用两阶段训练让代码大模型从理解到执行，性能突破新纪录
From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents
- 先无执行地学代码语义，再用真实执行反馈精炼工程能力
- 32B模型在SWE-bench上解决率达62.2%，开源模型中领先
- 仅练Python却能零样本跨语言工作，适合做代码助手的开发者
我们提出SWE-ZERO到SWE-HERO的两阶段监督微调方案，通过蒸馏开源前沿大模型，在SWE-bench上取得当前最佳表现。该流程以进化式精炼替代高资源依赖：(1) SWE-ZERO利用大规模无执行轨迹掌握代码语义与仓库级推理能力；(2) SWE-HERO引入针对性执行反馈，将语义直觉转化为严谨工程流程。实验结果树立了同规模开源模型的新基准。我们发布了基于Qwen3-Coder-480B蒸馏的30万条SWE-ZERO和1.3万条SWE-HERO轨迹数据集，以及基于Qwen2.5-Coder系列的多款代理。值得注意的是，SWE-HERO-32B在SWE-bench Verified上达到62.2%的解决率；尽管仅用Python训练，其在SWE-bench Multilingual上仍实现44.1%的零样本跨语言表现，验证了该范式的强泛化能力。
原文摘要 · Abstract (English)
We introduce SWE-ZERO to SWE-HERO, a two-stage SFT recipe that achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs. Our pipeline replaces resource-heavy dependencies with an evolutionary refinement strategy: (1) SWE-ZERO utilizes large-scale, execution-free trajectories to master code semantics and repository-level reasoning, and (2) SWE-HERO applies targeted, execution-backed refinement to transition these semantic intuitions into rigorous engineering workflows. Our empirical results set a new benchmark for open-source models of comparable size. We release a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-Coder-480B, alongside a suite of agents based on the Qwen2.5-Coder series. Notably, SWE-HERO-32B achieves a 62.2% resolution rate on SWE-bench Verified. Furthermore, despite being trained exclusively on Python, our agents demonstrate robust zero-shot transferability on SWE-bench Multilingual, reaching 44.1% and confirming the paradigm's generalizability across diverse languages.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。