Research Papers 论文研究 1d ago Updated 1d ago 更新于 1天前 49

Think Big, Search Small: Where Capacity Matters in Hierarchical Search Agents? 想大搜小:层级搜索代理中容量何处重要?

Role factorization in hierarchical search agents consistently outperforms single-agent baselines, yielding significant improvements in exact match scores across various model scales. Capacity sensitivity is highly asymmetric; scaling the delegation backbone provides substantial performance gains (~11 EM points), whereas scaling the execution sub-agent offers minimal benefit (~2.6 EM points). A distilled 1.7B-parameter executor achieves parity with frontier sub-agents in accuracy while reducing t 分层搜索代理中,任务分解(Delegation)是能力瓶颈,其模型容量对性能影响远大于执行(Execution)角色。 角色分离架构在六个模型规模下均优于单一代理基线,精确匹配率提升4.5至8.6分。 通过质量过滤轨迹蒸馏训练的1.7B参数执行器,在保持前沿精度的同时减少了37%的子代理Token消耗。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Role factorization in hierarchical search agents consistently outperforms single-agent baselines, yielding significant improvements in exact match scores across various model scales.
  • Capacity sensitivity is highly asymmetric; scaling the delegation backbone provides substantial performance gains (~11 EM points), whereas scaling the execution sub-agent offers minimal benefit (~2.6 EM points).
  • A distilled 1.7B-parameter executor achieves parity with frontier sub-agents in accuracy while reducing token consumption by 37%, optimizing the efficiency-accuracy trade-off.

Why It Matters

This research challenges the common practice of using uniformly scaled models for all agent roles, providing empirical evidence that intelligent resource allocation is critical for performance. It offers a practical blueprint for developers to optimize cost and latency by focusing computational power on task decomposition rather than retrieval execution.

Technical Details

  • Role Factorization: The study decomposes hierarchical search into three distinct roles: delegation (task decomposition), execution (retrieval and evidence extraction), and answer generation (held fixed as a control).
  • Experimental Design: Controlled capacity sweeps were conducted on both delegation and execution axes across five multi-hop QA benchmarks and six different model scales.
  • Key Findings: Delegation capacity was identified as the primary bottleneck, with scaling it driving the majority of performance improvements compared to execution scaling.
  • Optimization Strategy: Quality-filtered trajectory distillation was used to train a smaller 1.7B-parameter executor, demonstrating that smaller models can match larger ones when optimized correctly.

Industry Insight

AI practitioners should prioritize investing in larger, more capable models for the planning and decomposition stages of multi-agent systems, while utilizing smaller, distilled models for execution tasks to maximize efficiency. This asymmetric scaling strategy can significantly reduce operational costs without compromising answer accuracy. Future agent architectures should explicitly separate and optimize these roles rather than relying on homogeneous model deployments.

TL;DR

  • 分层搜索代理中,任务分解(Delegation)是能力瓶颈,其模型容量对性能影响远大于执行(Execution)角色。
  • 角色分离架构在六个模型规模下均优于单一代理基线,精确匹配率提升4.5至8.6分。
  • 通过质量过滤轨迹蒸馏训练的1.7B参数执行器,在保持前沿精度的同时减少了37%的子代理Token消耗。

为什么值得看

该研究揭示了多智能体系统中资源分配的关键不对称性,为构建高效分层搜索代理提供了具体的容量配置指南。它证明了“重规划、轻执行”的策略能在不牺牲准确性的前提下显著降低计算成本,对优化Agent架构具有直接指导意义。

技术解析

  • 角色解耦与实验设计:将分层搜索分解为负责任务分解的“委派角色”、负责检索与证据提取的“执行角色”,以及作为控制变量的“答案生成角色”。在五个多跳QA基准上进行受控容量扫描。
  • 容量敏感性非对称性:实验发现扩大委派骨干模型的容量可使精确匹配(EM)提升约11分,而扩大执行子代理仅提升约2.6分,明确识别出任务分解是当前系统的能力瓶颈。
  • 小模型蒸馏优化:提出使用质量过滤轨迹蒸馏技术训练1.7B参数的执行器,使其精度媲美前沿子代理,同时将子代理Token消耗降低37%,优化了帕累托前沿。

行业启示

  • 架构优化方向:在开发多智能体系统时,应将主要算力集中在顶层的任务规划和分解模块,而非平均分配给所有子代理,以实现性价比最大化。
  • 成本控制策略:利用蒸馏技术将执行层模型小型化是可行的降本路径,可在维持高性能的同时大幅降低推理成本和延迟。
  • 标准化评估范式:通过固定部分角色并单独扫描其他角色容量的方法,为评估复杂AI系统中各组件的贡献度提供了可复现的实验框架。

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 LLM 大模型 Research 科学研究