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

Existing coding agents are primarily trained to modify or repair existing codebases, leaving the task of constructing complete programs from scratch largely unsolved. The authors introduce MindForge, a pipeline designed to bridge this gap by creating scalable, source-free training environments. These environments provide agents with only a compiled reference executable and its documentation, forcing the model to navigate the entire software development life cycle—from specification inference and architecture design to implementation and debugging—without access to the original source code.

To build these environments, the authors automatically process open-source command-line programs, ensuring they are reproducible, self-contained, and free of source-level leakage. They then use a strong teacher agent (GLM-5.2) to generate 1,001 long-horizon development trajectories. These trajectories are refined through two procedures: infrastructure-noise recovery, which salvages interrupted runs, and a reasoning rewrite mechanism, which repairs incoherent model reflections caused by the removal of erroneous tool calls.

Fine-tuning the Qwen3.6-27B model on these trajectories significantly boosts its performance on ProgramBench, increasing the average test pass rate from 37.98% to 49.51%. This performance allows the 27B-parameter model to outperform larger models like DeepSeek V4 Pro and approach the capabilities of frontier models such as Opus 4.7. Beyond the primary benchmark, the model demonstrates strong generalization across seven unseen software engineering tasks, including repository generation, bug fixing, and cross-language issue resolution, with notable absolute gains in benchmarks like RepoZero-C2Rust and SWE-bench Pro.

This work demonstrates that small language models can achieve high-level software engineering proficiency when trained on high-quality, multi-stage development trajectories rather than just isolated code-completion or bug-fixing tasks. By providing a scalable, source-free pipeline, MindForge offers a robust method for training agents that can handle the complexities of end-to-end program construction, effectively democratizing access to high-performance coding agents that do not require massive parameter counts.