Research Papers 论文研究 19h ago Updated 18h ago 更新于 18小时前 49

Recursive Harness Self-Improvement 递归工具链自改进

Introduces Recursive Harness Self-Improvement (RHI), a method that optimizes agent harnesses via iterative pairwise feedback on their own revision history. Demonstrates that RHI significantly raises the performance ceiling of low-reasoning-effort agents, surpassing maximum-reasoning-effort baselines. Achieves up to a 60% reduction in inference costs while improving execution trace quality for future model training. Identifies improved task-specific context management and inter-agent information 提出递归增强环自改进(RHI)框架,将增强环视为可迭代优化的提示级规范,而非静态推理支架。 通过基于自身修订历史的成对反馈机制,仅需少量迭代即可显著提升低推理能力智能体的性能上限。 在30项跨领域合成任务中,RHI使推理成本降低高达60%,且表现超越最大推理努力设置。 性能提升主要源于更有效的智能体间信息流和任务特定上下文管理,而非延长推理轨迹。 形式化了RHI的信息论隐式优化目标,为模型与增强环协同进化范式下的持续学习提供了实用算法。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces Recursive Harness Self-Improvement (RHI), a method that optimizes agent harnesses via iterative pairwise feedback on their own revision history.
  • Demonstrates that RHI significantly raises the performance ceiling of low-reasoning-effort agents, surpassing maximum-reasoning-effort baselines.
  • Achieves up to a 60% reduction in inference costs while improving execution trace quality for future model training.
  • Identifies improved task-specific context management and inter-agent information flow as the primary drivers of performance gains, rather than extended reasoning traces.

Why It Matters

This research shifts the perspective of AI harnesses from static inference-time scaffolds to dynamic, data-generating components essential for model evolution. By enabling lightweight, user-constructed harness optimization, it offers a scalable path for continual learning and efficient agent deployment without the high costs associated with updating provider-built infrastructure.

Technical Details

  • Methodology: RHI represents the harness as a prompt-level specification of the agent loop and refines it iteratively using pairwise feedback derived from its own revision history.
  • Evaluation Scope: Tested across 30 synthetic machine-learning research tasks covering quantitative finance, robotics, and pharmacy domains.
  • Performance Metrics: The approach allows low-reasoning-effort agents to exceed the performance of maximum-reasoning-effort settings while cutting inference costs by up to 60%.
  • Mechanism Analysis: Gains are attributed to enhanced context management and more effective information flow between agents, formalized through an information-theoretic hypothesis regarding RHI's implicit optimization objective.

Industry Insight

  • Cost-Efficiency Strategy: Organizations should prioritize optimizing agent workflows and harness structures over simply scaling up model reasoning capabilities to achieve better ROI on inference costs.
  • Data-Centric AI: Treating harness execution traces as valuable training data for future model iterations creates a positive feedback loop, making harness design a critical component of long-term model development strategies.
  • Agile Agent Development: The lightweight nature of RHI suggests that user-defined, task-specific harness optimizations can be rapidly deployed and updated, reducing reliance on heavy, centralized infrastructure updates.

TL;DR

  • 提出递归增强环自改进(RHI)框架,将增强环视为可迭代优化的提示级规范,而非静态推理支架。
  • 通过基于自身修订历史的成对反馈机制,仅需少量迭代即可显著提升低推理能力智能体的性能上限。
  • 在30项跨领域合成任务中,RHI使推理成本降低高达60%,且表现超越最大推理努力设置。
  • 性能提升主要源于更有效的智能体间信息流和任务特定上下文管理,而非延长推理轨迹。
  • 形式化了RHI的信息论隐式优化目标,为模型与增强环协同进化范式下的持续学习提供了实用算法。

为什么值得看

本文揭示了从“使用增强环”到“优化增强环以生成高质量训练数据”的范式转变,强调了执行痕迹对基础模型未来进化的影响。对于AI从业者而言,RHI提供了一种低成本、高效率的方法来增强现有智能体性能,同时为构建可持续进化的AI系统提供了理论依据和实践路径。

技术解析

  • 核心机制:RHI将增强环表示为智能体循环的提示级规范,利用成对反馈(pairwise feedback)基于其自身的修订历史进行迭代优化,实现了轻量级的持续改进。
  • 实验验证:在涵盖量化金融、机器人学和制药学的30项合成机器学习研究任务中进行了评估,证明少量迭代即可大幅缩小低推理能力智能体与高推理能力智能体之间的性能差距。
  • 效率优势:通过优化上下文管理和智能体间信息流,RHI在减少高达60%推理成本的同时,实现了比最大推理努力设置更高的性能表现。
  • 理论贡献:文章将RHI的行为形式化为信息论假设,阐明了其隐式优化目标,即最大化执行痕迹的信息价值,从而支持模型与增强环的协同进化。

行业启示

  • 从静态工具到动态组件:企业应重新评估智能体框架中的增强环设计,将其视为可学习和进化的数据生成组件,而不仅仅是推理时的临时支架。
  • 低成本高性能策略:通过优化提示结构和信息流而非单纯增加计算资源或推理步数,可以在显著降低运营成本的同时提升智能体效能,特别是在资源受限场景下。
  • 持续学习的新路径:RHI提供的迭代优化方法为构建能够自我完善、适应新任务的持续学习系统提供了可行方案,有助于解决基础模型更新成本高、周期长的问题。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Agent Agent Training 训练