arxiv:2607.27146

MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis

Published on Jul 29

· Submitted by

Boyuan Chen on Jul 30

Centre for Software Excellence

Upvote

Authors:

Boyuan Chen ,

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, as existing environment-construction frameworks focus only on a single phase in software development. To address this gap, 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, spanning long-horizon repository generation and translation, bug fixing, feature implementation, and cross-language issue resolution, with absolute gains of 31.00 points on RepoZero-C2Rust, 14.16 on DeepSWE, 10.70/4.56 on NL2Repo-Bench (with/without tests), 5.04 on SWE-bench Verified, 5.93 on SWE-bench Pro, 5.22 on SWE-bench Multilingual, and 4.94 on FeatBench.

View arXiv page View PDF Add to collection

Community

nemocbb

Paper author Paper submitter about 13 hours ago

We believe this is the first data pipeline that targets at the end-to-end software engineeing cycle.

With this pipeline, we:

🔧 Automatically constructed 562 cleanroom environments across 6 compiled languages
🔧 Collected 1,001 whole-life-cycle trajectories from a teacher agent (GLM-5.2), averaging 181 turns and 177K tokens each
🔧 Refined them into clean supervision via infrastructure-noise recovery and reasoning rewrites

We then fine-tuned Qwen3.6-27B on these trajectories. The result — MindForge-27B:

📈 Jumps from 37.98% to 49.51% on ProgramBench (+30% relative), surpassing DeepSeek V4 Pro and reaching the performance band of GLM-5.1 and Claude Opus 4.7
📈 Outperforms its base model on 152 of 200 tasks — a broad gain, not a few lucky wins
📈 Generalizes to 7 benchmarks it never trained on: +31 points on C→Rust repository translation, a 9× improvement on DeepSWE, +5 on SWE-bench Verified, and more — all statistically significant

Perhaps most interesting: behavior analysis shows the model learned to genuinely engineer. It works roughly twice as long per task while making proportionally fewer errors per command, and nearly doubles the rate at which reasoning and failure recovery turn into actual code edits.