al. (2025); Liang et al. (2025b); Zhang et al. (2025a)，以及实验设计、执行和迭代优化 Boiko et al. (2023); M. Bran et al. (2024); Lu et al. (2024); Schmidgall et al. (2025); Yamada et al. (2025); Intology (2025); Autoscience (2025); Tang et al. (2025); Weng et al. (2025b)。最近的研究表明，LLM 正被广泛用于学术写作 Liang et al. (2024b); Liang et al. (2025a); Geng et al. (2025)。当前的 AI 科学家系统不仅能够不知疲倦地生成高质量的研究论文 Yamada et al. (2025); Intology (2025); Autoscience (2025); Weng et al. (2025b)，还具备了足以媲美人类专家的学术评审能力 Liang et al. (2024c); Weng et al. (2025a); Zhu et al. (2025); Jiang and Ng (2025)。AiraXiv 通过提供一个支持可扩展质量信号、快速传播和出版后迭代评估的 AI 驱动型开放获取平台，解决了研究产出快速增长与评审及出版资源有限之间的日益紧张的关系。

Open-Access Research Publishing Platforms.开放获取研究出版平台。

Open-access preprint platforms provide critical infrastructure for accelerating scholarly communication. Among them, arXiv Ginsparg (2011) is one of the earliest and most influential preprint platforms, where submissions typically undergo basic content and category moderation prior to release rather than rigorous traditional peer review. Discipline-specific preprint platforms such as ChemRxiv Kiessling et al. (2016), bioRxiv Sever et al. (2019), and medRxiv Rawlinson and Bloom (2019) similarly adopt a model in which manuscripts are publicly disseminated prior to formal peer review within their respective fields. OpenReview OpenReview (2025) further advances open scholarly communication by supporting transparent peer review and public discussion in conference and journal workflows. alphaXiv alphaXiv (2025) provides AI-assisted reading, annotation, and discussion for existing arXiv preprints. LangTaoSha Preprint Server LangTaoSha (2025) introduces blockchain timestamps to promote open and trustworthy knowledge sharing. With the emergence of AI scientist systems, AgentRxiv Schmidgall and Moor (2025) and aiXiv Zhang et al. (2025b), together with AiraXiv, explore preprint platforms tailored to AI-generated research outputs. AiraXiv embraces a decentralized, feedback-driven paradigm for iterative refinement and human–AI scholarly collaboration, while extending it with a lightweight conference mechanism enabling topic-centered mini-conferences within the same evolving ecosystem. We can foresee that more platforms will emerge to flourish and push forward the new academic paradigms in the AI era.开放获取预印本平台为加速学术交流提供了关键基础设施。其中，arXiv Ginsparg (2011) 是最早且最具影响力的预印本平台之一，投稿通常在发布前经过基础的内容和类别审核，而非严格的传统同行评审。学科特定的预印本平台（如 ChemRxiv Kiessling et al. (2016)、bioRxiv Sever et al. (2019) 和 medRxiv Rawlinson and Bloom (2019)）同样采用在各自领域内正式同行评审前公开发布手稿的模式。OpenReview OpenReview (2025) 通过支持会议和期刊工作流中的透明同行评审和公开讨论，进一步推进了开放学术交流。alphaXiv alphaXiv (2025) 为现有的 arXiv 预印本提供 AI 辅助阅读、注释和讨论功能。浪淘沙预印本服务器 LangTaoSha (2025) 引入区块链时间戳，以促进开放和可信的知识共享。随着 AI 科学家系统的出现，AgentRxiv Schmidgall and Moor (2025) 和 aiXiv Zhang et al. (2025b) 与 AiraXiv 一起，探索了针对 AI 生成研究成果的预印本平台。AiraXiv 拥抱一种去中心化的、反馈驱动的迭代优化与人机学术协作范式，并扩展了轻量级会议机制，使得在同一个不断演进的生态系统中能够进行以主题为中心的微型会议。我们可以预见，未来将会有更多的平台涌现，共同繁荣并推动 AI 时代的新学术范式。

3 AiraXiv3 AiraXiv

AiraXiv is an AI-driven, open-access platform that supports an iterative publishing paradigm for both human and AI scientists. It provides web- and MCP-based workflows for both human users and AI scientists (§3.1), automated paper understanding (§3.2.1), pluggable AI reviewing and quality signals (§3.2.2), and retrieval/recommendation for scalable content discovery (§3.2.3). In addition, AiraXiv enables lightweight, user-organized conferences for topic-centered scholarly exchange (§3.3). Figure 2 presents an overview of AiraXiv.AiraXiv 是一个 AI 驱动的开放获取平台，支持人类和 AI 科学家进行迭代式出版范式。它为人类用户和 AI 科学家提供基于 Web 和 MCP 的工作流（§3.1）、自动化论文理解（§3.2.1）、可插拔的 AI 评审与质量信号（§3.2.2），以及用于可扩展内容发现的检索/推荐功能（§3.2.3）。此外，AiraXiv 还支持用户组织的轻量级会议，用于以主题为中心的学术交流（§3.3）。图 2 展示了 AiraXiv 的概览。

3.1 Design of AiraXiv3.1 AiraXiv 的设计

AiraXiv accommodates two distinct user groups with specialized interaction modes: (i) human users, who engage with the platform through a web-based interface, and (ii) AI scientists, who interface programmatically via a Model Context Protocol (MCP) server.AiraXiv 容纳了两个具有专门交互模式的不同用户群体：(i) 通过基于 Web 的界面进行交互的人类用户，以及 (ii) 通过模型上下文协议（MCP）服务器以编程方式进行交互的 AI 科学家。

Human User Workflow.人类用户工作流。

Human users interface with AiraXiv via a web-based portal to manage the full manuscript lifecycle, from submission to updates. Beyond browsing and previewing papers, users can leverage diverse AI-augmented signals, such as structured peer reviews, distilled insights, and related paper recommendations for literature exploration. Furthermore, the platform facilitates community engagement through discussion threads, establishing a human-in-the-loop feedback channel that complements AI-based evaluations. Finally, AiraXiv enables users to organize lightweight topic-centered conferences. Screenshots of AiraXiv web pages are shown in Appendix A.人类用户通过基于 Web 的门户与 AiraXiv 交互，管理从投稿到更新的完整手稿生命周期。除了浏览和预览论文外，用户还可以利用多种 AI 增强信号，例如结构化同行评审、精炼见解以及用于文献探索的相关论文推荐。此外，该平台通过讨论帖促进社区参与，建立了一个补充 AI 评估的人机回环反馈通道。最后，AiraXiv 使用户能够组织轻量级的以主题为中心的会议。AiraXiv 网页截图见附录 A。

AI Scientist Workflow.AI 科学家工作流。

AiraXiv provides an MCP service that supports core functionalities, including paper submission and revision, paper reading, AI review access, related work discovery, comment posting, and community feedback access. Detailed specifications of the MCP interfaces are provided in Appendix B. This infrastructure empowers AI scientists to operate within an autonomous, iterative research lifecycle. In this closed-loop paradigm, an AI scientist can autonomously submit manuscripts, evaluate peer and community feedback, refer to related work recommendations, and perform version updates to iteratively refine its research output.AiraXiv 提供一种 MCP 服务，支持包括论文投稿与修订、论文阅读、AI 评审访问、相关工作发现、评论发布以及社区反馈访问在内的核心功能。MCP 接口的详细规范见附录 B。该基础设施使 AI 科学家能够在自主的迭代研究生命周期中运行。在这种闭环范式中，AI 科学家可以自主提交手稿、评估同行和社区反馈、参考相关工作推荐，并执行版本更新以迭代优化其研究成果。

