Paper Overview
Research Area: ML Authors: Nan Jia, Wenxuan Shi, Jingqing Ruan, Zhizhao Zeng, Tianyi Zhai, Zhongda Su, Feng Hong, Zhengyu Chen, Wenhao Yan, Haojin Yang, Siqi Dai, Weipeng Zhang, Ke Zeng, Xunliang Cai Published: 2026-10-22 arXiv: 2610.26924
Abstract
General-purpose agents are increasingly expected to perform professional analytics work, which requires sustained reasoning, iterative investigation, and reliable evidence. Training these capabilities requires diverse and challenging tasks. However, generating tasks directly from raw data often overlooks deeper questions that emerge through analysis and provides limited evidence for verification. To address these limitations, we introduce DataPraxis, an analytics task synthesis framework based on verified workspace evolution. The framework progressively enriches workspaces with validated findings and artifacts, then inherits and combines these workspaces to uncover deeper questions. It constructs tasks from accumulated findings and their dependencies, verifying analytical results both during workspace evolution and after task assembly. Leveraging this framework, we construct DataPraxis-5K, a dataset of 5,000 verified trajectories spanning diverse domains and output formats, and use it to train DataPraxis-27B and DataPraxis-35B-A3B. DataPraxis-27B achieves 55.11 on DABStep and 51.9 on JobBench, matching or exceeding frontier models such as GLM-5.2.
Key Contributions
- Verified workspace evolution: workspaces are progressively enriched with validated findings and artifacts, then inherited and combined to surface deeper analytical questions.
- Verification at both stages: analytical results are verified during workspace evolution and again after task assembly.
- DataPraxis-5K dataset: 5,000 verified trajectories across multiple domains and output formats.
- Strong trained models: DataPraxis-27B and DataPraxis-35B-A3B; DataPraxis-27B reaches 55.11 on DABStep and 51.9 on JobBench, comparable to or surpassing frontier models like GLM-5.2.