DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch
TL;DR. Code agents are moving past localized bug fixes toward building whole repositories from a high-level spec. Training for that is hard because verifiable whole-repo-generation data barely exists. DeNovoSWE is a 4,818-instance dataset where each instance requires generating a complete repository from documentation, built automatically through a sandboxed agentic workflow (no human annotation) using a divide-and-conquer plus critic-repair philosophy and difficulty-aware trajectory filtering. Fine-tuning Qwen3-30B-A3B on it raises the BeyondSWE-Doc2Repo score from 5.8% to 47.2%.
Source: HuggingFace Daily Papers · arxiv 2606.10728
Key findings
- Whole-repo, not patch-level. Each instance is "generate a complete repository from its documentation," a genuinely long-horizon task, versus SWE-bench's single-issue patch contract.
- Automatic construction. A sandboxed agentic workflow curates instances without human labels, using divide-and-conquer decomposition and a critic-repair loop; difficulty-aware filtering balances quality against diversity.
- Large effect. Qwen3-30B-A3B on BeyondSWE-Doc2Repo goes 5.8% → 47.2%, an 8x relative jump, showing the data bottleneck (not the model) was limiting long-horizon SWE.
How this relates to prior wiki knowledge
DeNovoSWE is the data-substrate member of the 06-11 self-improvement cluster alongside Arbor (research over a hypothesis tree) and RACES (composable verified environments). Where RACES manufactures RL environments and Arbor manufactures research strategy, DeNovoSWE manufactures training data for long-horizon coding. All three answer the self-evolving agents page's standing question of where long-horizon training signal comes from. It also pairs with Claw-SWE-Bench (06-11, the harness-aware SWE benchmark): DeNovoSWE builds the training data, Claw-SWE-Bench measures whether the resulting harness uses it well.
→ Raw: raw/huggingface/2026-06-11-denovoswe-scaling-long-horizon-environments-for-generating-e.md