Paper Instruments于2026年9月23日发布了Paper Office，一套Python包套件，旨在使AI Agent能够安全、准确地操作Word、PowerPoint和Excel文件1。在涵盖5个模型和61个任务的测试中，Paper包取得了92.5%的成功率，相比无定制工具的上游包的80.7%和Anthropic对应Office技能的69.5%有显著提升1。
按文件格式划分，Paper包在Word文档（27个任务）的成功率为96.3%，PowerPoint演示文稿（15个任务）为92.0%，Excel表格（19个任务）为87.4%1。测试涵盖的模型包括Opus 5、GLM 5.3 Flash、GPT 5.6 Sol、Grok 4.6和DeepSeek v4.1 Flash1。
该套件通过改进的API设计和专用工具，大幅减少了Agent直接编辑文件内部结构的必要性1。在Paper Office的运行中，Agent编写原始代码编辑的频率仅为1.6%，而上游无技能包为78.7%，Anthropic技能为50.5%1。
Paper Instruments has released Paper Office, a suite of Python packages designed to enable AI agents to safely and accurately manipulate Word, PowerPoint, and Excel files.1 The toolkit achieved a 92.5% success rate across 61 tasks tested on five different models, substantially outperforming alternative approaches.1
In comparative testing, Paper Office demonstrated clear advantages over existing solutions.1 The upstream package without specialized skills achieved an 80.7% success rate, while Anthropic's corresponding Office skills reached 69.5%.1 Notably, agents using Paper Office required raw code editing of file structures in only 1.6% of cases, compared to 78.7% for the unskilled upstream package and 50.5% for Anthropic's skills implementation.1 This reduction reflects Paper Office's improved API design and dedicated tools that minimize the need for direct file manipulation.
The evaluation encompassed 27 Word document tasks, 15 PowerPoint presentations, and 19 Excel spreadsheet operations across models including Opus 5, GLM 5.3 Flash, GPT 5.6 Sol, Grok 4.6, and DeepSeek v4.1 Flash.1 Performance varied by format, with Word documents achieving 96.3% success, PowerPoint presentations 92.0%, and Excel spreadsheets 87.4%.1 Paper Office was released on September 23, 2026.1
评论
还没有评论，欢迎留下第一条。