Marc Tommasi

Achieving fairness in synthetic data generation isn't just about bias correction—it's about ensuring utility parity across sensitive groups, and this paper reveals how to do it effectively.

You can shrink a privacy-sensitive LLM by 4500x and still get human-level agreement on data privacy assessments.

Forget hand-crafted rules: this work learns to prompt language models for optimal text anonymization, adapting to diverse privacy needs and outperforming static methods.