Recursive Harness Self-Improvement
- source_url: https://arxiv.org/abs/2607.15524
- source_type: paper
- platform: arxiv
- author: Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Matei Zaharia, Yujin Tang
- original_date: 2026-07-17
- added_date: 2026-07-22
- arxiv_id: 2607.15524
- arxiv_categories: cs.LG, cs.AI
- pdf_url: https://arxiv.org/pdf/2607.15524v1
- category: agents
- tags: agent-harness, self-improvement, model-harness-coevolution, trace-quality, rhi, arxiv
- quality_score: 4
摘要（中文）
在 model–harness 共演化视角下，harness 不只是推理时脚手架，也是会生成轨迹、影响未来基座模型训练数据的组件。这催生 harness-in-the-loop 学习：同时为即时 agent 性能与未来训练轨迹质量优化 harness。持续更新厂商内置 scaffold 成本高；作者研究任务特定、轻量、仅需少量迭代的用户侧 harness 优化。提出 Recursive Harness Self-Improvement（RHI）：把 harness 表示为 agent 循环的 prompt 级规范，并用自身修订历史的成对反馈迭代精炼。在覆盖量化金融、机器人与药学等 30 个合成 ML 研究任务上，少量 RHI 迭代即可显著抬高低推理力度 agent 的性能上限，超过对应最大推理设置，并最高节省约 60% 推理成本；增益主要来自更有效的跨 agent 信息流与任务特定上下文管理，而非更长推理轨迹。文末将行为形式化为信息论假设下的隐式优化目标，把 RHI 定位为 model–harness 共演化范式中的实用持续学习算法。
Summary (English)
Under model--harness co-evolution, harnesses are not merely inference-time scaffolds but data-generating components whose execution traces can shape future foundation models. This motivates harness-in-the-loop learning: optimizing harnesses for both immediate agent performance and the quality of traces used for future model training. However, continually updating provider-built scaffolds is costly and labor-intensive. We therefore investigate whether optimizing user-constructed harnesses in a task-specific manner can improve execution-trace quality while remaining computationally lightweight and requiring only a few update iterations. To this end, we introduce Recursive Harness Self-Improvement (RHI), which represents the harness as a prompt-level specification of the agent loop and iteratively refines it using pairwise feedback over its own revision history. Across 30 synthetic machine-learning research tasks spanning quantitative finance, robotics, and pharmacy, a few RHI iterations suffice to substantially raise the performance ceiling of low-reasoning-effort agents, exceeding the corresponding maximum-reasoning-effort setting while reducing inference cost by up to 60%. We show that these gains arise primarily from improved task-specific context management through more effective inter-agent information flow rather than longer reasoning traces. Finally, we formalize this behavior as an information-theoretic hypothesis for RHI's implicit optimization objective, suggesting RHI as a practical algorithm for continual learning within the paradigm of model--harness co-evolution.
One-liner
RHI 以 prompt 级 harness 规范+修订历史成对反馈轻量自改进，抬高低推理 agent 上限并最高省约 60% 推理成本。
原文 / 元数据抓取
Recursive Harness Self-Improvement
作者: Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee, Matei Zaharia, Yujin Tang
原文链接: https://arxiv.org/abs/2607.15524
PDF: https://arxiv.org/pdf/2607.15524v1
发布时间: 2026-07-17
更新时间: 2026-07-17
分类: cs.LG, cs.AI
Under model--harness co-evolution, harnesses are not merely inference-time scaffolds but data-generating components whose execution traces can shape future foundation models. This motivates harness-in-the-loop learning: optimizing harnesses for both immediate agent performance and the quality of traces used for future model training. However, continually updating provider-built scaffolds is costly and labor-intensive. We therefore investigate whether optimizing user-constructed harnesses in a task-specific manner can improve execution-trace quality while remaining computationally lightweight and requiring only a few update iterations. To this end, we introduce Recursive Harness Self-Improvement (RHI), which represents the harness as a prompt-level specification of the agent loop and iteratively refines it using pairwise feedback over its own revision history. Across 30 synthetic machine-learning research tasks spanning quantitative finance, robotics, and pharmacy, a few RHI iterations suffice to substantially raise the performance ceiling of low-reasoning-effort agents, exceeding the corresponding maximum-reasoning-effort setting while reducing inference cost by up to 60%. We show that these gains arise primarily from improved task-specific context management through more effective inter-agent information flow rather than longer reasoning traces. Finally, we formalize this behavior as an information-theoretic hypothesis for RHI's implicit optimization objective, suggesting RHI as a practical algorithm for continual learning within the paradigm of model--harness co-evolution.
Obsidian intake evidence excerpt
该内容文件由 AAIF content-fetcher 根据 active/high-score entry 与 OpenCLI arXiv 元数据补齐。
- entry_id: 4a8a2e27
- title: Recursive Harness Self-Improvement
- source: https://arxiv.org/abs/2607.15524
- existing_summary_zh: 在 modelharness 共演化视角下提出 Recursive Harness Self-Improvement（RHI）：把 harness 当作 prompt 级 agent 循环规范，用自身修订历史的成对反馈轻量迭代优化；在 30 个合成 ML 研究任务上，少量迭代即可抬高低推理力度 agent 上限并超过最大推理设置，同时最高节省约 60% 推理成本，增益主要来自任务特定上下文与跨 agent 信息流管理收录理由：给出可操作的 harness-in-the-loop 学习路径，连接即时性能与未来训练轨迹质量