Role of shared person–email regularities in unseen-identity recovery

Determine the contribution of shared person–email regularities to the recovery gains achieved by ReGap when adaptation identities are disjoint from the memorized and evaluation identities and no exact target answers appear in the generated or selected supervision.

Background

ReGap is evaluated in an Unseen-Identity Control in which the identities used to construct adaptation data have zero overlap with the memorized and evaluation identities, and the generated candidate pools and selected examples contain no exact target answers. Nevertheless, recovery improves substantially for the evaluated GPT-2 models.

The authors attribute this result to recovery that does not rely on direct identity reuse or exact-answer copying, but explicitly leave unresolved how much of the gain is instead explained by shared person–email regularities present in the structured Enron-derived task. Resolving this issue would clarify whether the observed recovery reflects broader task-distributional structure rather than target-specific latent information.

References

This rules out direct identity reuse and exact-answer copying as explanations for these gains, while leaving open the contribution of shared person–email regularities.

— Sleeping Secrets: How Fine-Tuning Reawakens Privacy Risks in Language Models  (2610.01365 - Li et al., 1 Oct 2026) in Section 5.3, Query-information ablation discussion (p. 6)