Agent-R² Trains Productivity Agents on Real-World Files

Introduction

A productivity agent needs to do far more than call a tool or produce fluent text. Real work often begins with annual reports, spreadsheets, PDFs, and other heterogeneous files. The agent must understand the user’s intent, follow a multi-step professional workflow, reconcile evidence across files, and produce a deliverable that satisfies detailed requirements. Agent-R² targets the data bottleneck behind these failures: how to build scalable post-training data that resembles actual work.

How the framework works

Rather than asking a language model to invent an entire working environment, Agent-R² anchors synthesis in real sources:

• It starts with occupational activity seeds to identify plausible work intents.
• It extracts file-oriented workflow evidence from web corpora.
• It retrieves and verifies authentic public attachments.
• It constructs instructions, deliverables, and fine-grained rubrics.
• It executes expert trajectories in a sandbox and revises the task when rubrics are unreasonable or requirements remain unmet.

From about 10,000 web-corpus candidates, the authors distill around 4,000 task instances. Each task contains an average of 2.51 real, multi-format attachments and 15.3 deliverable-level rubric items. Expert trajectories average 18 steps and 41.8k tokens. This pipeline connects the work intent, the actual file context, and the definition of a successful output.

Results

The researchers first apply supervised fine-tuning to Qwen3-4B-Thinking-2507 and then use rubric-based reinforcement learning, producing Agent-R²-4B. The resulting model reaches 51.0% Pass@1 on GDPval, ahead of Qwen3-235B-A22B-Thinking-2507 at 48.0%. On OfficeQA, performance rises from 2.2% for the base model to 14.8%, also exceeding the 235B model’s 13.8%.

Further analysis reports a 90.9% completion rate and more effective tool steps. In an ablation, real files score 43.3 versus 41.4 for synthetic files and lead to 2.54 times as many expert tool calls. Removing the rubric reward lowers the reinforcement-learning result from 51.0% to 49.5%.

Why it matters

The study points to a shift in agent training priorities. Scaling model size or trajectory volume alone may be less useful than making the training environment faithful to real work. Authentic files introduce formatting inconsistencies, missing information, and cross-document verification—the kinds of friction that force an agent to search, operate, and check its work instead of relying on superficial patterns. Deliverable-level rubrics also turn vague notions of quality into inspectable objectives.

The approach still has boundaries. The tasks are built from public web sources and should not automatically be treated as representative of every enterprise workflow. Coverage, copyright and privacy, and the extent to which GDPval and OfficeQA capture office performance all warrant further study. The authors have released code and a model, making broader replication possible.

Source: Hugging Face Daily Papers