DRFlow (DRBench Agent): leaderboard
DRFlow (ServiceNow, 2026) evaluates deep-research agents on predicting personalized enterprise workflows, the concrete action steps a user must take, from scattered private sources in five domains. This board runs the planning-based DRBench Agent on each backbone model over the 50-task original split and scores step F1, the task-wise harmonic mean of step recall and precision against the reference workflow. Higher is better.
Metric: Step F1 (%) on the 50-task original split of DRFlow (enterprise deep-research tasks in five domains: predict a personalized workflow of action steps from scattered heterogeneous sources): the task-wise harmonic mean of step recall and precision against the reference workflow (steps matched by an LLM judge), averaged over tasks; the DRBench Agent, a planning-based enterprise deep-research baseline with staged planning and adaptive execution; higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 54.17 | #176 |
| 2 | Claude Opus 4.5 | 51.09 | #156 |
| 3 | DeepSeek V3.2 | 45.37 | #406 |
| 4 | Gemini 3.1 Pro (Preview) | 40.99 | #94 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=drflow-drbench-agent · How It Works · Data refreshed daily, snapshot 2026-11-13.