ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?

2026-06-01

AI Summary

Existing evaluation methods struggle to assess whether differentially private (DP) synthetic text effectively transfers novel knowledge and capabilities from the original sensitive corpus. To address this gap, this work proposes ContinuousBench—the first dynamic evaluation framework centered on “capability gain.” It features quarterly automated updates of training corpora and corresponding question sets, creating data-dependent tasks that cannot be solved without access to the underlying corpus while remaining compatible with DP learning. The benchmark comprises two tracks, Geminon and News, and integrates DP text synthesis, programmatic data generation, and a standardized training-evaluation pipeline, leveraging large-scale independent records to model knowledge. Experiments show that non-private synthetic methods significantly transfer knowledge, whereas state-of-the-art DP approaches—even at ε=100—fail to achieve effective capability transfer.

Abstract

Differentially private (DP) text synthesis promises to unlock sensitive corpora for model training, but it remains unclear whether DP synthetic data transmits genuinely new knowledge and capabilities present only in those corpora. This is because existing evaluations rely on tasks that are nearly solvable without training, so strong benchmark performance does not establish that DP synthesis can substitute original data access. Thus, we introduce ContinuousBench, a continuously and automatically-regenerated benchmark that measures capability gain from DP synthetic text. Each quarter, a new release pairs a never-before-seen training corpus with a derived QA set, constructed to be: (1) unsolvable sans-corpus; and (2) learnable under DP, as the tested knowledge is supported by hundreds of independent records. Researchers produce DP synthetic data from the training corpus and run our standardized training and evaluation harness on their synthetic data to measure gains. We instantiate two tracks: Geminon, a procedurally-generated dataset about fictional creatures; and News, a stream of newly crawled public news articles. Although standard benchmarks are nearly saturated, on ContinuousBench we find that non-private synthesis transfers substantial knowledge from the original corpus, while state-of-the-art DP synthesis methods generally fail to do so, even at $\varepsilon=100$.

Problem

• differentially private text synthesis
• capability evaluation
• synthetic data utility
• benchmarking
• knowledge transfer

Innovation

• differentially private synthesis
• capability evaluation
• synthetic text
• ContinousBench
• knowledge transfer