Researchers propose schema-free data synthesis method using LLM agents to generate enterprise data

Enterprise AI deployments face persistent constraints due to legal and business restrictions on accessing real-world systems, data, and database schemas, limiting the scale at which tool-calling agents can be trained and evaluated.

Existing tabular data synthesis methods are fundamentally limited by structural validity and schema availability, while procedure-based approaches often lack distributional fidelity without extensive per-domain authoring.

The proposed Synthesis Through Simulation (STS) paradigm introduces a schema-free approach where an LLM agent generates data by executing operations against policy-enforcing APIs within simulated enterprise environments.

Because data generation occurs through the same environment that defines validity, STS guarantees structural validity by construction and decouples validity enforcement from distribution modeling, enabling independent optimization of each component.

The Generalist Populator (GP), a domain-agnostic agent within STS, achieves 0.88 average marginal fidelity and 100% constraint satisfaction across all ten simulated environments without requiring access to database schemas.

In comparative evaluations, statistical synthesizers were inapplicable to seven of the ten environments due to necessary seed data requirements, and schema-privileged agents failed 82% of trajectories in the airline environment because of brittle task composition in tightly coupled workflows.

The researchers open-sourced the full framework, all ten simulated environments, and the generated datasets to support reproducibility and further research.