内容来源：https://research.google/blog/improving-the-academic-workflow-introducing-two-ai-agents-for-better-figures-and-peer-review/

内容总结：

谷歌云发布两款AI科研助手 革新论文图表绘制与同行评审流程

2026年4月8日，谷歌云研究科学家Jinsung Yoon与总监Tomas Pfister共同宣布推出两款旨在优化科研工作流程的人工智能体：PaperVizAgent（论文可视化智能体）与ScholarPeer（学者同行评审智能体）。这两款工具旨在将科研人员从繁重的图表制作与初阶评审工作中解放出来，更专注于科学创新本身。

当前，学术研究面临两大普遍挑战：一是如何高效生成符合顶级期刊会议要求的复杂方法论图示与精确统计图表；二是日益增长的论文投稿量使传统同行评审系统承受巨大压力，导致评审疲劳与质量波动。谷歌云此次推出的AI体正是为了应对这些痛点。

PaperVizAgent：一键生成出版级学术图表
该智能体能够根据论文方法章节的文本描述及用户对图表传达意图的说明，自动生成可直接用于出版的学术插图。其核心是一个由五个专业AI体协同工作的系统：检索体搜集相关文献图表作为参考；规划体组织内容；风格体确保输出符合学术审美；可视化体生成图像或统计绘图代码；评审体则对输出进行技术准确性校验并触发迭代优化。评估显示，其在忠实度、简洁性、可读性与美观度四个维度上综合得分（60.2）显著超越现有主流基准工具，并首次突破了设定的人类基线水平（50.0）。

ScholarPeer：模拟资深学者的自动化评审框架
该智能体并非简单进行文本生成，而是模仿资深研究者的审稿流程，构建了一个基于动态网络文献检索与多角度验证的评审系统。它通过“子领域历史学家”智能体建立领域背景叙事，通过“基线侦察员”智能体主动核查作者可能遗漏的对比基线或数据集，并利用多维度问答引擎严格验证论文技术主张。测试表明，其生成的评审意见不仅批判性强、贴合实际，而且深度扎根于现有文献体系，大幅缩小了AI反馈与人类评审在多样性和深度上的差距。

谷歌云强调，这两款工具目前仍为实验性研究原型，其输出成果仅供学术探索参考，不应作为编辑或出版决策的唯一依据。团队展望未来能构建一个深度融合于科研全流程的AI辅助生态系统，持续推动科学交流的质量与效率提升。

（完）

中文翻译：

提升学术工作流效率：两款AI助手助力图表绘制与论文评审
2026年4月8日
Google Cloud研究科学家Jinsung Yoon与总监Tomas Pfister

我们推出两款AI助手以优化科研流程：专攻学术图表绘制的可视化助手PaperVizAgent，以及可对学术论文（含内嵌图表）进行自动化严格评审的ScholarPeer。

快速导览
人工智能的飞速发展正以前所未有的速度推动学术研究演进。严谨的学术工作流远不止构思课题与撰写论文，许多研究者常面临研究成果可视化的挑战。尽管AI能辅助文本创作，但为顶级期刊会议绘制复杂方法论图示与精确统计图表仍具难度。同时，科学界依赖同行评审保障学术成果的严谨性，然而论文数量的激增使评审系统承压，导致审稿疲劳与评价标准不一。随着语言模型与多智能体系统日益成熟，我们认识到它们不仅能作为研究对象，更可成为科研流程的积极参与者。

为此，我们推出两项创新智能体框架：（1）学术图表可视化助手PaperVizAgent（原名PaperBanana）；（2）自动化论文评审助手ScholarPeer。这些助手专为辅助学术研究全周期设计，助力科研人员聚焦创新而非事务性工作。评估显示，PaperVizAgent持续生成专业级图表，显著优于主流基准模型（GPT-Image-1.5、Nano-Banana-Pro、Paper2Any）；而ScholarPeer能提供基于文献的深度批判性评审，超越当前最先进的自动化评审系统。

PaperVizAgent：生成可直接发表的学术图表
PaperVizAgent是从学术文本生成出版级图示的自主框架，通过弥合技术描述与视觉传达的鸿沟，让研究者能直接从手稿创建专业图表。启动流程需研究者提供两项输入：

• 源文本：通常包含研究方法章节的技术细节
• 传达意图：描述图表核心信息的详细说明文字

该框架协调五类专项AI智能体协同工作：（1）检索器（2）规划器（3）风格设计师（4）可视化器（5）评审器。首先，检索器与规划器收集参考文献并组织内容；随后风格设计师合成美学规范以确保符合学术标准；可视化器继而渲染图像或生成统计图表对应的Python代码；最终评审器对照原文评估输出结果，若发现不一致之处则向可视化器提供定向反馈，触发迭代优化循环。通过多轮精修，这一多智能体系统确保最终成果兼具视觉吸引力与技术精确性。

