Research Papers 论文研究 4h ago Updated 2h ago 更新于 2小时前 47

MetaRoute-Bench: Evaluating Meta-Decision Policies for Agentic Workflow Routing MetaRoute-Bench:评估智能体工作流路由的元决策策略

MetaRoute-Bench is an open, inspectable framework for comparing meta-decision policies in agentic workflow routing under a shared execution model A task-aware compositional policy achieves 79.4% success rate, outperforming static (76.7%), one-shot routing (67.4%), and direct answering (52.9%) The benchmark spans 180 synthetic task profiles, eight routing policies, and 30 paired random seeds, generating 43,200 traces Ablation studies reveal that restricting route composition to one operation and 提出MetaRoute-Bench,一个开放、可检查的基准测试框架,用于在共享执行模型下评估智能体工作流中的元决策策略 任务感知的组合策略在43,200个轨迹中达到79.4%成功率,优于静态策略(76.7%)、一次性路由(67.4%)和直接回答(52.9%) 组合策略相比强静态策略提升2.7个百分点,但伴随4.7%成本增加和6.4%延迟增加 消融实验表明,限制路由组合为单一操作和移除验证机制会导致最大性能损失 主要贡献是可重复的评估方法和路由策略权衡分析,结果来自离线执行模型而非生产部署

62
Hot 热度
74
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • MetaRoute-Bench is an open, inspectable framework for comparing meta-decision policies in agentic workflow routing under a shared execution model
  • A task-aware compositional policy achieves 79.4% success rate, outperforming static (76.7%), one-shot routing (67.4%), and direct answering (52.9%)
  • The benchmark spans 180 synthetic task profiles, eight routing policies, and 30 paired random seeds, generating 43,200 traces
  • Ablation studies reveal that restricting route composition to one operation and removing verification cause the largest performance losses
  • Results come from a seeded offline execution model; the primary contribution is a reproducible evaluation method and routing-policy tradeoff analysis

Why It Matters

Agentic AI systems increasingly rely on meta-decision routing—choosing between answering directly, decomposing tasks, invoking tools, executing code, or delegating—but these decisions are typically embedded in opaque orchestration frameworks and evaluated only through aggregate accuracy. MetaRoute-Bench provides the first open, reproducible benchmark for systematically comparing routing policies under identical execution conditions, enabling practitioners to understand the cost-latency-accuracy tradeoffs that matter in production deployments.

Technical Details

  • Benchmark composition: 180 synthetic task profiles across data analysis, research, and document processing domains; eight distinct routing policies; 30 paired random seeds yielding 43,200 execution traces
  • Routing policies evaluated: task-aware compositional policy, workload-specific static policy, one-shot task routing, and direct answering baseline
  • Performance results: compositional policy achieves 79.4% success vs. 76.7% (static), 67.4% (one-shot), 52.9% (direct); the 2.7-point improvement over static comes at 4.7% higher mean cost and 6.4% higher latency (paired 95% CI: ±2.0 points)
  • Ablation findings: largest performance drops occur when route composition is limited to a single operation and when intermediate verification is removed, highlighting the importance of multi-step composition and self-checking
  • Execution model: seeded offline simulation rather than live deployment; all artifacts (task generation, policies, traces, tests, analysis code) are released for community validation

Industry Insight

  • Agentic workflow routing is a critical but under-evaluated dimension of AI system design; benchmarking routing policies separately from task execution reveals tradeoffs that aggregate accuracy metrics obscure
  • Multi-step route composition and verification are essential for high success rates—simplifying to single-operation routing or removing self-checks produces disproportionate performance degradation
  • The open, reproducible nature of MetaRoute-Bench sets a template for future agentic system evaluation, encouraging the community to move beyond black-box orchestration toward transparent, comparable policy analysis

TL;DR

  • 提出MetaRoute-Bench,一个开放、可检查的基准测试框架,用于在共享执行模型下评估智能体工作流中的元决策策略
  • 任务感知的组合策略在43,200个轨迹中达到79.4%成功率,优于静态策略(76.7%)、一次性路由(67.4%)和直接回答(52.9%)
  • 组合策略相比强静态策略提升2.7个百分点,但伴随4.7%成本增加和6.4%延迟增加
  • 消融实验表明,限制路由组合为单一操作和移除验证机制会导致最大性能损失
  • 主要贡献是可重复的评估方法和路由策略权衡分析,结果来自离线执行模型而非生产部署

为什么值得看

这篇论文为智能体系统的元决策评估提供了标准化的基准测试框架,填补了该领域缺乏统一评估方法的空白。研究揭示了路由策略设计中的关键权衡,为构建高效智能体系统提供了实证依据。

技术解析

  • MetaRoute-Bench初始基准包含180个合成任务配置文件,涵盖数据分析、研究和文档处理三大领域,支持8种路由策略和30个配对随机种子,共生成43,200个执行轨迹
  • 任务感知的组合策略通过整合多步骤决策(分解、工具调用、代码执行、专家委托、验证、恢复)实现79.4%成功率,相比一次性任务路由(67.4%)和直接回答(52.9%)优势显著
  • 消融实验显示路由组合能力和验证机制是性能关键因素:限制为单操作和移除验证会导致最大性能下降
  • 评估基于种子离线执行模型,确保结果可复现;研究团队开源了任务生成、策略实现、轨迹数据、测试代码和分析工具

行业启示

  • 智能体系统的元决策策略设计需要在准确率、成本和延迟之间进行权衡,组合策略虽提升2.7%准确率但增加约5-6%开销
  • 验证机制和路由组合能力是智能体可靠性的核心要素,架构设计应优先考虑这些组件
  • 标准化的基准测试框架有助于推动该领域的可比研究,建议从业者关注后续在线系统验证结果

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。

Agent Agent Evaluation 评测 Benchmark 基准测试 LLM 大模型 Research 科学研究