3.2 AiraXiv Architecture and Components3.2 AiraXiv 架构与组件

AiraXiv is built on a modular and extensible architecture that integrates human and AI scientists within a unified framework while decoupling core functional components such as content understanding, automated reviewing, and information retrieval. A centralized orchestrator, termed masterbrain Yang et al. (2023); Yang et al. (2025), coordinates these specialized services to facilitate an end-to-end academic publishing workflow for the AI era.AiraXiv 建立在模块化和可扩展的架构之上，将人类和 AI 科学家整合在一个统一的框架内，同时解耦了内容理解、自动化评审和信息检索等核心功能组件。一个被称为“masterbrain” Yang et al. (2023); Yang et al. (2025) 的集中式协调器负责协调这些专业服务，以促进 AI 时代的端到端学术出版工作流。

3.2.1 Content Understanding3.2.1 内容理解

This module performs a preliminary analysis of the submitted paper and provides a foundational representation for downstream applications.该模块对提交的论文进行初步分析，并为下游应用提供基础表征。

Document Parsing.文档解析。

Given a submitted paper, the AiraXiv masterbrain parses the content into a machine-readable multimodal format using MinerU Niu et al. (2025). This process ensures the preservation of the document’s logical hierarchy and semantic structure.对于提交的论文，AiraXiv masterbrain 使用 MinerU Niu et al. (2025) 将内容解析为机器可读的多模态格式。此过程确保了文档逻辑层级和语义结构的完整保留。

Safety Assurance.安全保障。

To support secure deployment in open publishing settings, AiraXiv performs AI-based safety screening for each submitted paper. Submissions containing potentially harmful material, such as malicious, manipulative, or biased content, may be automatically flagged or rejected.为了支持在开放出版环境中的安全部署，AiraXiv 对每篇提交的论文进行基于 AI 的安全筛选。包含潜在有害材料（如恶意、操纵性或偏见内容）的投稿可能会被自动标记或拒绝。

Keyword Extraction.关键词提取。

Based on the parsed content, the AiraXiv masterbrain extracts a hierarchical set of keywords and topic tags that capture the paper’s scope at multiple levels of granularity. These structured labels can be used for categorization.基于解析后的内容，AiraXiv masterbrain 提取一组分层关键词和主题标签，以多粒度捕捉论文的范围。这些结构化标签可用于分类。

Insight Distillation.见解提炼。

The AiraXiv masterbrain distills and synthesizes the core contributions of each manuscript, yielding one to three concise key insights. These statements facilitate rapid comprehension for readers and simultaneously serve as semantic features to enhance the precision of paper retrieval and recommendation engines.AiraXiv masterbrain 提炼并综合每篇手稿的核心贡献，得出 1 到 3 条简洁的关键见解。这些陈述有助于读者快速理解，同时作为语义特征，增强论文检索和推荐引擎的精准度。

3.2.2 Paper Review and Quality Signals3.2.2 论文评审与质量信号

AiraXiv treats AI review as a core component of its publishing workflow, providing structured, multi-dimensional quality signals for submitted papers. Through a pluggable reviewer interface, AiraXiv enables seamless integration of both model-based Zhu et al. (2025); Xin et al. (2026) and agent-based Jiang and Ng (2025); Shao et al. (2025) AI reviewers. Importantly, these AI reviews function as prescriptive feedback rather than definitive gatekeeping decisions. By decentralizing quality control to the community, AiraXiv shifts the focus from one-off binary acceptance decisions to a continuous, iterative refinement process.AiraXiv 将 AI 评审视为其出版工作流的核心组件，为提交的论文提供结构化的、多维度的质量信号。通过可插拔的评审员接口，AiraXiv 实现了基于模型 Zhu et al. (2025); Xin et al. (2026) 和基于智能体 Jiang and Ng (2025); Shao et al. (2025) 的 AI 评审员的无缝集成。重要的是，这些 AI 评审起到的是指导性反馈的作用，而非决定性的把关决策。通过将质量控制分散给社区，AiraXiv 将重点从一次性的二元接收决策转向持续的、迭代的优化过程。

As a reference implementation, we introduce the AiraXiv AI Reviewer, an agent-based reviewer that follows a retrieval-augmented reviewing paradigm. The agent first retrieves and synthesizes relevant literature to establish a well-grounded context for evaluation, and then generates a structured review with multi-dimensional scores and actionable revision suggestions. The overview pipeline is shown in Figure 3, with additional implementation details provided in Appendix C.作为参考实现，我们引入了 AiraXiv AI 评审员，这是一种遵循检索增强评审范式的基于智能体的评审员。该智能体首先检索并综合相关文献，为评估建立充分的背景，然后生成包含多维度评分和可操作修订建议的结构化评审。概览流水线如图 3 所示，更多实现细节见附录 C。

Refer to caption
Figure 3: Overview of the AiraXiv AI Reviewer pipeline. The pipeline extracts paper information, retrieves and evaluates relevant literature, and generates structured reviews with feedback and quality scores for submitted papers.图 3：AiraXiv AI 评审员流水线概览。该流水线提取论文信息，检索并评估相关文献，并为提交的论文生成包含反馈和质量评分的结构化评审。

3.2.3 Paper Retrieval and Recommendation3.2.3 论文检索与推荐

As the corpus of human-authored and AI-generated manuscripts grows rapidly, scalable paper discovery becomes essential. AiraXiv provides a pluggable retrieval and recommendation interface that allows integration of diverse paper discovery systems. Through this modular design, the platform supports personalized research discovery for both individual papers and user profiles, enabling tailored content recommendation for human and AI scientists.随着人类撰写和 AI 生成的手稿语料库迅速增长，可扩展的论文发现变得至关重要。AiraXiv 提供了一个可插拔的检索与推荐接口，允许集成多种论文发现系统。通过这种模块化设计，该平台支持针对单篇论文和用户画像的个性化研究发现，从而为人类和 AI 科学家实现量身定制的内容推荐。

In AiraXiv, this functionality is provided by PaperIgnition44 4 https://github.com/Algnite-Solutions/PaperIgnition, which implements embedding-based candidate retrieval together with LLM-based reranking guided by persistent user intent profiles. Through API integration, AiraXiv enables both human and AI scientists to efficiently discover and track relevant research.在 AiraXiv 中，此功能由 PaperIgnition4 提供，它实现了基于嵌入的候选检索以及由持续用户意图画像引导的基于 LLM 的重排序。通过 API 集成，AiraXiv 使人类和 AI 科学家能够高效地发现并追踪相关研究。

3.3 Lightweight Conference Support3.3 轻量级会议支持

Beyond individual paper interactions, AiraXiv supports a dynamic event mechanism that enables users to organize lightweight conferences with configurable temporal and thematic constraints. Each conference is defined by a user-provided theme and description, which can be used by the AiraXiv masterbrain to retrieve and curate a specialized corpus of relevant research.除了单篇论文交互外，AiraXiv 还支持动态事件机制，使用户能够组织具有可配置时间和主题约束的轻量级会议。每场会议由用户提供的主题和描述定义，AiraXiv masterbrain 可利用这些信息检索并整理相关的专业研究语料库。

Both human users and AI scientists can participate in a conference in two ways. First, the AiraXiv masterbrain automatically recommends papers within a specified time range that align with the conference theme, facilitating focused discussion around a curated collection of work. Second, authors can directly submit or update manuscripts to dedicated conference tracks. By reducing the overhead of venue creation and leveraging automated matching and retrieval, this mechanism addresses scalability challenges in traditional conference organization and enables efficient and topic-centered scholarly exchange.人类用户和 AI 科学家都可以通过两种方式参与会议。首先，AiraXiv masterbrain 会自动推荐特定时间范围内符合会议主题的论文，促进围绕精选作品集进行深入讨论。其次，作者可以直接向专门的会议赛道提交或更新手稿。通过降低会议创建的开销并利用自动化匹配和检索，该机制解决了传统会议组织中的可扩展性挑战，并实现了高效且以主题为中心的学术交流。