成果评估
在综合实验中，PaperVizAgent在忠实度、简洁性、可读性与美学四大关键维度（采用0-100评分制，分数越高越好）上持续超越主流基准方法（包括直接提示、少样本提示及可视化前沿技术Paper2Any）。评估采用经人工生成图表校准的大语言模型作为评判器，并以50.0分作为人类表现基准线。

PaperVizAgent取得60.2分的综合成绩，显著优于GPT-Image-1.5、Nano-Banana-Pro及Paper2Any等基准模型，更是唯一超越人类基准线的框架。在细分维度中，该系统于简洁性与美学表现尤为突出，统计图表生成能力也达到人类水平。这些成果标志着自动化图示生成领域的重大突破。

ScholarPeer：模拟资深学者评审
ScholarPeer是具备情境感知与搜索能力的多智能体框架，通过模拟资深研究者工作流实现自动化、高水准的论文评审。

区别于将评审视为简单文本生成任务的标准语言模型，ScholarPeer采用情境获取与主动验证的双轨流程：其子领域历史学家智能体动态构建领域叙事，使评审扎根于实时网络文献；基线侦察器作为对抗性审核员，专门核查作者可能遗漏的数据集或对比基准；多维度问答引擎则严格验证论文技术主张，确保形成基于事实的深度批判。最终评审报告包含详细总结、优势分析、缺陷指正及作者答疑环节，完全对标专家级同行评审标准。

在大型公开数据集测试中，ScholarPeer相较于前沿自动化评审方法取得显著胜率。更重要的是，该系统的主动验证机制极大缩小了AI反馈与人类评审多样性之间的差距，生成兼具批判性、真实性且深植文献的评审意见。

科学界的未来展望
PaperVizAgent与ScholarPeer是我们探索AI辅助科研的阶段性成果。通过攻克出版周期中两个独立而关键的环节，这些工具将成为提升科研质量的协作伙伴，与其他技术共同加速知识传播。

尽管现有框架已为学界带来切实效益，但这仅是探索之旅的起点。我们展望未来能构建丰富互联的AI助手生态系统，将其无缝融入科研工作流的每个环节，并持续深耕这一领域。

致谢
感谢Palash Goyal、Dawei Zhu、Mihir Parmar、Rui Meng Yiwen Song、Yale Song、Hamid Palangi、Xiyu Wei、Sujian Li及Burak Gokturk对本研究的宝贵贡献。

免责声明
PaperVizAgent与ScholarPeer均为实验性研究原型，非成熟产品工具。其自动生成的反馈、图表及评审意见仅供科研探索使用，不应作为编辑或出版决策的最终依据。

英文来源：

Improving the academic workflow: Introducing two AI agents for better figures and peer review
April 8, 2026
Jinsung Yoon, Research Scientist, and Tomas Pfister, Director, Google Cloud
Introducing two AI agents to streamline academic research. These include: PaperVizAgent, a visualizer agent for drawing academic figures, and ScholarPeer, a reviewer agent that automatically and rigorously evaluates academic papers.
Quick links
Academic research is evolving at an unprecedented pace driven by the rapid advancements in AI. The academic research workflow is notoriously rigorous, involving far more than just conceptualizing an idea and writing a paper. One hurdle many researchers face is how to effectively visualize their research. While AI can draft text, creating the complex methodology diagrams and precise statistical plots required for top-tier conferences and journals is significantly more difficult. Furthermore, the scientific community relies on the peer review process to maintain the integrity of published research. However, the exponential growth of paper submissions has severely strained this system, leading to reviewer fatigue and inconsistent evaluations. As language models and multi-agent systems become more sophisticated, we see their potential not just as subjects of study, but as active participants in the scientific process itself.
To that end, we introduce two novel agentic frameworks: (i) PaperVizAgent (formally known as PaperBanana), a visualizer agent for drawing academic figures, and (ii) ScholarPeer, a reviewer agent that automatically and rigorously evaluates academic papers, including inlined diagrams). These agents are designed specifically to assist with the academic research lifecycle to empower scientists to focus on innovation rather than administrative overhead. Our evaluations show PaperVizAgent consistently generates expert quality figures that significantly outperform leading baselines (GPT-Image-1.5, Nano-Banana-Pro, Paper2Any) while ScholarPeer delivers highly critical, literature-grounded reviews that beat state-of-the-art automated reviewers.
PaperVizAgent: Generating publication-ready figures
PaperVizAgent is an autonomous framework designed to generate publication-ready academic illustrations from academic text. By bridging the gap between technical descriptions and visual communication, PaperVizAgent allows researchers to create professional-grade figures directly from their manuscripts. To initiate the process, a researcher provides two inputs:

