Multi-step SWE 基准发布，最强模型长链任务解决率仅 18.5%
Multi-step SWE: Can Coding Agents Keep Building on Their Own Code Over Long Horizons?
AI 导读
研究者发布 Multi-step SWE 长程基准，含 69 个软件工程任务，每个任务是在单一代码库上 5-15 步的依赖链，后续步骤须基于智能体此前生成的代码，覆盖四类工程工作和九种语言。九个前沿模型中最强模型解决率仅 18.5%，69 个任务中有 38 个无任何模型解决，基准与代码已公开。
正文 · 原文
Abstract:Developers rarely give a coding agent a single request: they ask it to build a component and then to extend, upgrade, and optimize it, each time on code the agent itself wrote. Most benchmarks, however, score one task at a time. We introduce Multi-step SWE, a long-horizon benchmark of 69 software engineering tasks, each a chain of 5-15 dependent steps on a single codebase. Every step is a complete, multi-hour engineering request that starts from the code the agent produced in the preceding steps, so early design decisions and errors carry forward. The tasks span four families of engineering work and nine languages, include both from-scratch builds and existing open-source codebases, and ground every step in an existing source of truth that defines both the instruction and its verifier. Across nine frontier models, the strongest model reaches a solve rate of only 18.5%, and 38 of the 69 tasks are not solved by any model. We release Multi-step SWE publicly at this https URL.
| Comments: | Preprint. Code and benchmark: this https URL |
| Subjects: | Software Engineering (cs.SE); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2611.07166 [cs.SE] |
| (or arXiv:2611.07166v1 [cs.SE] for this version) | |
| https://doi.org/10.48550/arXiv.2611.07166 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xiangkai Zeng [view email]
[v1]
Wed, 4 Nov 2026 21:49:40 UTC (219 KB)
来源：arXiv cs.AI · arxiv.org