Agent-R^2: Scaling Agents for Realistic Productivity Work via Real-Source Data Synthesis
A 4B-parameter model trained on 4,000 tasks beat Qwen3-235B-A22B-Thinking-2507 on two productivity benchmarks, the authors report.
Agent-R² builds training tasks from real occupational examples, workflow evidence and web-sourced files. The pipeline distilled 10,000 candidates into 4,000 tasks with multi-format attachments and detailed criteria for evaluating deliverables. After supervised fine-tuning and rubric-based reinforcement learning, the model scored 51.0% versus Qwen’s 48.0% on GDPval and 14.8% versus 13.8% on OfficeQA. The authors also report fewer shortcuts: producing plausible-looking deliverables before adequately examining files, completing the work or checking results.
HF Daily Papers' note
Agent-R² builds training tasks from real occupational examples, workflow evidence and web-sourced files. The pipeline distilled 10,000 candidates into 4,000 tasks with multi-format attachments and detailed criteria for evaluating deliverables. After supervised fine-tuning and rubric-based reinforcement learning, the model scored 51.0% versus Qwen’s 48.0% on GDPval and 14.8% versus 13.8% on OfficeQA. The authors also report fewer shortcuts: producing plausible-looking deliverables before adequately examining files, completing the work or checking results.
HF Daily Papers' note
score 5