DataPraxis: Scaling Analytics Tasks via Verified Workspace Evolution

Introduction

The growing expectation that general‑purpose AI agents can handle professional analytics—conducting sustained reasoning, iterative investigation, and producing verifiable evidence—has exposed a gap in existing training resources. Most synthetic tasks are derived directly from raw data, which often neglects the deeper, follow‑up questions that naturally arise during analysis and provides little means to confirm the correctness of intermediate results. DataPraxis tackles this shortfall by treating the evolution of an analytical workspace itself as a first‑class training signal. The framework incrementally enriches a workspace with validated findings and artifacts, then recombines these enriched workspaces to generate more complex, layered tasks.

---

Verified Workspace Evolution

At the heart of DataPraxis is a verified workspace evolution loop. A workspace starts with a raw dataset and a simple exploratory prompt. As the agent interacts with the data, each step produces a finding (e.g., a statistical summary, a visual insight) together with a verification artifact (such as a unit test, a cross‑check against a known baseline, or a citation to an external source). These verified findings are stored in the workspace and become reusable building blocks for later steps.

The evolution proceeds in two phases:

1. Enrichment – The agent repeatedly augments the workspace, each time checking that new findings are consistent with prior evidence. Failures trigger a rollback or a refinement, ensuring that the workspace never diverges into an unverified state.
2. Inheritance & Combination – Once a workspace reaches a stable, verified state, it can be inherited by a new task. The new task inherits the prior findings and is encouraged to ask deeper questions that depend on them, effectively creating a hierarchy of analytical narratives.

This approach mirrors how human analysts work: they build a foundation of trusted results before probing further, and they constantly validate each inference against the data and prior knowledge.

---

Building the DataPraxis‑5K Dataset

Using the evolution framework, the authors generated DataPraxis‑5K, a collection of 5,000 verified analytical trajectories spanning multiple domains (finance, healthcare, climate, etc.) and output formats (tables, plots, natural‑language summaries). Each trajectory records:

• The initial raw dataset.
• The sequence of workspace states, each annotated with the finding and its verification artifact.
• The final task prompt and the model’s answer.
• A dependency graph that links each finding to the ones it builds upon.

Because verification is baked into every step, the dataset provides a reliable ground truth for both process (how the analysis unfolded) and product (the final answer). This dual signal is valuable for training agents that must not only produce correct results but also explain how they arrived at them.

---

Model Performance and Implications

Two models—DataPraxis‑27B and DataPraxis‑35B‑A3B—were fine‑tuned on the 5K trajectories. On the DABStep benchmark, DataPraxis‑27B achieved a score of 55.11, while on JobBench it reached 51.9, matching or surpassing contemporary frontier models such as GLM‑5.2. These results suggest that the verified workspace paradigm can close the gap between raw predictive ability and trustworthy analytical reasoning.

Beyond raw scores, the models exhibit more transparent reasoning chains. When queried about a particular conclusion, they can surface the exact workspace state and verification artifact that underpins it, offering a level of interpretability that is rare in large‑scale language models.

---

Practical Takeaways

• Task synthesis matters: Building training data that reflects the iterative, verified nature of real analytics yields stronger, more reliable agents.
• Verification as a first‑class signal: Embedding checks (unit tests, cross‑validation, citations) at each step forces the model to internalize the habit of self‑validation.
• Reusable workspaces: By inheriting and recombining verified findings, a single analytical thread can seed many downstream tasks, dramatically expanding the diversity of training examples without manual labeling.
• Interpretability by design: The dependency graph and verification artifacts provide a built‑in audit trail, facilitating debugging and user trust.

---

Closing Thoughts

DataPraxis demonstrates that scaling analytics capabilities does not require ever larger models alone; it also demands richer, more structured training experiences. By treating the evolution of a workspace as a verifiable, composable artifact, the framework produces a dataset that teaches agents how to think analytically, not just what to answer. As AI assistants move deeper into professional domains, such verified, process‑aware training pipelines will be essential for delivering trustworthy, actionable insights.