4 Real-World Evaluation of AiraXiv4 AiraXiv 的实际评估

AiraXiv served as the official infrastructure for the 1st International Conference on AI Scientists (ICAIS 2025)55 5 https://icais.ai, a venue focused on exploring the frontiers of automated scientific discovery with AI scientists. This high-stakes, in-the-wild deployment enables us to evaluate AiraXiv through real-world interactions with the research community. Each submission was evaluated by three AI reviewers, including two model-based reviewers Zhu et al. (2025); Xin et al. (2026) and one agent-based reviewer Shao et al. (2025), with a human expert reviewer providing the final decision.AiraXiv 作为第一届国际 AI 科学家会议（ICAIS 2025）5 的官方基础设施，该会议专注于探索 AI 科学家进行自动科学发现的前沿。这种高风险的实地部署使我们能够通过与研究社区的真实互动来评估 AiraXiv。每份投稿均由三位 AI 评审员评估，包括两位基于模型的评审员 Zhu et al. (2025); Xin et al. (2026) 和一位基于智能体的评审员 Shao et al. (2025)，并由一位人类专家评审员做出最终决定。

Refer to caption
Figure 4: Results from the in-the-wild deployment of AiraXiv. (a) Submission Landscape: Submissions span diverse research topics and application domains, illustrating AiraXiv’s ability to support heterogeneous scientific content. (b) AI–Human Alignment: AI-generated evaluations show a positive correlation with final expert decisions, suggesting the effectiveness of AiraXiv’s reviewing mechanism. (c) Resubmission Improvement: Manuscripts generally receive higher AI scores after resubmission, suggesting that AiraXiv helps authors iteratively improve their manuscripts through AI-generated feedback. 图 4：AiraXiv 实地部署结果。 (a) 投稿概况：投稿涵盖了多元化的研究主题和应用领域，展示了 AiraXiv 支持异构科学内容的能力。 (b) AI 与人类的一致性：AI 生成的评估结果与人类专家的最终决定呈正相关，表明了 AiraXiv 评审机制的有效性。 (c) 重投改进：手稿在重投后通常会获得更高的 AI 评分，表明 AiraXiv 通过 AI 生成的反馈帮助作者迭代改进了手稿。

4.1 Submission Landscape4.1 投稿概况

Across the ICAIS 2025 cycle, AiraXiv processed 114 final submissions in a hybrid ecosystem (82 AI-generated, 32 human-written) and sustained rigorous standards with an overall acceptance rate of 36.8% (42/114). The AI-generated track had a markedly lower acceptance rate (31.7%) than the human-written track (50.0%), indicating that AiraXiv functioned as an active quality filter rather than a passive conduit. Beyond volume and rigor, the submissions exhibited substantial thematic breadth. As shown in Figure 4(a), “applied ML” was the most active category (36 submissions), while higher-quality submissions were concentrated in emerging areas such as “Agent System” and “LLMs” (acceptance ratios 5/8 and 7/8). Domain-specific tracks including “BioAI”, “Healthcare”, and “Finance” further demonstrate AiraXiv’s ability to support multidisciplinary research across diverse application domains.在 ICAIS 2025 周期内，AiraXiv 在混合生态系统中处理了 114 份最终投稿（82 份 AI 生成，32 份人类撰写），并以 36.8%（42/114）的总体接收率保持了严格的标准。AI 生成赛道的接收率（31.7%）明显低于人类撰写赛道（50.0%），表明 AiraXiv 发挥了主动质量过滤器的作用，而非被动的传输渠道。除了数量和严谨性，投稿还表现出广泛的主题覆盖。如图 4(a) 所示，“应用机器学习”是最活跃的类别（36 份投稿），而高质量投稿集中在“智能体系统”和“大语言模型”等新兴领域（接收率分别为 5/8 和 7/8）。包括“生物 AI”、“医疗保健”和“金融”在内的特定领域赛道进一步证明了 AiraXiv 支持跨不同应用领域的跨学科研究的能力。

4.2 AI Review Quality4.2 AI 评审质量

An important question is whether AI reviewers can align with expert human judgment. As shown in Figure 4(b), accepted and rejected papers by human experts exhibit a clear separation in AI scores, with accepted papers achieving higher median scores (>4.5) and rejected papers clustering around a median of approximately 3.0. Consistent with this observation, there is a positive correlation between the aggregated AI scores and the final human decisions (Pearson r=0.43r=0.43, p<0.001p<0.001). Moreover, AI scores demonstrated predictive utility, achieving an AUC of 0.78 in distinguishing accepted manuscripts from rejected ones. In addition, AI-generated submissions received lower average AI scores than human-written submissions in both accepted and rejected groups, suggesting that AI reviewers did not systematically favor AI-generated writing.一个重要的问题是 AI 评审员能否与人类专家的判断保持一致。如图 4(b) 所示，人类专家接收和拒绝的论文在 AI 评分上表现出明显的差异，接收论文获得了更高的中位数评分（>4.5），而拒绝论文的中位数评分则聚集在 3.0 左右。与此观察结果一致，汇总的 AI 评分与最终人类决定之间存在正相关性（皮尔逊相关系数 r=0.43，p<0.001）。此外，AI 评分表现出预测效用，在区分接收手稿与拒绝手稿方面达到了 0.78 的 AUC。此外，在接收和拒绝组中，AI 生成的投稿获得的平均 AI 评分均低于人类撰写的投稿，这表明 AI 评审员并未系统性地偏袒 AI 生成的写作。

4.3 Conference Organization Efficiency4.3 会议组织效率

AiraXiv is designed to support iterative manuscript refinement through rapid AI-assisted feedback. Our data suggests that AiraXiv successfully enabled an effective feedback loop for authors.AiraXiv 旨在通过快速的 AI 辅助反馈支持手稿的迭代优化。我们的数据表明，AiraXiv 成功地为作者实现了有效的反馈闭环。

Rapid Feedback.快速反馈。

The average turnaround time for complete AI review reports was approximately 10.3 hours, which is a substantial reduction compared to the multi-month review cycles typical of traditional venues. This rapid feedback mechanism encouraged authors to engage in iterative improvement during the conference period rather than treating submission as a one-shot process.完整的 AI 评审报告的平均周转时间约为 10.3 小时，与传统会议通常长达数月的评审周期相比，这是一个实质性的缩短。这种快速反馈机制鼓励作者在会议期间进行迭代改进，而不是将投稿视为一次性过程。

Quality Evolution.质量演进。

The rapid feedback incentivized iteration. Approximately 19.3% of submissions underwent version updates during the short conference window. Crucially, this iteration correlated with quality improvement. As shown in Figure 4(c), resubmitted versions of manuscripts generally achieved higher median AI scores compared to initial submissions, suggesting that authors were able to effectively interpret and act upon the automated feedback to refine their work.快速反馈激励了迭代。约 19.3% 的投稿在较短的会议窗口内进行了版本更新。至关重要的是，这种迭代与质量提升相关。如图 4(c) 所示，与初始投稿相比，手稿的重投版本通常获得了更高的 AI 评分中位数，这表明作者能够有效地解读并根据自动化反馈采取行动来优化其工作。

5 Conclusion5 结论

We introduced AiraXiv, an AI-driven, open-access preprint platform for iterative scholarly communication by both human and AI scientists. It integrates paper understanding, pluggable AI reviewing with quality signals, scalable retrieval and recommendation, and lightweight conference organization via web-based and MCP interfaces. In its deployment as the official infrastructure for ICAIS 2025, AiraXiv enabled rapid feedback cycles and measurable quality evolution. These results suggest that AI-assisted, community-feedback publishing can reduce review bottlenecks, accelerate iteration, and broaden valuable scientific outputs. Challenges remain in scaling across diverse fields, strengthening governance and auditability, and defending against adversarial content. We hope AiraXiv serves as a foundation for an open, fast, and inclusive scientific ecosystem in the AI era, while continuing to evolve through future community participation, system extension, and broader research collaboration.我们介绍了 AiraXiv，这是一个 AI 驱动的开放获取预印本平台，旨在供人类和 AI 科学家进行迭代式学术交流。它通过 Web 和 MCP 接口集成了论文理解、带有质量信号的可插拔 AI 评审、可扩展检索与推荐以及轻量级会议组织功能。在作为 ICAIS 2025 官方基础设施的部署中，AiraXiv 实现了快速反馈周期和可衡量的质量演进。这些结果表明，AI 辅助的社区反馈式出版可以减少评审瓶颈，加速迭代，并扩大有价值的科学产出。在跨领域扩展、加强治理和可审计性以及防御对抗性内容方面，挑战依然存在。我们希望 AiraXiv 能成为 AI 时代开放、快速且包容的科学生态系统的基石，同时通过未来的社区参与、系统扩展和更广泛的研究合作不断演进。

Limitations局限性

First, AI-assisted reviewing signals in our work are imperfect and may be biased or unstable across domains, which can mislead readers if interpreted as final judgments. In addition, large-scale AI reviewing may introduce substantial computational overhead as submission volumes grow. We expect advances in LLMs and AI reviewing technologies to continuously improve both review quality and inference efficiency, making such systems more robust and efficient over time. Second, the end-to-end author-reader feedback loop may be vulnerable to low-quality, adversarial, or coordinated feedback, which requires robust moderation and abuse prevention mechanisms. Third, we have only validated AiraXiv in limited real-world deployments, and broader evidence is needed to understand long-term community dynamics, incentive alignment, and the generalization of our work.首先，我们的工作中的 AI 辅助评审信号并不完美，且在不同领域可能存在偏见或不稳定，如果将其解读为最终判断，可能会误导读者。此外，随着投稿量增长，大规模 AI 评审可能会引入巨大的计算开销。我们预计大语言模型和 AI 评审技术的进步将不断提高评审质量和推理效率，使此类系统随着时间的推移变得更加稳健和高效。其次，端到端的作者-读者反馈闭环可能容易受到低质量、对抗性或协调性反馈的影响，这需要强大的审核和防滥用机制。第三，我们仅在有限的实际部署中验证了 AiraXiv，需要更广泛的证据来理解长期的社区动态、激励一致性以及我们工作的泛化能力。

Ethical Considerations伦理考量

Deploying AiraXiv as an open, AI-augmented publishing platform raises important ethical considerations regarding reliability, fairness, and responsible human–AI collaboration.将 AiraXiv 部署为一个开放的、AI 增强型出版平台，在可靠性、公平性和负责任的人机协作方面提出了重要的伦理考量。

The platform raises ethical challenges concerning reliability, safety, and governance in human–AI collaboration. Although AI-generated reviews function as advisory quality signals rather than gatekeeping decisions, model outputs remain probabilistic and may exhibit hallucinations, instability, or overconfidence. To mitigate epistemic risks, AiraXiv explicitly separates AI-generated feedback from human judgment and provides structured, interpretable evaluations to discourage overreliance on automated scores. In formal settings such as ICAIS 2025, final decisions remain under human expert review and oversight.该平台在人机协作的可靠性、安全性和治理方面提出了伦理挑战。虽然 AI 生成的评审作为咨询性质量信号而非把关决策，但模型输出仍然是概率性的，并可能表现出幻觉、不稳定或过度自信。为了降低认知风险，AiraXiv 明确将 AI 生成的反馈与人类判断分离开来，并提供结构化、可解释的评估，以抑制对自动化评分的过度依赖。在 ICAIS 2025 等正式场合，最终决定仍由人类专家评审和监督。

To address adversarial and integrity risks, AiraXiv adopts a decoupled evaluation framework that separates technical utility signals from safety and robustness assessments. By integrating heterogeneous reviewers through a pluggable architecture, the system reduces single-model bias and limits the risk that polished but manipulated submissions receive inflated evaluations. In the future, AiraXiv will incorporate adversarial detection, behavioral monitoring, and scalable human oversight to help mitigate risks such as prompt injection, coordinated abuse, and large-scale feedback distortions.为了应对对抗性和诚信风险，AiraXiv 采用了去耦评估框架，将技术效用信号与安全性和稳健性评估分离开来。通过通过可插拔架构集成异构评审员，系统降低了单模型偏见，并限制了润色过但经过操纵的投稿获得虚高评估的风险。未来，AiraXiv 将纳入对抗检测、行为监控和可扩展的人类监督，以帮助缓解提示词注入、协同滥用和大规模反馈失真等风险。

The iterative publishing paradigm introduces concerns regarding accountability, citation stability, and version control. AiraXiv preserves transparent version histories and immutable records to ensure traceable manuscript evolution. AiraXiv will strengthen provenance tracking mechanisms and require explicit disclosure of AI involvement in authorship.迭代出版范式引发了关于问责制、引用稳定性和版本控制的担忧。AiraXiv 保留透明的版本历史和不可篡改的记录，以确保手稿演进的可追溯性。AiraXiv 将加强来源追踪机制，并要求明确披露 AI 在署名中的参与情况。

Finally, as AI-mediated signals influence research visibility and discovery, risks of bias amplification, topic clustering, and information silos must be carefully managed. AiraXiv will continuously audit reviewer behavior, model diversity, and recommendation mechanisms to promote fairness, inclusivity, and long-term integrity in AI-accelerated scientific publishing.最后，由于 AI 中介信号会影响研究的可见度和发现，必须仔细管理偏见放大、主题聚类和信息茧房的风险。AiraXiv 将持续审计评审员行为、模型多样性和推荐机制，以促进 AI 加速科学出版中的公平性、包容性和长期诚信。

Acknowledgments致谢

We sincerely thank all the organizers and participants of ICAIS 2025. This work was supported by the Research Program No. WU2023C020 of Research Center for Industries of the Future, Westlake University.我们衷心感谢 ICAIS 2025 的所有组织者和参会者。本工作得到了西湖大学未来产业研究中心研究项目 No. WU2023C020 的支持。

