跳转至

重新思考自进化智能体：我们是否仍然需要既定的优化流程？

文章背景与核心概要

传统的自进化智能体通常依赖于刚性、预先设定的优化流程来处理诸如收集证据、修改持久化工件、选择候选对象以及确定停止准则等任务。随着前沿大语言模型（如 GPT-5.5）作为优化器的能力不断突破，本文深入探讨了当使用前沿模型时，这些特定任务的流程是否依然必不可少。

为了验证这一假设，作者引入了开放式优化（Open-Ended Optimization, OEO）框架。该框架固定了优化目标、允许的交互、资源预算、数据边界和评估机制，同时赋予优化器在线动态组合改进过程的自主权。研究结果表明，由 GPT-5.5 驱动的 OEO 在多项基准测试中击败了传统的预设流程，并且在大幅减少计算资源消耗的同时，展现出了强大的自我改进路径规划能力。

Self-evolving agents traditionally rely on rigid, prescribed optimization pipelines to handle tasks like gathering evidence, revising persistent artifacts, selecting candidates, and determining stopping criteria. This paper questions whether these task-specific procedures remain necessary when a frontier model is utilized as the optimizer.

To explore this, the authors introduce Open-Ended Optimization (OEO)—a framework that fixes the objective, permitted interactions, resource budget, data boundary, and evaluation, while empowering the optimizer to dynamically compose the improvement process online.

Key Findings & Insights:

  • Performance: Driven by GPT-5.5, OEO outperformed prescribed baseline approaches (SkillOpt, a staged pipeline with bounded edits; and GEPA, a reflective evolutionary search) across 14 head-to-head comparisons over 8 benchmark-target-model settings, securing 12 wins, 1 tie, and only 1 narrow loss of 0.21 percentage points.
  • Efficiency: OEO achieved these gains using a median of just 34.3% of the target-interaction token budget configured for SkillOpt.
  • Capability Boundaries: The delegation model has limits—SkillOpt outperforms OEO when paired with a medium-tier optimizer, and weaker optimizers fail to operate through the unchanged OEO interface entirely.
  • Trajectory Analysis: The fully instrumented OEO-SkillOpt comparison reveals that rigid pipelines primarily alter how the optimization process proceeds (the trajectory) rather than fundamentally changing the final behavior.

Ultimately, the research recasts prescribed pipelines as capability-dependent scaffolding: while external constraints remain essential, a sufficiently capable optimizer can successfully forge its own path from measurable feedback to persistent improvement.

摘要

自进化智能体通常围绕预设的优化流程构建：框架决定如何收集证据、修改持久化工件、选择候选对象以及何时停止。我们探讨了当前沿模型充当优化器时，这种特定于任务的程序是否仍然必要。我们引入了开放式优化（OEO），它保持目标、允许的交互、资源预算、数据边界和评估固定，同时允许优化器在线组合改进过程。我们将 OEO 与两种互补的预设方法进行了比较：SkillOpt（具有有界编辑的分阶段管道）和 GEPA（反射式演化搜索）。在 8 个基准-目标模型设置下的 14 次直接比较中，由 GPT-5.5 驱动的 OEO 取得了 12 胜、1 平以及仅以 0.21 个百分点微弱落败 1 次的战绩。它使用的目标交互 Token 预算中位数仅为 SkillOpt 配置预算的 34.3%。一次性、零交互的控制实验表明，这些收益并非由单一的先验驱动重写所解释。然而，任务委托存在能力边界：当配备中等强度的优化器时，SkillOpt 的表现优于 OEO，而较弱的优化器完全无法通过未更改的 OEO 接口进行操作。在完全配备仪器的 OEO-SkillOpt 对比中，轨迹分析进一步表明，预设流程改变优化进行的方式（轨迹）比改变最终行为更为一致。总之，这些发现将预设流程重新定义为依赖于能力的脚手架：外部约束仍然必不可少，但足够强大的优化器可以自主开辟从可测量反馈到持续改进的路线。

Self-evolving agents are usually built around prescribed optimization pipelines: the framework decides how to gather evidence, revise a persistent artifact, select candidates, and stop. We ask whether this task-specific procedure remains necessary when a frontier model acts as the optimizer. We introduce Open-Ended Optimization (OEO), which keeps the objective, permitted interactions, resource budget, data boundary, and evaluation fixed while allowing the optimizer to compose the improvement process online. We compare OEO with two complementary prescribed approaches: SkillOpt, a staged pipeline with bounded edits, and GEPA, a reflective evolutionary search. Across 14 head-to-head comparisons over 8 benchmark-target-model settings, GPT-5.5-driven OEO records 12 wins, 1 tie, and 1 narrow loss of 0.21 percentage points. It uses a median 34.3 percent of SkillOpt's configured target-interaction token budget. A one-shot, zero-interaction control shows that the gains are not explained by a single prior-driven rewrite. However, delegation has a capability boundary: SkillOpt outperforms OEO with a medium optimizer, and a weak optimizer cannot operate through the unchanged OEO interface. In the fully instrumented OEO-SkillOpt pair, trajectory analysis further shows that prescription changes how optimization proceeds more consistently than it changes final behavior. Together, these findings recast prescribed pipelines as capability-dependent scaffolding: essential constraints remain external, but a sufficiently capable optimizer can compose the route from measurable feedback to persistent improvement.

访问论文与资源

获取论文原文及相关开源资源，请访问以下链接：

外部与文献计量工具

External & Bibliographic Tools