Inference-Efficient Zero-Shot Differentially Private Text Synthesis

Real-world text is largely unstructured and spans diverse domains, yet its analysis is often impeded by privacy constraints. Differentially private (DP) synthetic data generation via LLM APIs has emerged as a practical approach to mitigate privacy concerns. However, state-of-the-art methods incur massive computational overhead. In this work, we introduce OASIS, an inference-efficient, zero-shot DP synthesis framework that generates documents with a single API query. Without in-context examples, OASIS achieves competitive downstream utility against evolutionary algorithms evaluated on the OpenReview dataset.