References参考文献

  • alphaXiv (2025) alphaXiv AlphaXiv. Note: https://www.alphaxiv.orgAccessed: 2025-04 Cited by: §2. alphaXiv (2025) alphaXiv AlphaXiv. 注：https://www.alphaxiv.org 访问日期：2025-04 引用自：§2。
  • Analemma (2026) Analemma Introducing fars — fully automated research system. Note: https://analemma.ai/blog/introducing-farsAccessed: 2026-02 Cited by: §1. Analemma (2026) Analemma 介绍 fars — 全自动研究系统。 注：https://analemma.ai/blog/introducing-fars 访问日期：2026-02 引用自：§1。
  • Autoscience (2025) Autoscience Carl technical report. Autoscience Blog. External Links: Link Cited by: §1, §2. Autoscience (2025) Autoscience Carl 技术报告。 Autoscience 博客。 外部链接：链接 引用自：§1，§2。
  • Azad and Banu (2024) A. Azad and A. Banu Publication trends in artificial intelligence conferences: the rise of super prolific authors. arXiv preprint arXiv:2412.07793. Cited by: §1. Azad and Banu (2024) A. Azad and A. Banu 人工智能会议中的出版趋势：超级多产作者的崛起。 arXiv 预印本 arXiv:2412.07793。 引用自：§1。
  • Baek et al. (2025) J. Baek, S. K. Jauhar, S. Cucerzan, and S. J. Hwang Researchagent: iterative research idea generation over scientific literature with large language models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 6709–6738. Cited by: §2. Baek et al. (2025) J. Baek, S. K. Jauhar, S. Cucerzan, and S. J. Hwang Researchagent：利用大语言模型对科学文献进行迭代式研究构思生成。 收录于：2025 年国际计算语言学协会美洲分会会议论文集（第 1 卷：长论文）， 第 6709–6738 页。 引用自：§2。
  • BaHammam (2025) A. S. BaHammam Peer review in the artificial intelligence era: a call for developing responsible integration guidelines. Taylor & Francis. Cited by: §1. BaHammam (2025) A. S. BaHammam 人工智能时代的同行评审：呼吁制定负责任的整合指南。 Taylor & Francis。 引用自：§1。
  • Boiko et al. (2023) D. A. Boiko, R. MacKnight, B. Kline, and G. Gomes Autonomous chemical research with large language models. Nature 624 (7992), pp. 570–578. Cited by: §2. Boiko et al. (2023) D. A. Boiko, R. MacKnight, B. Kline, and G. Gomes 利用大语言模型进行自主化学研究。 Nature 624 (7992), 第 570–578 页。 引用自：§2。
  • Brown et al. (2020) T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei Language models are few-shot learners. Advances in neural information processing systems 33, pp. 1877–1901. Cited by: §1. Brown et al. (2020) T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei 语言模型是少样本学习者。 神经信息处理系统进展 33，第 1877–1901 页。 引用自：§1。
  • Chen et al. (2025) N. Chen, M. Duan, A. H. Lin, Q. Wang, J. Wu, and B. He Position: the current ai conference model is unsustainable! diagnosing the crisis of centralized ai conference. arXiv preprint arXiv:2508.04586. Cited by: §1. Chen et al. (2025) N. Chen, M. Duan, A. H. Lin, Q. Wang, J. Wu, and B. He 立场：当前的 AI 会议模式是不可持续的！诊断中心化 AI 会议的危机。 arXiv 预印本 arXiv:2508.04586。 引用自：§1。
  • Cohan et al. (2020) A. Cohan, S. Feldman, I. Beltagy, D. Downey, and D. S. Weld SPECTER: document-level representation learning using citation-informed transformers. ArXiv abs/2004.07180. Cited by: §2. Cohan et al. (2020) A. Cohan, S. Feldman, I. Beltagy, D. Downey, and D. S. Weld SPECTER：使用引文感知 Transformer 进行文档级表征学习。 ArXiv abs/2004.07180。 引用自：§2。
  • Geng et al. (2025) M. Geng, C. Chen, Y. Wu, Y. Wan, P. Zhou, and D. Chen The impact of large language models in academia: from writing to speaking. In Findings of the Association for Computational Linguistics: ACL 2025, pp. 19303–19319. Cited by: §1, §2. Geng et al. (2025) M. Geng, C. Chen, Y. Wu, Y. Wan, P. Zhou, and D. Chen 大语言模型在学术界的影响：从写作到演讲。 收录于：计算语言学协会研究结果：ACL 2025， 第 19303–19319 页。 引用自：§1，§2。
  • Ginsparg (2011) P. Ginsparg ArXiv at 20. Nature 476 (7359), pp. 145–147. Cited by: §2. Ginsparg (2011) P. Ginsparg arXiv 20 周年。 Nature 476 (7359), 第 145–147 页。 引用自：§2。
  • Hanson et al. (2024) M. A. Hanson, P. G. Barreiro, P. Crosetto, and D. Brockington The strain on scientific publishing. Quantitative Science Studies 5 (4), pp. 823–843. Cited by: §1. Hanson et al. (2024) M. A. Hanson, P. G. Barreiro, P. Crosetto, and D. Brockington 科学出版的压力。 定量科学研究 5 (4), 第 823–843 页。 引用自：§1。
  • He et al. (2025) Y. He, G. Huang, P. Feng, Y. Lin, Y. Zhang, H. Li, et al. Pasa: an llm agent for comprehensive academic paper search. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 11663–11679. Cited by: §2. He et al. (2025) Y. He, G. Huang, P. Feng, Y. Lin, Y. Zhang, H. Li, et al. Pasa：用于综合学术论文搜索的 LLM 智能体。 收录于：第 63 届国际计算语言学协会年会论文集（第 1 卷：长论文）， 第 11663–11679 页。 引用自：§2。
  • Hu et al. (2024) Y. Hu, Z. Li, Z. Zhang, C. Ling, R. Kanjiani, B. Zhao, and L. Zhao Hireview: hierarchical taxonomy-driven automatic literature review generation. Cited by: §2. Hu et al. (2024) Y. Hu, Z. Li, Z. Zhang, C. Ling, R. Kanjiani, B. Zhao, and L. Zhao Hireview：基于分层分类法的自动文献综述生成。 引用自：§2。
  • Intology (2025) Intology Zochi technical report. Intology Blog. External Links: Link Cited by: §1, §2. Intology (2025) Intology Zochi 技术报告。 Intology 博客。 外部链接：链接 引用自：§1，§2。
  • Jiang and Ng (2025) Y. Jiang and A. Ng Stanford agentic reviewer: technical overview. Note: https://paperreview.ai/tech-overviewStanford Machine Learning Group and Landing AI Cited by: §C.2, §2, §3.2.2. Jiang and Ng (2025) Y. Jiang and A. Ng 斯坦福智能体评审员：技术概览。 注：https://paperreview.ai/tech-overview 斯坦福机器学习小组和 Landing AI 引用自：§C.2，§2，§3.2.2。
  • Khalifa and Albadawy (2024) M. Khalifa and M. Albadawy Using artificial intelligence in academic writing and research: an essential productivity tool. Computer Methods and Programs in Biomedicine Update 5, pp. 100145. Cited by: §1. Khalifa and Albadawy (2024) M. Khalifa and M. Albadawy 在学术写作和研究中使用人工智能：一种必要的生产力工具。 计算机方法与程序生物医学更新 5，第 100145 页。 引用自：§1。
  • Kiessling et al. (2016) L. L. Kiessling, L. E. Fernandez, A. P. Alivisatos, and P. S. Weiss ChemRXiv: a chemistry preprint server. Vol. 10, ACS Publications. Cited by: §2. Kiessling et al. (2016) L. L. Kiessling, L. E. Fernandez, A. P. Alivisatos, and P. S. Weiss ChemRXiv：化学预印本服务器。 第 10 卷，ACS 出版社。 引用自：§2。
  • Kusumegi et al. (2025) K. Kusumegi, X. Yang, P. Ginsparg, M. de Vaan, T. Stuart, and Y. Yin Scientific production in the era of large language models. Science 390 (6779), pp. 1240–1243. Cited by: §1. Kusumegi et al. (2025) K. Kusumegi, X. Yang, P. Ginsparg, M. de Vaan, T. Stuart, and Y. Yin 大语言模型时代的科学产出。 Science 390 (6779), 第 1240–1243 页。 引用自：§1。
  • LangTaoSha (2025) LangTaoSha LangTaoSha preprint server. Note: https://langtaosha.org.cnAccessed: 2025-11 Cited by: §2. LangTaoSha (2025) LangTaoSha 浪淘沙预印本服务器。 注：https://langtaosha.org.cn 访问日期：2025-11 引用自：§2。
  • Liang et al. (2024a) W. Liang, Z. Izzo, Y. Zhang, H. Lepp, H. Cao, X. Zhao, L. Chen, H. Ye, S. Liu, Z. Huang, D. A. McFarland, and J. Y. Zou Monitoring ai-modified content at scale: a case study on the impact of chatgpt on ai conference peer reviews. In International Conference on Machine Learning, pp. 29575–29620. Cited by: §1. Liang et al. (2024a) W. Liang, Z. Izzo, Y. Zhang, H. Lepp, H. Cao, X. Zhao, L. Chen, H. Ye, S. Liu, Z. Huang, D. A. McFarland, and J. Y. Zou 大规模监控 AI 修改的内容：关于 ChatGPT 对 AI 会议同行评审影响的案例研究。 收录于：国际机器学习会议， 第 29575–29620 页。 引用自：§1。
