Agent post-training data has a structural problem: it is usually generated by the same model family the agent is supposed to outperform. Workflows are hand-written into prompts, the files an agent is asked to read are synthetic, and the evaluation is a checklist rather than a deliverable. Train on that and you get agents that are fluent about work and weak at doing it.
Agent-R², posted to arXiv on 5 November, changes where the material comes from.
The Move
The framework’s core claim is in its name: synthesis is grounded in real-source data at every stage, not just at the prompt level. It uses real occupational seeds, corpus-derived workflow evidence, and authentic web-sourced files — the attachments an agent would meet on an actual job, not documents another model made up.
The funnel is aggressive. From 10,000 web-corpus candidates, the pipeline distills 4,000 task instances, each pairing real multi-format attachments with fine-grained, deliverable-level rubrics. The rubrics matter as much as the data: an instance is judged on the artefact produced, not on the steps taken.
Training uses supervised fine-tuning followed by rubric-based reinforcement learning, producing Agent-R²-4B.
What It Buys
The headline is a scale inversion. Trained on only 4,000 instances, Agent-R²-4B surpasses Qwen3-235B-A22B-Thinking-2507 on both GDPval (51.0% vs. 48.0%) and OfficeQA (14.8% vs. 13.8%) — a 4B model beating a frontier-scale thinking model on real professional work.
The analysis section explains why, and it is not simply “more data”. The framework mitigates shortcut behaviour, where agents emit a superficially plausible deliverable before actually grounding, constructing, or verifying it, and it promotes sustained exploration. Synthetic data teaches plausible-looking completion; real files with real rubrics punish it.
Why It Matters
The obvious objection is contamination — mined web corpora may leak benchmark material. The paper’s answer is the rubric structure rather than the corpus size: with 4,000 instances, memorising the source set is not a viable strategy, and the surviving gains point to the grounding and evaluation format rather than to knowledge leakage.
The deeper implication is about where agent research effort goes. If a 4B model trained on four thousand carefully-grounded real tasks can match a 235B model on productivity benchmarks, then the bottleneck for office-grade agents may not be model capacity at all — it may be that nobody had bothered to build training data from real work.
Paper: arXiv:2611.08219