Saved in:
Bibliographic Details
Main Authors: Bouzid, Nassima M., Yuan, Dehao, Nguyen, Nam H., Pereira, Mayana
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.15461
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913040584146944
author Bouzid, Nassima M.
Yuan, Dehao
Nguyen, Nam H.
Pereira, Mayana
author_facet Bouzid, Nassima M.
Yuan, Dehao
Nguyen, Nam H.
Pereira, Mayana
contents LLM-based simulators offer a promising path for generating complex synthetic data where traditional differentially private (DP) methods struggle with high-dimensional user profiles. But can LLMs faithfully reproduce statistical distributions from DP-protected inputs? We evaluate this using PersonaLedger, an agentic financial simulator, seeded with DP synthetic personas derived from real user statistics. We find that PersonaLedger achieves promising fraud detection utility (AUC 0.70 at epsilon=1) but exhibits significant distribution drift due to systematic LLM biases--learned priors overriding input statistics for temporal and demographic features. These failure modes must be addressed before LLM-based methods can handle the richer user representations where they might otherwise excel.
format Preprint
id arxiv_https___arxiv_org_abs_2604_15461
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Evaluating LLM Simulators as Differentially Private Data Generators
Bouzid, Nassima M.
Yuan, Dehao
Nguyen, Nam H.
Pereira, Mayana
Machine Learning
Computation and Language
Cryptography and Security
68T07, 68P27
I.2; G.3; K.4; I.6; J.1
LLM-based simulators offer a promising path for generating complex synthetic data where traditional differentially private (DP) methods struggle with high-dimensional user profiles. But can LLMs faithfully reproduce statistical distributions from DP-protected inputs? We evaluate this using PersonaLedger, an agentic financial simulator, seeded with DP synthetic personas derived from real user statistics. We find that PersonaLedger achieves promising fraud detection utility (AUC 0.70 at epsilon=1) but exhibits significant distribution drift due to systematic LLM biases--learned priors overriding input statistics for temporal and demographic features. These failure modes must be addressed before LLM-based methods can handle the richer user representations where they might otherwise excel.
title Evaluating LLM Simulators as Differentially Private Data Generators
topic Machine Learning
Computation and Language
Cryptography and Security
68T07, 68P27
I.2; G.3; K.4; I.6; J.1
url https://arxiv.org/abs/2604.15461