Agent-R^2: Scaling Agents for Realistic Productivity Work via Real-Source Data Synthesis
Abstract
A real-source data synthesis framework trains a 4B agent on 4k task instances with supervised fine-tuning and rubric-based reinforcement learning, improving realistic productivity work and file-grounded reasoning over larger models.
Large language models have made rapid progress in agentic reasoning and tool use, raising expectations for agents that can take over realistic productivity work from humans. However, current agents still struggle to finish such tasks, where success requires following long-horizon professional workflows, understanding heterogeneous file contexts, and producing usable deliverables under fine-grained requirements. A key bottleneck is the lack of scalable post-training data that reflects real work intent, authentic file environments, and deliverable-level evaluation. To close this gap, we introduce Agent-R^2, a real-source data synthesis framework that scales training data for realistic productivity work. Instead of relying on manually specified workflows or LLM-generated files, Agent-R^2 grounds synthesis throughout the pipeline in real occupational seeds, corpus-derived workflow evidence, and authentic web-sourced files. From 10k web-corpus candidates, Agent-R^2 distills 4k task instances that pair real multi-format attachments with fine-grained, deliverable-level rubrics. On this data, we train Agent-R^2-4B via supervised fine-tuning and rubric-based reinforcement learning, improving both end-to-end productivity performance and file-grounded reasoning. Trained on only 4k instances, Agent-R^2-4B surpasses Qwen3-235B-A22B-Thinking-2507 on both GDPval (51.0% vs. 48.0%) and OfficeQA (14.8% vs. 13.8%). Further analysis shows that Agent-R^2 mitigates shortcut behavior, where agents produce superficially plausible deliverables before sufficiently grounding, constructing, or verifying them, while promoting sustained exploration.
Community
Real productivity work means turning heterogeneous files (annual reports, spreadsheets, PDFs) into deliverables with the right format and evidence. LLM agents still struggle here: workflows are long, files are messy, and success criteria are fine-grained. A key bottleneck is the lack of scalable training data that captures real work intent, real file environments, and deliverable-level evaluation.
💡 We propose Agent-R², a real-source data synthesis framework for cowork agents. Starting from occupational activity seeds, it extracts file-oriented workflows from web corpora, retrieves and verifies real public attachments, then synthesizes instructions, fine-grained rubrics, and reference deliverables in a sandbox.
🛠Data
- 4k task instances distilled from 10k web candidates
- Avg. 2.51 real multi-format attachments and 15.3 deliverable-level rubrics per task
- Expert trajectories: avg. 18.0 steps, 41.8k tokens
- Rubric-based verification: unreasonable rubrics are revised, and unmet ones trigger one more round of expert revision
⚡ Results (Qwen3-4B-Thinking-2507 → SFT + rubric-based RL → Agent-R²-4B)
- GDPval: 51.0% Pass@1, surpassing Qwen3-235B-A22B-Thinking-2507 (48.0%), +18.6 over the 4B base
- OfficeQA: 2.2% → 14.8% Pass@1, also above Qwen3-235B (13.8%)
- Less shortcut behavior: completion rate reaches 90.9%, with more effective tool steps
- Ablations: real files beat synthetic ones (43.3 vs. 41.4), with 2.54× more expert tool calls (20.4 vs. 8.0); removing the rubric reward drops RL from 51.0 to 49.5
💻 Code: https://github.com/ZJU-REAL/Agent-R2
🤗 Model: https://huggingface.co/teng-pan/Agent-R2-4B
Get this paper in your agent:
hf papers read 2611.08219 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper