MindForge is an automated pipeline for turning open-source command-line programs into source-free software-engineering environments that expose only a compiled reference executable and documentation. The authors use GLM-5.2 as a teacher to generate program-synthesis trajectories, then fine-tune Qwen3.6-27B. On ProgramBench, the model’s average test pass rate rises from 37.98% to 49.51%. The paper also reports consistent gains across seven unseen benchmarks covering repository generation and translation, bug fixing, feature implementation, and multilingual issue resolution, including a 31.00-point gain on RepoZero-C2Rust.
No heat snapshots are available in the last 24 hours.