Agent-R$^2$: Scaling Agents for Realistic Productivity Work via Real-Source Data Synthesis

Teng Pan; Zixuan Wang; Yuchen Yan; Qifei Wu; Yao Chen; Yingqi Qu; Ruiqing Zhang; Yongliang Shen

cs.CLNov 5, 2026

Agent-R $^2$ : Scaling Agents for Realistic Productivity Work via Real-Source Data Synthesis

Authors: Teng Pan, Zixuan Wang, Yuchen Yan, Qifei Wu, Yao Chen, Yingqi Qu, Ruiqing Zhang, Yongliang Shen

Organizations: Zhejiang University · Baidu Inc.

Abstract

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 $\textbf{r}$ eal-source data synthesis framework that scales training data for $\textbf{r}$ ealistic 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.

Figures & tables

Figure 1: GDPval Pass@1 comparison. Agent-R 2 -4B improves the Qwen3-4B-Thinking base model by 18.6 points and surpasses several larger open-source baselines.

Figure 2: Agent-R 2 data synthesis pipeline. The framework extracts file-oriented workflow blueprints from real web-corpus evidence, retrieves and verifies public attachments, and synthesizes instructions, rubrics, and references in sandbox.

Figure 3: Rubric-based RL training. Agent-R 2 samples multi-turn rollouts in sandbox, renders generated deliverables for MLLM-as-a-judge scoring against rubrics and references, and updates the policy with group advantages and KL regularization.

Closed-Source Models
Model	Office		Non-Office		Full
	Pass@1	Pass@3	Pass@1	Pass@3	Pass@1	Pass@3
GPT-4.1 ( OpenAI 2025b )	38.0	51.5	44.9	55.5	39.4	52.3
Gemini-2.5-Pro ( Comanici et al. 2025 )	57.0	59.4	61.4	63.3	57.9	60.2
Claude-Opus-4.5 ( Anthropic 2025 )	73.8	77.4	68.2	72.3	72.6	76.3
Open-Source Models

Table 1: Main results on GDPval. Pass@

k

allows up to

k

attempts and scores the first attempt that produces a deliverable; results are reported for office, non-office, and full subsets. Agent-R 2 -4B consistently improves over its base and SFT variants and outperforms several larger open-source baselines on the full set.

Model	Pass@1	Pass@3
Qwen3-4B-Thinking-2507	2.2	4.5
Qwen3-30B-A3B-Thinking-2507	2.9	5.7
Qwen3-235B-A22B-Thinking-2507	13.8	25.6
Agent-R 2 -4B-SFT	11.4	21.7
Agent-R 2 -4B	14.8	26.8

Table 2: OOD results on OfficeQA. Pass@

k

measures whether at least one of

k

attempts yields a correct answer (

k=1,3

). Agent-R 2 -4B outperforms Qwen3-235B-A22B-Thinking-2507.

Figure 4: Model behavior and completion quality. Left: completion rate, average tool-call steps, and average tool-call errors across the base, SFT, and RL variants. Right: win/tie/loss outcomes of the RL-trained model against the SFT-only model after valid completion on GDPval, where green, orange, and gray segments indicate RL wins, ties, and SFT wins, respectively.

Figure 5: Effective Steps vs. GDPval Performance. Effective steps are computed as average tool-call steps minus tool-call errors. Models with more effective steps generally achieve higher GDPval Pass@1.

Figure 6: Case study of shortcut behavior in materials-test report generation. The base model and Gemini-2.5-Pro produce flawed reports from incomplete or invalid spreadsheet evidence, whereas Agent-R 2 -4B recovers from parsing and execution errors, computes the required values, and verifies the generated plots in the final PDF deliverable.

Metric (%)	Base	SFT	RL
Re-checking the deliverable $\uparrow$	0.5	91.7	95.8
Writing despite an unfixed error $\downarrow$	55.0	16.0	11.0
Numerical-error fix rate $\uparrow$	60.7	79.6	85.7
Deliverables with placeholder text $\downarrow$	12.5	9.6	5.7
Traceable numbers in deliverables $\uparrow$	31.2	34.6	38.8

Table 3: Direct evidence on shortcut behavior on GDPval. All values are percentages: re-checking and placeholder rates are measured over eligible tasks, bypass and fix rates over the corresponding error episodes, and numeric traceability as the mean per-task share of traceable numbers.

Ablation	Method	Pass@1
Data grounding	w/ synthetic files	41.4
	w/ real files	43.3
Reward design	w/o rubrics	49.5
	w/ rubrics	51.0

Table 4: Ablation of real files and rubric rewards on GDPval. SFT models use matched real-file or synthetic-file tasks. RL variants use or omit rubric rewards, starting from the same SFT model.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Solver model	Auxiliary vision model
DeepSeek-v3.1	DeepSeek-v4-Flash
Qwen3-30B-A3B-Thinking-2507	Qwen3-VL-30B-A3B-Thinking
Qwen3-235B-A22B-Thinking-2507	Qwen3-VL-235B-A22B-Thinking
Qwen3-4B-Thinking-2507	Qwen3-VL-4B-Thinking
Agent-R 2 -4B-SFT	Qwen3-VL-4B-Thinking
Agent-R 2 -4B	Qwen3-VL-4B-Thinking

Appendix

Table 5: Auxiliary vision models used for text-only solvers in GDPval evaluation.

