DRFLOW
收藏资源简介:
DRFLOW是由ServiceNow AI研究院与不列颠哥伦比亚大学联合构建的个性化工作流预测深度研究基准数据集,旨在评估智能体从异构数据源中推理并生成可执行操作序列的能力。该数据集涵盖100个跨领域任务,包含1,246个参考工作流步骤,并植根于超过3,900份来源文档,数据规模庞大且结构复杂。数据集通过端到端的合成管道生成,结合人工验证确保质量,模拟了企业环境中分散于文档、邮件和聊天记录等多源证据的真实场景。其核心应用在于推动智能体在政策合规、客户问题解决等场景中,从混杂信息中自动推断结构化、个性化的操作流程,填补了现有深度研究基准在可执行工作流预测领域的空白。
DRFLOW is a benchmark dataset for in-depth research on personalized workflow prediction, co-developed by ServiceNow AI Research and the University of British Columbia. It aims to evaluate the ability of AI Agents to reason over heterogeneous data sources and generate executable action sequences. This dataset covers 100 cross-domain tasks, contains 1,246 reference workflow steps, and is sourced from over 3,900 source documents, boasting a large scale and complex structure. Generated via an end-to-end synthetic pipeline and paired with manual validation to ensure data quality, it simulates real-world enterprise scenarios where evidence is scattered across multiple sources such as documents, emails, and chat logs. Its core applications lie in advancing AI Agents to automatically infer structured, personalized operational workflows from mixed information in scenarios like policy compliance and customer issue resolution, filling the gap in existing deep research benchmarks for executable workflow prediction.

- 1DRFLOW: A Deep Research Benchmark for Personalized Workflow PredictionServiceNow AI研究院; 不列颠哥伦比亚大学 · 2026年



