DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch

arXiv·medium signal

DeNovoSWE provides a 4,818-instance dataset for whole-repository generation from documentation. Fine-tuning Qwen3-30B-A3B on it lifts performance on the BeyondSWE-Doc2Repo benchmark from 5.8% to 47.2%, a large jump on a long-horizon, doc-to-repo task. Signals real progress toward agents that scaffold complete codebases, not just patches.