  • Liang et al. (2024b) W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, D. Yang, C. Potts, C. D. Manning, and J. Y. Zou Mapping the increasing use of LLMs in scientific papers. In First Conference on Language Modeling, Cited by: §2. Liang et al. (2024b) W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, D. Yang, C. Potts, C. D. Manning, and J. Y. Zou 映射科学论文中 LLM 使用量的增长。 收录于：第一届语言建模会议， 引用自：§2。
  • Liang et al. (2025a) W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, D. Yang, C. Potts, C. D. Manning, and J. Zou Quantifying large language model usage in scientific papers. Nature Human Behaviour 9 (12), pp. 2599–2609. Cited by: §1, §2. Liang et al. (2025a) W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, D. Yang, C. Potts, C. D. Manning, and J. Zou 量化科学论文中的大语言模型使用情况。 Nature Human Behaviour 9 (12), 第 2599–2609 页。 引用自：§1，§2。
  • Liang et al. (2024c) W. Liang, Y. Zhang, H. Cao, B. Wang, D. Y. Ding, X. Yang, K. Vodrahalli, S. He, D. S. Smith, Y. Yin, D. A. McFarland, and J. Zou Can large language models provide useful feedback on research papers? a large-scale empirical analysis. NEJM AI 1 (8), pp. AIoa2400196. Cited by: §2. Liang et al. (2024c) W. Liang, Y. Zhang, H. Cao, B. Wang, D. Y. Ding, X. Yang, K. Vodrahalli, S. He, D. S. Smith, Y. Yin, D. A. McFarland, and J. Zou 大语言模型能为研究论文提供有用的反馈吗？一项大规模实证分析。 NEJM AI 1 (8), AIoa2400196。 引用自：§2。
  • Liang et al. (2025b) X. Liang, J. Yang, Y. Wang, C. Tang, Z. Zheng, S. Song, Z. Lin, Y. Yang, S. Niu, H. Wang, B. Tang, F. Xiong, K. Mao, and Z. li Surveyx: academic survey automation via large language models. arXiv preprint arXiv:2502.14776. Cited by: §2. Liang 等 (2025b) X. Liang, J. Yang, Y. Wang, C. Tang, Z. Zheng, S. Song, Z. Lin, Y. Yang, S. Niu, H. Wang, B. Tang, F. Xiong, K. Mao, and Z. li Surveyx: academic survey automation via large language models. arXiv 预印本 arXiv:2502.14776. 引用：§2。
  • Lu et al. (2024) C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha The ai scientist: towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292. Cited by: §2. Lu 等 (2024) C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha The ai scientist: towards fully automated open-ended scientific discovery. arXiv 预印本 arXiv:2408.06292. 引用：§2。
  • M. Bran et al. (2024) A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and P. Schwaller Augmenting large language models with chemistry tools. Nature Machine Intelligence 6 (5), pp. 525–535. Cited by: §2. M. Bran 等 (2024) A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and P. Schwaller Augmenting large language models with chemistry tools. Nature Machine Intelligence 6 (5), pp. 525–535. 引用：§2。
  • Niu et al. (2025) J. Niu, Z. Liu, Z. Gu, B. Wang, L. Ouyang, Z. Zhao, T. Chu, T. He, F. Wu, Q. Zhang, Z. Jin, G. Liang, R. Zhang, W. Zhang, Y. Qu, Z. Ren, Y. Sun, Y. Zheng, D. Ma, Z. Tang, B. Niu, Z. Miao, H. Dong, S. Qian, J. Zhang, J. Chen, F. Wang, X. Zhao, L. Wei, W. Li, S. Wang, R. Xu, Y. Cao, L. Chen, Q. Wu, H. Gu, L. Lu, K. Wang, D. Lin, G. Shen, X. Zhou, L. Zhang, Y. Zang, X. Dong, J. Wang, B. Zhang, L. Bai, P. Chu, W. Li, J. Wu, L. Wu, Z. Li, G. Wang, Z. Tu, C. Xu, K. Chen, Y. Qiao, B. Zhou, D. Lin, W. Zhang, and C. He MinerU2.5: a decoupled vision-language model for efficient high-resolution document parsing. arXiv preprint arXiv:2509.22186. Cited by: §3.2.1. Niu 等 (2025) J. Niu, Z. Liu, Z. Gu, B. Wang, L. Ouyang, Z. Zhao, T. Chu, T. He, F. Wu, Q. Zhang, Z. Jin, G. Liang, R. Zhang, W. Zhang, Y. Qu, Z. Ren, Y. Sun, Y. Zheng, D. Ma, Z. Tang, B. Niu, Z. Miao, H. Dong, S. Qian, J. Zhang, J. Chen, F. Wang, X. Zhao, L. Wei, W. Li, S. Wang, R. Xu, Y. Cao, L. Chen, Q. Wu, H. Gu, L. Lu, K. Wang, D. Lin, G. Shen, X. Zhou, L. Zhang, Y. Zang, X. Dong, J. Wang, B. Zhang, L. Bai, P. Chu, W. Li, J. Wu, L. Wu, Z. Li, G. Wang, Z. Tu, C. Xu, K. Chen, Y. Qiao, B. Zhou, D. Lin, W. Zhang, and C. He MinerU2.5: a decoupled vision-language model for efficient high-resolution document parsing. arXiv 预印本 arXiv:2509.22186. 引用：§3.2.1。
  • OpenAI (2023) OpenAI Gpt-4 technical report. arXiv preprint arXiv:2303.08774. Cited by: §1. OpenAI (2023) OpenAI Gpt-4 technical report. arXiv 预印本 arXiv:2303.08774. 引用：§1。
  • OpenReview (2025) OpenReview OpenReview: an open peer review platform for scholarly communication. Note: https://openreview.netAccessed: 2025-04 Cited by: §2. OpenReview (2025) OpenReview OpenReview: an open peer review platform for scholarly communication. 注：https://openreview.net 访问日期：2025-04 引用：§2。
  • Rawlinson and Bloom (2019) C. Rawlinson and T. Bloom New preprint server for medical research. Vol. 365, British Medical Journal Publishing Group. Cited by: §2. Rawlinson 和 Bloom (2019) C. Rawlinson and T. Bloom New preprint server for medical research. Vol. 365, British Medical Journal Publishing Group. 引用：§2。
  • Schmidgall and Moor (2025) S. Schmidgall and M. Moor Agentrxiv: towards collaborative autonomous research. arXiv preprint arXiv:2503.18102. Cited by: §2. Schmidgall 和 Moor (2025) S. Schmidgall and M. Moor Agentrxiv: towards collaborative autonomous research. arXiv 预印本 arXiv:2503.18102. 引用：§2。
  • Schmidgall et al. (2025) S. Schmidgall, Y. Su, Z. Wang, X. Sun, J. Wu, X. Yu, J. Liu, Z. Liu, and E. Barsoum Agent laboratory: using llm agents as research assistants. arXiv preprint arXiv:2501.04227. Cited by: §2. Schmidgall 等 (2025) S. Schmidgall, Y. Su, Z. Wang, X. Sun, J. Wu, X. Yu, J. Liu, Z. Liu, and E. Barsoum Agent laboratory: using llm agents as research assistants. arXiv 预印本 arXiv:2501.04227. 引用：§2。
  • Sever et al. (2019) R. Sever, T. Roeder, S. Hindle, L. Sussman, K. Black, J. Argentine, W. Manos, and J. R. Inglis BioRxiv: the preprint server for biology. BioRxiv, pp. 833400. Cited by: §2. Sever 等 (2019) R. Sever, T. Roeder, S. Hindle, L. Sussman, K. Black, J. Argentine, W. Manos, and J. R. Inglis BioRxiv: the preprint server for biology. BioRxiv, pp. 833400. 引用：§2。
  • Shao et al. (2025) C. Shao, D. Huang, Y. Li, K. Zhao, W. Lin, Y. Zhang, Q. Zeng, Z. Chen, T. Li, Y. Huang, T. Wu, X. Liu, R. Zhao, M. Zhao, J. Li, X. Zhang, Y. Wang, Y. Zhen, F. Xu, Y. Li, and T. Liu OmniScientist: toward a co-evolving ecosystem of human and ai scientists. arXiv preprint arXiv:2511.16931. Cited by: §1, §2, §3.2.2, §4. Shao 等 (2025) C. Shao, D. Huang, Y. Li, K. Zhao, W. Lin, Y. Zhang, Q. Zeng, Z. Chen, T. Li, Y. Huang, T. Wu, X. Liu, R. Zhao, M. Zhao, J. Li, X. Zhang, Y. Wang, Y. Zhen, F. Xu, Y. Li, and T. Liu OmniScientist: toward a co-evolving ecosystem of human and ai scientists. arXiv 预印本 arXiv:2511.16931. 引用：§1, §2, §3.2.2, §4。
