MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis
Authors: Yihao Chen, Shi Chang, Khaled Chawa, Feng Lin, Boyuan Chen, Shaowei Wang, Ahmed E. Hassan
Published: 2026-07-29
Categories: cs.SE, cs.CL, cs.LG
arXiv: 2607.27146
Abstract
Coding agents have made substantial progress on software engineering tasks that modify existing codebases, including bug fixing and feature implementation. However, constructing a complete program from scratch remains a major challenge: even the frontier models evaluated on ProgramBench fully resolve fewer than 1% of tasks. One obstacle is the lack of scalable training environments for this from-scratch setting, spanning the whole software engineering life cycle. We introduce MindForge, an automated pipeline that converts open-source command-line programs into source-free environments that expose only a compiled reference executable and its documentation. Using MindForge, we construct training environments from repositories disjoint from those in ProgramBench, and curate a high-quality data recipe consisting of program synthesis trajectories using GLM-5.2 as the teacher agent. Fine-tuning Qwen3.6-27B on these trajectories increases its ProgramBench average test pass rate from 37.98% to 49.51%, achieving performance comparable to substantially larger frontier models. Moreover, the fine-tuned model consistently improves over the base model across all seven unseen software engineering benchmarks.
中文概述
这篇论文提出 MindForge，一个将开源命令行程序转换为无源码环境（只暴露编译后的参考可执行文件和文档）的自动化管线。使用 GLM-5.2 作为教师代理生成程序合成轨迹，微调 Qwen3.6-27B 后 ProgramBench 通过率从 37.98% 提升到 49.51%，达到与更大前沿模型相当的表现。在七个未见的软件工程基准上均有显著提升。
关键发现
- MindForge 将开源 CLI 程序转为无源码环境（只暴露可执行文件 + 文档）
- 使用 GLM-5.2 作为教师代理生成合成轨迹
- 微调 Qwen3.6-27B 后 ProgramBench 通过率 37.98% → 49.51%
- 七个跨场景 SE 基准均有显著提升（RepoZero-C2Rust +31, DeepSWE +14.16）
信息来源
- 论文链接: https://arxiv.org/abs/2607.27146
- PDF: https://arxiv.org/pdf/2607.27146v1
- 主要类别: cs.SE, cs.CL, cs.LG