MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis
MindForge turns open-source CLI programs into training environments where models must rebuild software from docs and a compiled reference, without seeing source.
The paper says this targets from-scratch program construction, a setting where frontier models solve fewer than 1% of ProgramBench tasks. Its pipeline generated source-free environments and used GLM-5.2 trajectories to fine-tune Qwen3.6-27B. The fine-tuned model raised ProgramBench average test pass rate from 37.98% to 49.51%. The authors also report gains across seven unseen software engineering benchmarks, including RepoZero-C2Rust, DeepSWE, NL2Repo-Bench, SWE-bench variants, and FeatBench. ArXiv · AI/CL/LG's note
The paper says this targets from-scratch program construction, a setting where frontier models solve fewer than 1% of ProgramBench tasks. Its pipeline generated source-free environments and used GLM-5.2 trajectories to fine-tune Qwen3.6-27B. The fine-tuned model raised ProgramBench average test pass rate from 37.98% to 49.51%. The authors also report gains across seven unseen software engineering benchmarks, including RepoZero-C2Rust, DeepSWE, NL2Repo-Bench, SWE-bench variants, and FeatBench. ArXiv · AI/CL/LG's note
score 5