RAAS：基于 GRPO 的大语言模型智能体系统架构搜索
RAAS: LLM Agentic System Architecture Search with GRPO
论文提出的「同查询候选群组对比 + 多试验聚合」稳定评估思路，可直接迁移到视觉智能体（如视觉工具调用、多模态推理流水线）的架构搜索与自动设计，缓解绝对分数受任务难度干扰的问题。
原文依据 evaluating cohorts of candidate architectures on identical queries, deriving context-aware merit signals through peer-group comparison
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论大语言模型（LLM）智能体系统通过协同工作流求解复杂任务，但其设计过程仍然十分耗费人力。智能体超网（Agentic Supernet）范式通过优化一个概率化架构空间来实现自动化，却受制于严重的评估不稳定性：绝对性能分数将架构优劣与查询难度纠缠在一起，而单次执行协议捕捉到的是执行随机性而非真实能力。这些不稳定性导致搜索动态不可靠：简单查询会抬高弱设计，困难查询则会压制强设计。我们提出 RAAS（鲁棒架构自适应搜索），通过两种协同机制建立更稳定、更公平的评估。上下文架构编排（CAO）在相同查询上评估候选架构群组，通过同组对比导出上下文感知的优劣信号，从而将质量与任务难度解耦。多试验评估合成（MTAS）通过聚合多次独立试验的性能来降低执行方差，产生统计上稳健的能力估计。二者共同为架构发现提供更可靠的信号。在涵盖数学推理、代码生成以及一个多步工具使用基准的六个基准上的实验表明，RAAS 稳定优于强基线，将 HumanEval pass@1 从 92.23% 提升至 96.31%，MATH 准确率从 52.08% 提升至 60.87%，同时保持实用效率。这些结果表明，在所研究的设定下，稳健评估是智能体架构搜索的一个有用要素。
英文原文摘要 1665 字符
Large Language Model (LLM) agentic systems solve complex tasks through coordinated workflows, but designing them remains labor-intensive. The Agentic Supernet paradigm automates this by optimizing a probabilistic architecture space, yet suffers from critical evaluation instabilities: absolute performance scores entangle architectural merit with query difficulty, while single-execution protocols capture execution randomness rather than true capability. These instabilities lead to unreliable search dynamics where simple queries inflate weak designs and challenging queries suppress strong ones.We introduce RAAS (Robust Architecture Adaptive Search), which establishes more stable and fair evaluation through two synergistic mechanisms. Contextual Architecture Orchestration (CAO) disentangles quality from task difficulty by evaluating cohorts of candidate architectures on identical queries, deriving context-aware merit signals through peer-group comparison. Multi-Trial Assessment Synthesis (MTAS) reduces execution variance by aggregating performance across multiple independent trials, producing statistically robust capability estimates.Together, these mechanisms provide more reliable signals for architecture discovery. Experiments on six benchmarks spanning mathematical reasoning, code generation, and one multi-step tool-use benchmark show that RAAS consistently improves over strong baselines, improving HumanEval pass@1 from 92.23% to 96.31% and MATH accuracy from 52.08% to 60.87%, while maintaining practical efficiency. These results suggest that robust evaluation is a useful ingredient for agentic architecture search in the studied settings.