SAP researchers unveiled Synthesis Through Simulation (STS), a schema‑free data‑generation framework that lets large language model agents produce realistic enterprise tables by interacting with policy‑enforcing APIs inside simulated environments. The approach guarantees structural validity while decoupling it from distribution modeling, achieving 0.88 average marginal fidelity and full constraint compliance across ten testbeds without needing database schemas.
Core Mechanics of STS
- STS employs a Generalist Populator (GP) agent that executes API calls in ten simulated enterprise settings.
- GP delivers 0.88 average marginal fidelity and 100% constraint compliance, outperforming statistical synthesizers and schema‑privileged agents.
- The framework is fully open‑source, with code, environments, and datasets on GitHub.
Why This Matters in Enterprise AI
Traditional tabular synthesis struggles with structural validity or requires detailed schemas, both of which are hard to obtain in regulated business contexts. STS sidesteps these bottlenecks by generating data directly within the same environment that defines validity, thereby eliminating the need for proprietary schema access and reducing legal exposure.
Next‑Gen Data Generation and AI Tooling
By decoupling validity enforcement from distribution modeling, STS paves the way for scalable training of tool‑calling agents across diverse domains. The open‑source release invites researchers to extend the ten environments, potentially enabling cross‑industry benchmarks and accelerating the adoption of AI‑driven enterprise workflows.