• Source context: Typically the method sections of a manuscript with technical details of the research.
• Communicative intent: A detailed figure caption that describes what the visual should convey.
  The PaperVizAgent framework orchestrates a collaborative team of five specialized AI agents including: (1) a retriever, (2) a planner, (3) a stylist, (4) a visualizer, and (5) a critic. First, the retriever and planner agents gather references (e.g., existing literature to reference relevant academic figures) and organize the content. Next, the stylist agent synthesizes aesthetic guidelines to ensure the output matches academic standards. The visualizer then renders an image or generates executable python code for statistical plots. Finally, the critic agent evaluates the output against the original text. If inconsistencies are found, the critic provides targeted feedback to the visualizer agent, triggering a loop of iterative refinement.Through iterative refinement, this multi-agent system ensures the final illustration is both visually appealing and technically accurate.
  Results
  In comprehensive experiments, PaperVisAgent consistently outperformed leading baselines — including direct prompting, few-shot prompting, and Paper2Any (a state-of-the-art approach for visualization). The system was rigorously evaluated using a comparative scoring metric (using a 0-100 scale, where a higher score is better) across four critical dimensions: faithfulness, conciseness, readability, and aesthetics. In this evaluation, we used an LLM judge that was calibrated using human-generated figures as inputs and a set human performance baseline of 50.0.
  PaperVizAgent achieved an impressive overall score of 60.2, significantly surpassing all evaluated baselines such as GPT-Image-1.5, Nano-Banana-Pro, and Paper2Any. Notably, it stands as the only framework to exceed the established human baseline of 50.0 in its overall rating. When breaking down the specific dimensions, the system particularly excels in Conciseness and Aesthetics, scoring well above the human threshold in both categories. It also achieved human-competitive results in generating statistical plots, proving its versatility. These results represent a significant leap forward in automated illustration.
  Emulating senior reviewers with ScholarPeer
  ScholarPeer is a context-aware, search-enabled multi-agent framework designed to automate and elevate the peer review process by following the workflow of a senior researcher.
  Unlike standard language models that treat reviewing as a simple text-generation task, ScholarPeer relies on a dual-stream process of context acquisition and active verification. It dynamically constructs a domain narrative using a sub-domain historian agent which grounds the review in live, web-scale literature. A baseline scout acts as an adversarial auditor, specifically hunting for datasets or comparative baselines the authors may have missed. Finally, a multi-aspect Q&A engine rigorously verifies the paper's technical claims, ensuring a deep and fact-based critique. The final review report includes a detailed summary, strengths, weaknesses, and questions for the authors, much like a standard expert peer review.
  ScholarPeer's performance demonstrates the immense potential of integrating active web search with multi-agent orchestration for academic evaluation. When tested on the extensive public datasets, ScholarPeer achieved significant win-rates against state-of-the-art automated reviewing approaches in side-by-side evaluations. More importantly, the system's active verification workflow drastically reduced the gap between AI-generated feedback and human-level diversity, producing reviews that are highly critical, realistic, and deeply grounded in existing literature.
  What’s next for the scientific community
  PaperVizAgent and ScholarPeer are part of our broader efforts exploring AI-assisted research more generally. By tackling two distinct but equally demanding phases of the publication lifecycle, these tools serve as collaborators that elevate the quality of scientific discourse and can, alongside other tools, accelerate the dissemination of knowledge.
  While these two frameworks offer immediate and tangible benefits to the academic community, they are just the beginning of our journey. We envision a future where researchers have access to a rich, interconnected ecosystem of AI assistants seamlessly integrated into every facet of the scientific workflow, and we are actively continuing our work in this space.
  Acknowledgements
  We would like to thank Palash Goyal, Dawei Zhu, Mihir Parmar, Rui Meng Yiwen Song, Yale Song, Hamid Palangi, Xiyu Wei, Sujian Li and Burak Gokturk for their valuable contributions to this work.
  Disclaimer
  PaperVizAgent and ScholarPeer are experimental research prototypes, not production-ready tools. Their automated feedback, figures, and reviews are intended only for research exploration and should not be relied upon as the definitive basis for editorial or publication decisions.