| 梁冬* **,史骁* ****,吕存驰* **,赵晓芳* ** *** *****.基于应用程序接口依赖关系图路径搜索的工作流生成方法[J].高技术通讯(中文),2026,36(1):41~52 |
| 基于应用程序接口依赖关系图路径搜索的工作流生成方法 |
| Workflow generation method with path search on application programming interface dependency graph |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 01. 004 |
| 中文关键词: 工作流生成； 思维链； 路径搜索； 大语言模型幻觉； 应用程序接口依赖关系图 |
| 英文关键词: workflow generation, chain-of-thought, path search, large language model hallucinations, application programming interface dependency graph |
| 基金项目: |
| 作者 | 单位 | | 梁冬* ** | （*中国科学院计算技术研究所北京 100190）
（**中国科学院大学北京 100049）
（***中国科学院大学南京学院南京 211135）
（****中科南京信息高铁研究院南京 211135）
（*****中科苏州智能计算技术研究院苏州 215028） | | 史骁* **** | | | 吕存驰* ** | | | 赵晓芳* ** *** ***** | |
|
| 摘要点击次数: 21 |
| 全文下载次数: 16 |
| 中文摘要: |
| 工作流编程日益成为重要的软件开发模式，智能辅助工作流生成服务能够显著提升用户开发效率。然而，现有的工作流平台仅提供基本的辅助功能，如候选节点推荐或边连接的限制。尽管大语言模型驱动的智能编程服务在代码生成上取得了显著进展，但在工作流生成中仍然面临诸多挑战。本文分析了大语言模型在工作流生成过程中常见的幻觉问题，并提出了一种基于应用程序接口（application programming interface，API）依赖关系图路径搜索的工作流生成方法。通过引导模型沿API依赖关系图搜索生成路径，限制其输出解码空间，有效缓解了模型幻觉现象。此外，本文引入多种路径搜索策略，包括出边路径搜索和入边路径搜索，为工作流生成提供了全新的思维链方法。实验结果表明，本文方法在工作流生成质量上优于先前的思维链技术，并显著降低了模型幻觉率。 |
| 英文摘要: |
| Workflow programming is increasingly becoming an important software development paradigm, and intelligent workflow generation services can offer users a more efficient development process. However, existing workflow development platforms only provide basic assistance, such as candidate node recommendations or edge connection constraints. Although large language model-driven intelligent programming services have made significant progress in code generation, there are still many challenges in the field of workflow generation. This paper analyzes the common hallucination problems of large language models during workflow generation and proposes a workflow generation method based on application programming interface (API) dependency graph-guided path search. By using API dependency graph to guide the path search, the model’s output decoding space is constrained, effectively mitigating hallucinations. Furthermore, we introduce various path search strategies, including outbound and inbound path searches, providing a novel chain-of-thought approach for workflow generation. Experimental results show that the proposed method achieves better performance compared to previous chain-of-thought techniques, and significantly reduces model hallucinations. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
