DRFLOW Benchmark Evaluates Personalized AI Workflow Prediction

This research introduces DRFLOW, a novel benchmark specifically designed to assess the capability of AI agents in predicting personalized workflows. Unlike traditional deep research systems that primarily generate reports or summaries, DRFLOW focuses on tasks requiring agents to identify concrete sequences of action steps, such as navigating complex enterprise procedures. The benchmark comprises 100 tasks across five distinct domains, featuring 1,246 reference workflow steps supported by over 3,900 diverse sources. It includes seven diagnostic metrics to thoroughly evaluate aspects like factual grounding, step recovery, structural ordering, condition resolution, and personalization. Alongside the benchmark, the paper presents DRFLOW-Agent (DRFA), a reference agent tailored for workflow prediction. While DRFA shows improvements over existing baselines, the results indicate that accurately predicting complete and personalized workflows remains a significant challenge for current deep research systems, underscoring the need for further advancements.