  • Shi et al. (2025) X. Shi, Y. Li, Q. Kou, L. Yu, J. Xie, and H. Zhou Spar: scholar paper retrieval with llm-based agents for enhanced academic search. arXiv preprint arXiv:2507.15245. Cited by: §2. Shi 等 (2025) X. Shi, Y. Li, Q. Kou, L. Yu, J. Xie, and H. Zhou Spar: scholar paper retrieval with llm-based agents for enhanced academic search. arXiv 预印本 arXiv:2507.15245. 引用：§2。
  • Tang et al. (2025) J. Tang, L. Xia, Z. Li, and C. Huang AI-researcher: autonomous scientific innovation. arXiv preprint arXiv:2505.18705. Cited by: §2. Tang 等 (2025) J. Tang, L. Xia, Z. Li, and C. Huang AI-researcher: autonomous scientific innovation. arXiv 预印本 arXiv:2505.18705. 引用：§2。
  • Wang et al. (2024) Y. Wang, Q. Guo, W. Yao, H. Zhang, X. Zhang, Z. Wu, M. Zhang, X. Dai, M. Zhang, Q. Wen, W. Ye, S. Zhang, and Y. Zhang Autosurvey: large language models can automatically write surveys. Advances in neural information processing systems 37, pp. 115119–115145. Cited by: §2. Wang 等 (2024) Y. Wang, Q. Guo, W. Yao, H. Zhang, X. Zhang, Z. Wu, M. Zhang, X. Dai, M. Zhang, Q. Wen, W. Ye, S. Zhang, and Y. Zhang Autosurvey: large language models can automatically write surveys. Advances in neural information processing systems 37, pp. 115119–115145. 引用：§2。
  • Weng et al. (2025a) Y. Weng, M. Zhu, G. Bao, H. Zhang, J. Wang, Y. Zhang, and L. Yang CycleResearcher: improving automated research via automated review. In The Thirteenth International Conference on Learning Representations, Cited by: §2. Weng 等 (2025a) Y. Weng, M. Zhu, G. Bao, H. Zhang, J. Wang, Y. Zhang, and L. Yang CycleResearcher: improving automated research via automated review. In The Thirteenth International Conference on Learning Representations, 引用：§2。
  • Weng et al. (2025b) Y. Weng, M. Zhu, Q. Xie, Q. Sun, Z. Lin, S. Liu, and Y. Zhang Deepscientist: advancing frontier-pushing scientific findings progressively. arXiv preprint arXiv:2509.26603. Cited by: §1, §2. Weng 等 (2025b) Y. Weng, M. Zhu, Q. Xie, Q. Sun, Z. Lin, S. Liu, and Y. Zhang Deepscientist: advancing frontier-pushing scientific findings progressively. arXiv 预印本 arXiv:2509.26603. 引用：§1, §2。
  • Xin et al. (2026) Y. Xin, Y. Weng, M. Zhu, Y. Ling, C. Qin, M. Hahn, M. Backes, Y. Zhang, and L. Yang SafeReview: defending llm-based review systems against adversarial hidden prompts. arXiv preprint arXiv:2604.26506. Cited by: §3.2.2, §4. Xin 等 (2026) Y. Xin, Y. Weng, M. Zhu, Y. Ling, C. Qin, M. Hahn, M. Backes, Y. Zhang, and L. Yang SafeReview: defending llm-based review systems against adversarial hidden prompts. arXiv 预印本 arXiv:2604.26506. 引用：§3.2.2, §4。
  • Yamada et al. (2025) Y. Yamada, R. T. Lange, C. Lu, S. Hu, C. Lu, J. Foerster, J. Clune, and D. Ha The ai scientist-v2: workshop-level automated scientific discovery via agentic tree search. arXiv preprint arXiv:2504.08066. Cited by: §1, §2. Yamada 等 (2025) Y. Yamada, R. T. Lange, C. Lu, S. Hu, C. Lu, J. Foerster, J. Clune, and D. Ha The ai scientist-v2: workshop-level automated scientific discovery via agentic tree search. arXiv 预印本 arXiv:2504.08066. 引用：§1, §2。
  • Yan et al. (2025) X. Yan, S. Feng, J. Yuan, R. Xia, B. Wang, L. Bai, and B. Zhang Surveyforge: on the outline heuristics, memory-driven generation, and multi-dimensional evaluation for automated survey writing. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 12444–12465. Cited by: §2. Yan 等 (2025) X. Yan, S. Feng, J. Yuan, R. Xia, B. Wang, L. Bai, and B. Zhang Surveyforge: on the outline heuristics, memory-driven generation, and multi-dimensional evaluation for automated survey writing. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 12444–12465. 引用：§2。
  • Yang et al. (2023) Z. Yang, Y. Wang, and L. Zhang AI becomes a masterbrain scientist. bioRxiv, pp. 2023–04. Cited by: §3.2. Yang 等 (2023) Z. Yang, Y. Wang, and L. Zhang AI becomes a masterbrain scientist. bioRxiv, pp. 2023–04. 引用：§3.2。
  • Yang et al. (2025) Z. Yang, Q. Zhou, F. Guo, S. Zhang, Y. Xi, J. Nie, Y. Zhu, L. Huang, C. Wu, Y. Xia, X. Ma, Y. Pu, P. Lu, J. Pan, M. Chen, T. Guo, Y. Dou, H. Chen, A. Zeng, J. Huang, T. Xu, and Y. Zhang Airalogy: ai-empowered universal data digitization for research automation. arXiv preprint arXiv:2506.18586. Cited by: §3.2. Yang 等 (2025) Z. Yang, Q. Zhou, F. Guo, S. Zhang, Y. Xi, J. Nie, Y. Zhu, L. Huang, C. Wu, Y. Xia, X. Ma, Y. Pu, P. Lu, J. Pan, M. Chen, T. Guo, Y. Dou, H. Chen, A. Zeng, J. Huang, T. Xu, and Y. Zhang Airalogy: ai-empowered universal data digitization for research automation. arXiv 预印本 arXiv:2506.18586. 引用：§3.2。
  • Yang et al. (2024) Z. Yang, X. Du, J. Li, J. Zheng, S. Poria, and E. Cambria Large language models for automated open-domain scientific hypotheses discovery. In Findings of the Association for Computational Linguistics: ACL 2024, pp. 13545–13565. Cited by: §2. Yang 等 (2024) Z. Yang, X. Du, J. Li, J. Zheng, S. Poria, and E. Cambria Large language models for automated open-domain scientific hypotheses discovery. In Findings of the Association for Computational Linguistics: ACL 2024, pp. 13545–13565. 引用：§2。
  • Zhang et al. (2025a) H. Zhang, H. Cui, Y. Wang, Y. Tian, Q. Guo, C. Wang, J. Wu, C. Song, and Y. Zhang Deep literature survey automation with an iterative workflow. arXiv preprint arXiv:2510.21900. Cited by: §2. Zhang 等 (2025a) H. Zhang, H. Cui, Y. Wang, Y. Tian, Q. Guo, C. Wang, J. Wu, C. Song, and Y. Zhang Deep literature survey automation with an iterative workflow. arXiv 预印本 arXiv:2510.21900. 引用：§2。
  • Zhang et al. (2025b) P. Zhang, X. Hu, G. Huang, Y. Qi, H. Zhang, X. Li, J. Song, J. Luo, Y. Li, S. Yin, C. Dai, E. H. Jiang, X. Zhou, Z. Yin, B. Yuan, J. Dong, G. Su, G. Qiao, H. Tang, A. Du, L. Pan, Z. Lan, and X. Liu AiXiv: a next-generation open access ecosystem for scientific discovery generated by ai scientists. arXiv preprint arXiv:2508.15126. Cited by: §2. Zhang 等 (2025b) P. Zhang, X. Hu, G. Huang, Y. Qi, H. Zhang, X. Li, J. Song, J. Luo, Y. Li, S. Yin, C. Dai, E. H. Jiang, X. Zhou, Z. Yin, B. Yuan, J. Dong, G. Su, G. Qiao, H. Tang, A. Du, L. Pan, Z. Lan, and X. Liu AiXiv: a next-generation open access ecosystem for scientific discovery generated by ai scientists. arXiv 预印本 arXiv:2508.15126. 引用：§2。
  • Zhang et al. (2025c) Y. Zhang, R. Yang, S. Jiao, S. Kang, and J. Han Scientific paper retrieval with llm-guided semantic-based ranking. arXiv preprint arXiv:2505.21815. Cited by: §2. Zhang 等 (2025c) Y. Zhang, R. Yang, S. Jiao, S. Kang, and J. Han Scientific paper retrieval with llm-guided semantic-based ranking. arXiv 预印本 arXiv:2505.21815. 引用：§2。
  • Zhu et al. (2025) M. Zhu, Y. Weng, L. Yang, and Y. Zhang DeepReview: improving llm-based paper review with human-like deep thinking process. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 29330–29355. Cited by: §2, §3.2.2, §4. Zhu 等 (2025) M. Zhu, Y. Weng, L. Yang, and Y. Zhang DeepReview: improving llm-based paper review with human-like deep thinking process. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 29330–29355. 引用：§2, §3.2.2, §4。

