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

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

arXiv自然语言 2026-11-16 22:20 13 阅读 查看原文
用户希望我作为一个专业的新闻排版工程师,将提供的一段纯文本新闻正文排版成语义化的HTML片段。 规则如下: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释。 2. 每个自然段落用`

`包裹,长文用`

`/`

`划分小节。 3. 列表内容用`
  • `,引用用`
    `,代码用`
    `,重要句子可用``。
    4. 保留原文所有信息,不得增删改内容。
    5. 不要输出``/``/``等文档级标签。
    
    分析原文内容:
    原文是一段关于“RepoFactory”的学术论文/技术新闻摘要。
    包含以下信息:
    - Training coding agents for long-horizon tasks requires environments with verifiable feedback.
    - Existing pipelines often follow a linear process that ends with validation, establishing usability but providing little guidance on improving training value.
    - Recursive self-improvement (RSI) often requires model rollouts or model training followed by downstream evaluation, with costs and feedback delays amplified on long-horizon tasks.
    - How can we use timely, reliable proxy signals to improve training value and synthesize environments at scale?
    - We present RepoFactory, an automated pipeline that converts real repositories into verifiable environments where agents build complete projects from scratch given a language specification.
    - Its Environment Evolution Loop optimizes a multiplicative objective combining validity, structural coupling, and implementation scope through proxy feedback.
    - RepoFactory produces 10,159 environments across 17 software domains.
    - Training Qwen3.5-35B-A3B on these environments yields RepoHorizen-35B-A3B, which achieves gains of 22.6 percentage points on NL2Repo-Bench and 8.5 points on average across five benchmarks covering long-horizon coding and agentic tasks in real-world production settings, compared with 17.5 and 6.4 points without environment evolution.
    - These results highlight the training value of RepoFactory environments and the additional gains from the Environment Evolution Loop.
    
    由于原文没有明显的列表、引用、代码,主要是一个长段落。为了满足“长文用`

    `/`

    `划分小节”的规则,我可以将这段文字合理地划分为几个小节,比如: - 背景与挑战 (背景与问题) - 方法与创新 (RepoFactory与环境进化循环) - 实验与结果 (实验与成果) 划分小节时,需要注意“不得增删改内容”,所以小节标题不能是原文没有的,或者如果加了标题,标题本身算不算“增”? 通常,排版工程师添加的小标题属于排版范畴,但为了严格遵守“不得增删改内容”,最好尽量不添加原文没有的文字作为标题,或者使用原文中的关键短语作为标题。 或者,原文本身只有一段,我可以不划分小节,只用`

    `,但规则提到“长文用`

    `/`

    `划分小节”。如果原文不长,可以不用。但为了体现排版能力,我可以用原文中的核心词组作为标题,比如:

    背景与挑战

    (这算增了内容) 为了绝对不增删改内容,我可以直接把原文分成几个`

    `,或者用原文的句子作为标题?不,标题如果直接是原文的句子,会打乱结构。 其实,规则说“长文用`

    `/`

    `划分小节”,如果原文不算长文,可以不用。原文大概180个词,算作一段。 我可以将其拆分为几个`

    `: 段落1:Training coding agents ... long-horizon tasks. 段落2:How can we use ... at scale? 段落3:We present RepoFactory ... proxy feedback. 段落4:RepoFactory produces ... Environment Evolution Loop. 重要句子可以用``,比如“We present RepoFactory...”和最后一句“These results highlight...”。 让我们仔细拆分原文为自然段落: 原文是一整段。为了排版,我可以按逻辑断句,用`

    `包裹。 P1: Training coding agents for long-horizon tasks requires environments with verifiable feedback. Existing pipelines often follow a linear process that ends with validation, establishing usability but providing little guidance on improving training value. Recursive self-improvement (RSI) often requires model rollouts or model training followed by downstream evaluation, with costs and feedback delays amplified on long-horizon tasks. P2: How can we use timely, reliable proxy signals to improve training value and synthesize environments at scale? P3: We present RepoFactory, an automated pipeline that converts real repositories into verifiable environments where agents build complete projects from scratch given a language specification. Its Environment Evolution Loop optimizes a multiplicative objective combining validity, structural coupling, and implementation scope through proxy feedback. P4: RepoFactory produces 10,159 environments across 17 software domains. Training Qwen3.5-35B-A3B on these environments yields RepoHorizen-35B-A3B, which achieves gains of 22.6 percentage points on NL2Repo-Bench and 8.5 points on average across five benchmarks covering long-horizon coding and agentic tasks in real-world production settings, compared with 17.5 and 6.4 points without environment evolution. P5: These results highlight the training value of RepoFactory environments and the additional gains from the Environment Evolution Loop. 这样既满足了每个自然段落用`

    `包裹,又使用了``来强调重要句子。 关于“长文用`

    `/`

    `划分小节”,由于原文较短,不强制使用,或者我可以使用`

    `来划分,但必须确保标题文字不违反“不得增删改内容”。如果添加“背景”、“方法”、“结果”这样的词,属于增加了原文没有的内容。因此,最好不添加额外的标题文字,直接使用`

    `。或者,如果非要用标题,可以用原文中的核心词汇,但这容易被视为增删内容。安全起见,仅使用`

    `和``。 再检查规则: 1. 只输出HTML片段本身,不要Markdown围栏、不要解释; 2. 每个自然段落用

    包裹,长文用

    /

    划分小节; 3. 列表内容用
    • ,引用用
      ,代码用
      ,重要句子可用;
      4. 保留原文所有信息,不得增删改内容;
      5. 不要输出//等文档级标签。
      
      原文没有列表、引用、代码。所以只需`

      `和``。 排版结果:

      Training coding agents for long-horizon tasks requires environments with verifiable feedback. Existing pipelines often follow a linear process that ends with validation, establishing usability but providing little guidance on improving training value. Recursive self-improvement (RSI) often requires model rollouts or model training followed by downstream evaluation, with costs and feedback delays amplified on long-horizon tasks.

      How can we use timely, reliable proxy signals to improve training value and synthesize environments at scale?

      We present RepoFactory, an automated pipeline that converts real repositories into verifiable environments where agents build complete projects from scratch given a language specification. Its Environment Evolution Loop optimizes a multiplicative objective combining validity, structural coupling, and implementation scope through proxy feedback.

      RepoFactory produces 10,159 environments across 17 software domains. Training Qwen3.5-35B-A3B on these environments yields RepoHorizen-35B-A3B, which achieves gains of 22.6 percentage points on NL2Repo-Bench and 8.5 points on average across five benchmarks covering long-horizon coding and agentic tasks in real-world production settings, compared with 17.5 and 6.4 points without environment evolution.

      These results highlight the training value of RepoFactory environments and the additional gains from the Environment Evolution Loop.

      检查是否有增删改内容: 原文: Training coding agents for long-horizon tasks requires environments with verifiable feedback. Existing pipelines often follow a linear process that ends with validation, establishing usability but providing little guidance on improving training value