Benchmarking Hybrid Deep Research Across Database Querying and Web Search

跨数据库查询与网络搜索的混合深度研究基准测试

Abstract: While autonomous agents have made significant strides in “deep research” by iteratively navigating the open web to synthesize information, real-world problem-solving is rarely confined to a single environment. Complex analytical tasks inherently require agents to weave together evidence from both ambiguous unstructured text (e.g., the open web) and highly precise structured data (e.g., relational databases).

摘要： 尽管自主智能体在通过迭代浏览开放网络以综合信息的“深度研究”方面取得了显著进展，但现实世界的问题解决过程很少局限于单一环境。复杂的分析任务本质上要求智能体将来自模糊非结构化文本（如开放网络）和高精度结构化数据（如关系数据库）的证据交织在一起。

However, existing benchmarks evaluate these modalities in isolation, failing to capture the critical “handoff” - the ability to preserve constraints when moving evidence between systems. We introduce HybridDeepResearch, to our knowledge the first deep-research benchmark that requires both web search and SQL to form a complete, verifiable answer.

然而，现有的基准测试通常孤立地评估这些模态，未能捕捉到关键的“交接”过程——即在系统间迁移证据时保持约束条件的能力。我们引入了 HybridDeepResearch，据我们所知，这是第一个要求同时使用网络搜索和 SQL 来形成完整、可验证答案的深度研究基准测试。

The benchmark contains 380 tool-dependent tasks grounded in LiveSQLBench-Base-Lite databases and public web corpora, validated through automated checks and human review, and covering three reasoning patterns: SQL2S, S2SQL, and Parallel.

该基准测试包含 380 个依赖工具的任务，这些任务基于 LiveSQLBench-Base-Lite 数据库和公共网络语料库，通过自动化检查和人工审核进行验证，并涵盖了三种推理模式：SQL2S（SQL 转搜索）、S2SQL（搜索转 SQL）和并行推理。

Evaluations across proprietary and open-weight models under various agentic scaffolds reveal that even state-of-the-art models like GLM-5.2, Claude-Sonnet-4.6 and GPT-5 achieve only about 50-54% Pass@8 on the hard subset. Notably, results show that directional reasoning is substantially more difficult than parallel intersection, highlighting that bridging structured and unstructured information spaces without losing constraints remains a major open challenge for agentic systems.

在各种智能体框架下，对专有模型和开源权重模型的评估显示，即使是 GLM-5.2、Claude-Sonnet-4.6 和 GPT-5 等最先进的模型，在困难子集上的 Pass@8 也仅达到约 50-54%。值得注意的是，结果表明定向推理比并行交叉推理要困难得多，这凸显了在不丢失约束条件的情况下连接结构化和非结构化信息空间，对于智能体系统而言仍然是一个重大的开放性挑战。

Code and datasets are publicly available at GitHub (this https URL) and Hugging Face (this https URL).

代码和数据集已在 GitHub（此链接）和 Hugging Face（此链接）上公开。