跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 5 天前AI 評分48

RepoFactory:以環境演化循環擴展長期編碼智能體的訓練環境

From Repositories to Training Environments: Scaling Long-Horizon Coding Agents with Environment Evolution Loop

AI 導讀

RepoFactory 從真實軟件倉庫建構完整專案開發環境,透過環境演化循環,以驗證執行和程式碼分析提供即時、低成本的代理指標,持續改進任務環境。

正文

Coding agents are taking on longer-horizon, more complex development tasks, and training also needs to cover the complete project development process. 🚀

However, existing environment construction pipelines often stop at execution and validation, lacking further optimization of how much training value the task itself can provide and how it can be improved. 🤔

We hope to introduce feedback-driven iteration from recursive self-improvement (RSI) into environment construction, so that tasks continuously increase their training value during the construction process. 🔄

However, related methods often rely on model rollouts or downstream evaluations after model training to judge whether modifications are effective. For long-horizon coding tasks, a single complete solution involves a large amount of implementation and debugging, further increasing computational cost and feedback cycles. ⏳💰

This prompts us to ask: can we avoid relying on high-cost, long-cycle downstream feedback and instead find reliable, low-cost, immediate proxy metrics to guide environment improvement? 🎯

RepoFactory builds complete project development environments from real software repositories and uses the product of reasonableness, implementation scope, and structural coupling as the environment evolution objective. 🏭

Validation execution and code analysis provide timely, low-cost proxy feedback, guiding the agent to jointly modify task instructions, reference implementations, and tests, thereby ensuring environment validity while increasing the implementation work and component collaboration required by the task. 🔍✅

Each round of modification yields new feedback that continues to guide the next iteration, without waiting for a complete rollout of the solving model or downstream training. 🔁

Ultimately, we constructed 10,159 environments covering 17 software domains. 📊

RepoHorizen-35B-A3B, trained on these environments, improves by 22.6 percentage points on NL2Repo-Bench compared with Qwen3.5-35B-A3B, and by 8.5 percentage points on average across five benchmarks. 📈

Among these, environment evolution brings an additional gain of 5.1 and 2.1 percentage points on these two metrics, respectively. 🎉

We are releasing an initial batch of 200 non-evolved environments and corresponding Docker images. 🐳📦

Friends working on coding agents, training data, and environment construction are welcome to get in touch. 🤝✨

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co