ACTG-ARL: Differentially Private Conditional Text Generation with RL-Boosted Control

Generating high-quality synthetic text under differential privacy (DP) is critical for training and evaluating language models without compromising user privacy. Prior work on synthesizing DP *datasets* often fail to preserve key statistical attributes, suffer utility loss from the noise required by DP, and lack fine-grained control over generation. To address these challenges, we make two contributions. First, we introduce a hierarchical framework that decomposes DP synthetic text generation into two subtasks: *feature learning* and *conditional text generation*. This design explicitly incorporates learned features into the generation process and simplifies the end-to-end synthesis task. Through systematic ablations, we identify the most effective configuration: a rich tabular schema as feature, a DP tabular synthesizer, and a DP fine-tuned conditional generator, which we term ACTG (**A**ttribute-**C**onditioned **T**ext **G**eneration). Second, we propose Anchored RL (ARL), a post-training method that improves the instruction-following ability of ACTG for conditional generation. ARL combines RL to boost control with an SFT anchor on best-of-$N$ data to prevent reward hacking. Together, these components form our end-to-end algorithm **ACTG-ARL**, which advances both the quality of DP synthetic text (+20% MAUVE over prior work) and the control of the conditional generator under strong privacy guarantees. Our code is at https://github.com/actg-arl/ACTG-ARL.

Modern AI is trained on vast amounts of human-generated text, such as emails, medical records, and customer reviews, but exposing raw data to algorithms can leak personal information. One alternative is to share synthetic text: AI-generated stand-ins that preserve patterns in real data while provably protecting individuals. Existing methods, however, often struggle to retain key statistical properties of the original data and offer limited control over the type of text generated. We propose a two-stage recipe. First, we learn a privacy-preserving “skeleton” of the data: a structured summary that captures attributes such as topic, sentiment, or length. Second, we train a language model to flesh out this skeleton into realistic text that respects the requested attributes. To make the model follow instructions reliably, we add a reinforcement-learning step that rewards desirable behavior, together with a safeguard that prevents the model from gaming the reward. Together, these techniques produce shareable text that is higher-quality and more controllable than prior privacy-preserving methods. This makes it easier for hospitals, banks, and other institutions to extract insights from sensitive text, and to train AI models on it, without putting users at risk.