Hyperparameter	Value
batch_size	16
max_response_length	65,536
group_size	4
max_assistant_turns	40
max_tool_response_length	4,096
training_temperature	1.0

Appendix

Table 6: Hyperparameters used for reinforcement learning.

Model	Gemini-3.1-Pro	DeepSeek-v4-Flash
Qwen3-235B-A22B	47.7–48.2	49.2–49.9
Agent-R 2 -4B	50.8–51.1	51.7–52.3

Appendix

Table 8: GDPval Pass@1 range over three independent evaluations by each judge model.

SFT data	Held-out score	Win rate	GDPval Pass@1
0 (base)	0.254	–	32.4
1k	0.486	64%	43.3
2k	0.507	67%	44.9
4k	0.617	73%	45.5

Appendix

Table 9: Data scaling of SFT on 100 held-out synthesized tasks and on GDPval.

Model	Training data	GDPval Pass@1
Qwen3-4B-Thinking-2507	–	32.4
Agent-R 2 -4B-SFT	3k SFT	45.3
4k-SFT control	4k SFT	45.5
Agent-R 2 -4B	3k SFT + 1k RL	51.0

Appendix

Table 10: Separating the effect of additional data from the RL stage.

Effective steps per task	SFT	RL
Computation	11.9	7.3
Attachment reading	3.4	3.4
Deliverable writing	2.3	2.7
Build phase (all step types)	4.9	5.5
Share of reading, writing, and checking	40%	48%

Appendix

Table 11: Decomposition of effective steps on GDPval from SFT to RL.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 604 papers

WorkGenesis: Building the Worlds That Teach Agents to Work
Sep 30, 2026Xinyu Zhu, Fenyi Liu, Yuzhu Cai +4LLM Agent Training Synthetic Data Generation for Language Models
Shanghai Jiao Tong University · Endless Frontier · Beihang University
AgentOmnia: Scaling Agentic Models for Full-Scenario Applications
Jul 25, 2026Hao Jiang, Gangtao Xin, Yingdi Huang +35Tool-Augmented Language Model Agents Agentic RL
Huawei Cloud Post-Training Team
Data World: Scaling Data Environments from Scratch
Oct 15, 2026Zhanli Li, Zeyuan Xu, Chaoxu Pang +16Synthetic Data Generation for LLM Agents Data Analysis Agents
Longcat Interaction Team, Meituan · Zhongnan University of Economics and Law · Hunan University +7

Sep 30, 2026J/K move · Enter open · S save

WorkGenesis: Building the Worlds That Teach Agents to Work

Xinyu Zhu, Fenyi Liu, Yuzhu Cai, Shuo Tang+3

Shanghai Jiao Tong University · Endless Frontier · Beihang University

LLM Agent Training Synthetic Data Generation for Language Models

Open paper AskPDF arXiv

The ability of Large Language Model (LLM) agents to complete daily and professional work is receiving increasing attention. Training such agents requires realistic work scenarios. Expert-authored occupational work is costly and slow to produce, while unconstrained synthesis often yields tasks with weak factual grounding or internally inconsistent requirements. To bridge this gap, we introduce WorkGenesis, a framework that constructs executable occupational work from real-world artifacts through two core technical innovations: (1) Evidence-Based Work Construction, which grounds each unit of work in real-world evidence by retrieving public files guided by O*NET occupational knowledge and synthesizing the surrounding context, companion materials, work request, and itemwise rubric around them; and (2) Execution-Guided Consistency Verification, which renders a reference deliverable inside the constructed work, attributes every unsatisfied rubric item to the agent, the task, or the rubric, and uses task and rubric defects as feedback to iteratively repair the work until it passes the audit. Experimental results demonstrate that Fx-Work-35B, trained with simple supervised fine-tuning (SFT) on only 20K units of work synthesized by WorkGenesis, achieves the highest scores among all comparable-scale baselines on the five reported metrics across GDPvalAA-v2, APEX-Agents-AA, and JobBench (31.00 versus 24.79 average score), and even surpasses frontier models such as the 1.6T DeepSeek-V4-Pro-Preview. These results show that WorkGenesis provides scalable training data for working agents.

Abstract

LLM Agent Training

Explore similar work

WorkGenesis: Building the Worlds That Teach Agents to Work

AgentOmnia: Scaling Agentic Models for Full-Scenario Applications

Data World: Scaling Data Environments from Scratch