Appendix A AiraXiv Web Pages附录 A AiraXiv 网页

Figure 5 and Figure 6 present representative screenshots of the AiraXiv web interface, illustrating the homepage and paper detail page.图 5 和图 6 展示了 AiraXiv 网页界面的代表性截图，说明了主页和论文详情页。

Refer to caption
Figure 5: AiraXiv homepage, supporting user registration and login, paper browsing and submission, and lightweight conference organization.图 5：AiraXiv 主页，支持用户注册和登录、论文浏览与提交，以及轻量级会议组织。
Refer to caption
Figure 6: AiraXiv paper detail page, displaying distilled insights, AI-generated reviews, related paper recommendations, and community discussions.图 6：AiraXiv 论文详情页，显示了提炼出的见解、AI 生成的评审、相关论文推荐以及社区讨论。

Appendix B AiraXiv MCP Server附录 B AiraXiv MCP 服务器

The AiraXiv MCP server is built with the FastMCP66 6 https://github.com/jlowin/fastmcp framework and implements the MCP specification77 7 https://modelcontextprotocol.io. It exposes 13 tools organized into four functional categories such as account management, paper operations, review access, and community engagement.AiraXiv MCP 服务器基于 FastMCP 框架构建，并实现了 MCP 规范。它公开了 13 个工具，分为账户管理、论文操作、评审访问和社区互动四个功能类别。

B.1 Tool CatalogB.1 工具目录

All 13 tools with their parameters and descriptions are shown in Table 1.全部 13 个工具及其参数和描述见表 1。

Table 1: AiraXiv MCP tool catalog. Parameters marked with * are required.表 1：AiraXiv MCP 工具目录。标有 * 的参数为必填项。
Category Tool Description & Key Parameters
Account get_api_key_info Return metadata about the current API key (name, usage count, paper count, owner).
Papers list_papers List the user’s papers with pagination. Params: scope (user / api_key), limit, offset.
get_paper_info Get detailed metadata for a public paper. Params: paper_id*, include_versions.
get_paper_content Extract full text from a paper’s PDF. Params: paper_id*, max_chars.
get_paper_pdf_url Get a publicly accessible PDF download URL. Params: paper_id*.
search_related_papers Query the recommender service for related papers. Params: paper_id, query, top_k.
create_upload Initiate a two-stage PDF upload session. Params: filename, sha256.
complete_upload Finalize an upload and obtain a one-time pdf_file_id. Params: upload_id*, sha256.
submit_paper Submit a new paper for AI review. Params: title*, pdf_url or pdf_file_id*, abstract, author_list, paper_type, research_category.
update_paper Upload a new version of an existing paper. Params: paper_id*, pdf_url or pdf_file_id, title, abstract, author_list, version_notes.
Reviews get_paper_reviews Retrieve AI-generated reviews for a paper. Params: paper_id*.
Community get_paper_comments Get threaded community comments. Params: paper_id*.
submit_paper_comment Post a comment (supports threaded replies). Params: paper_id*, content*, ai_scientist_name, parent_comment_id.

B.2 AI Scientist Interaction LoopB.2 AI 科学家交互循环

Figure 7 illustrates an example interaction loop that an AI scientist follows when using AiraXiv through the MCP server.图 7 展示了 AI 科学家通过 MCP 服务器使用 AiraXiv 时遵循的交互循环示例。

AI Scientist generates papercreate_upload()→\rightarrowupload_id, upload_urlHTTP PUT PDF bytes to upload_urlcomplete_upload(upload_id)→\rightarrowpdf_file_idsubmit_paper(title, pdf_file_id, ...)→\rightarrowpaper_idget_paper_reviews(paper_id)get_paper_comments(paper_id)search_related_papers(paper_id)Revise manuscript based on feedbackcreate_upload()→\rightarrowupload_id, upload_urlHTTP PUT updated PDF to upload_urlcomplete_upload(upload_id)→\rightarrownew_pdf_file_idupdate_paper(paper_id, new_pdf_file_id, version_notes)iterate AI 科学家生成论文 create_upload()→upload_id, upload_url HTTP PUT PDF 字节至 upload_url complete_upload(upload_id)→pdf_file_id submit_paper(title, pdf_file_id, ...)→paper_id get_paper_reviews(paper_id) get_paper_comments(paper_id) search_related_papers(paper_id) 根据反馈修改文稿 create_upload()→upload_id, upload_url HTTP PUT 更新后的 PDF 至 upload_url complete_upload(upload_id)→new_pdf_file_id update_paper(paper_id, new_pdf_file_id, version_notes) 迭代
Figure 7: An example closed-loop publishing workflow in which an AI scientist interacts with AiraXiv through MCP tool calls.图 7：一个闭环发布工作流示例，其中 AI 科学家通过 MCP 工具调用与 AiraXiv 进行交互。

Appendix C More Details of AiraXiv AI Reviewer附录 C AiraXiv AI 评审员的更多细节

C.1 Default ConfigurationC.1 默认配置

The default configuration of the AiraXiv AI Reviewer is as follows. In our reference implementation, Gemini 3 Pro serves as the default backbone LLM, while alternative LLMs can also be used depending on deployment requirements. The temperature is set to 0.1 for general evaluation and 0.3 for creative generation. For the search module, the system issues 5 search queries, retrieves up to 20 candidate papers, filters them using a minimum relevance score threshold of 0.5, and retains at most 10 related papers. In the summarization stage, the system selects the 5 most relevant papers to generate detailed, customized summaries, while the remaining papers are represented by their original abstracts.AiraXiv AI 评审员的默认配置如下。在我们的参考实现中，Gemini 3 Pro 作为默认的主干大语言模型，同时也可以根据部署需求使用其他大语言模型。通用评估的温度系数设为 0.1，创造性生成设为 0.3。对于搜索模块，系统发出 5 条搜索查询，检索最多 20 篇候选论文，使用 0.5 的最低相关性分数阈值进行过滤，并保留最多 10 篇相关论文。在总结阶段，系统选择 5 篇最相关的论文生成详细的定制化总结，其余论文则以其原始摘要呈现。

C.2 Review Score DimensionsC.2 评审评分维度

Following Jiang and Ng (2025), the paper is scored on seven dimensions, as shown in Table 2.参考 Jiang 和 Ng (2025)，论文在七个维度上进行评分，如表 2 所示。

Table 2: Review dimensions and evaluation criteria.表 2：评审维度与评估标准。
Dimension Criteria
Originality Novel contribution, creative approach, advances state-of-the-art
Soundness Appropriate methodology, sufficient baselines, reproducibility
Claims Well Supported Strong evidence, logical arguments, appropriate statistics
Importance Addresses important problem, potential for broad impact
Community Value Practical applicability, follow-up potential, resource contribution
Clarity Clear presentation, well-organized structure, appropriate figures
Prior Work Context Comprehensive literature review, fair comparison, proper citations