A 4B agent trained on 4,000 real-source office tasks edges past a 235B model on GDPval, 51.0% to 48.0%
Synopsis
Zhejiang University and Baidu introduce Agent-R2, a pipeline that synthesizes productivity-work training data from real sources: it distills 4,000 task instances from 10,000 web-corpus candidates, each pairing authentic multi-format attachments with fine-grained rubrics. Trained on this data with supervised fine-tuning and rubric-based reinforcement learning, Agent-R2-4B reaches 51.0% Pass@1 on GDPval versus 48.0% for Qwen3-235B-A22B-Thinking-2507, and 14.8% versus 13.8% on OfficeQA.
Interpretation
Training data no longer comes from hand-written workflows or LLM-generated files; workflows are mined from web corpora and public files are searched and verified as attachments, so tasks resemble real office work. Earlier agent training data covered capabilities in isolation, such as instruction following, document question answering, or single-type file generation, and lacked end-to-end supervision over real work objectives, cross-file evidence, professional workflow execution and deliverable constraints. In a matched 1,000-task comparison, the SFT model trained on real files scored 43.3 GDPval Pass@1 versus 41.4 for synthetic files; expert trajectories on real files used 2.54 more tool calls on average (20.4 versus 8.0).
Agent-R2-4B builds on Qwen3-4B-Thinking-2507, first supervised fine-tuning on 3,000 tasks and then rubric-based reinforcement learning on 1,000 non-overlapping tasks, reaching 51.0% Pass@1 on the full GDPval set. The same base model scores 32.4% Pass@1 on the full GDPval set, and the SFT-only Agent-R2-4B-SFT scores 45.3%. On the full GDPval set, Agent-R2-4B's 51.0% exceeds Qwen3-235B-A22B-Thinking-2507's 48.0% and Qwen3-VL-235B-A22B-Thinking's 49.7%; on out-of-domain OfficeQA it lifts the base model from 2.2% to 14.8%, above Qwen3-235B-A22B-Thinking-2507's 13.8%.
The trained agent reduces shortcut behavior, where an agent emits a superficially plausible deliverable before sufficiently grounding, constructing or verifying it, and instead sustains exploration. The base model averages only 6.7 tool-call steps, writes a deliverable despite an unfixed error in 55.0% of build-phase cases, and has 31.2% traceable numbers in deliverables. Across all 220 GDPval tasks, Agent-R2-4B re-checks 95.8% of deliverables, writes despite an unfixed error in 11.0% of cases, fixes 85.7% of numerical errors, reaches 38.8% traceable numbers, and completes 90.9% of tasks.
Perspective
The demonstrated setting is office-type productivity tasks in a sandbox, with attachments drawn from real public PDF, docx, xlsx and pptx files, evaluated on GDPval and OfficeQA; non-office deliverables such as audio and video are stated as outside the main focus.
GDPval scoring relies on an LLM judge, and per-task scores vary across repeats with median absolute differences of 3.2 and 3.1 points, though aggregate scores stay stable; in a manual audit of 100 synthesized tasks, workflow-question alignment was 86% and task feasibility 85%.
