Research Papers 论文研究 12h ago Updated 8h ago 更新于 8小时前 45

SearchAtlas: Analyzing Agentic Search Strategies via Evidential Query Graphs SearchAtlas:通过证据查询图分析代理搜索策略

SearchAtlas is a framework that converts LLM search agent trajectories into structured evidential query graphs, mapping how evidence propagates from queries to final answers The automated parsing pipeline achieves 86.0% mean edge F1 against human-annotated graphs with consistent performance across repeated runs Analysis of five search agents across three benchmarks reveals systematic differences in search scale and evidence aggregation patterns SearchAtlas exposes three key process failures: fra 提出SearchAtlas框架,将LLM搜索代理的原始轨迹转换为结构化证据图,边表示证据从查询到最终答案的传播路径 自动解析管道在人类标注图上达到86.0%的边F1分数,且跨重复运行保持一致性 分析五个搜索代理在三个基准上的表现,揭示搜索规模和证据聚合的系统性差异 发现三类过程失败:碎片化答案支持、问题约束未传递到答案、未验证的参数知识进入响应 过程诊断分数比LLM judge(基于原始轨迹或排序查询列表)更能预测答案正确性,且包含答案准确性之外的信息

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • SearchAtlas is a framework that converts LLM search agent trajectories into structured evidential query graphs, mapping how evidence propagates from queries to final answers
  • The automated parsing pipeline achieves 86.0% mean edge F1 against human-annotated graphs with consistent performance across repeated runs
  • Analysis of five search agents across three benchmarks reveals systematic differences in search scale and evidence aggregation patterns
  • SearchAtlas exposes three key process failures: fragmented answer support, unreachable question constraints, and unverified parametric knowledge entering responses
  • Process-diagnostic scores from the graph analysis correlate more strongly with answer correctness than LLM judges evaluating raw trajectories or ordered query lists

Why It Matters

This work addresses a critical gap in LLM agent evaluation: the overreliance on final-answer accuracy while ignoring the reasoning process. By making search strategies interpretable through graph structures, SearchAtlas gives practitioners a diagnostic tool to identify exactly where and why agents fail, enabling targeted improvements rather than blind benchmarking.

Technical Details

  • Graph Construction: SearchAtlas transforms raw search trajectories into structured graphs where nodes represent queries and evidence, and edges encode the propagation of evidence from retrieval to final answer generation
  • Parsing Pipeline: An automated pipeline parses agent trajectories into these graphs, achieving 86.0% mean edge F1 against human-annotated ground truth, with consistency across repeated runs
  • Evaluation Scope: Five distinct search agents were analyzed across three benchmarks, examining differences in search scale (number of queries) and evidence aggregation strategies
  • Diagnostic Metrics: The framework quantifies process failures including fragmented answer support (evidence not cohesively linked), question constraint coverage (whether all constraints are addressed), and parametric knowledge leakage (unverified internal knowledge used as evidence)
  • Comparative Analysis: Process-diagnostic scores were compared against LLM judge evaluations on both raw trajectories and ordered query lists, demonstrating superior correlation with actual answer correctness

Industry Insight

  • Agent evaluation frameworks should incorporate process-level diagnostics alongside outcome metrics; final accuracy alone masks systematic reasoning failures that prevent generalization
  • The graph-based interpretability approach could be adapted for other agentic workflows beyond search, such as code generation or multi-step planning, where traceability of evidence is equally critical
  • Teams building production search agents should monitor evidence propagation completeness and constraint coverage as early-warning signals, since process failures strongly predict answer errors even when the final answer appears correct

TL;DR

  • 提出SearchAtlas框架,将LLM搜索代理的原始轨迹转换为结构化证据图,边表示证据从查询到最终答案的传播路径
  • 自动解析管道在人类标注图上达到86.0%的边F1分数,且跨重复运行保持一致性
  • 分析五个搜索代理在三个基准上的表现,揭示搜索规模和证据聚合的系统性差异
  • 发现三类过程失败:碎片化答案支持、问题约束未传递到答案、未验证的参数知识进入响应
  • 过程诊断分数比LLM judge(基于原始轨迹或排序查询列表)更能预测答案正确性,且包含答案准确性之外的信息

为什么值得看

当前LLM搜索代理评估过度依赖最终答案准确率,忽视了搜索过程的质量分析。SearchAtlas提供了一种可解释的过程诊断工具,帮助从业者识别搜索策略中的系统性缺陷,对提升搜索代理的可靠性和可调试性具有重要价值。

技术解析

  • SearchAtlas将搜索轨迹解析为结构化图,节点代表查询和证据片段,边表示证据在推理链中的传播关系,实现从非结构化轨迹到可分析图结构的转换
  • 自动解析管道在人类标注的图上进行评估,达到86.0%的边F1分数,且跨重复运行保持一致性,证明解析方法的可靠性
  • 在三个基准上分析五个搜索代理,揭示不同代理在搜索规模和证据聚合策略上的系统性差异
  • 识别三类过程失败模式:碎片化答案支持(证据不连贯)、问题约束未传递(查询约束未影响最终答案)、未验证参数知识(模型内部知识未经检索验证直接进入响应)
  • 过程诊断分数与答案正确性的关联强度超过LLM judge,且审计显示两者存在不可约的差异,证明过程分析具有独立诊断价值

行业启示

  • 搜索代理评估应从单一的结果准确率转向过程质量评估,过程诊断能揭示答案准确性无法捕捉的缺陷
  • 证据传播的可解释性分析为优化搜索策略提供了新的调试维度,可指导代理设计改进
  • 过程失败模式与错误答案的高度相关性表明,改进搜索过程本身比事后修正答案更有效,应优先优化证据检索和聚合机制

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。

LLM 大模型 Agent Agent Evaluation 评测 Research 科学研究