Amalgam: Hybrid LLM-PGM Synthesis Algorithm for Accuracy and Realism

2026-03-28

AI Summary

Existing synthetic data generation methods struggle to simultaneously ensure statistical fidelity for analytical validity and realism under complex data patterns: probabilistic graphical models (PGMs) often fail to capture intricate structures, while large language models (LLMs) tend to introduce distributional skew. This work proposes Amalgam, a hybrid synthesis framework that uniquely integrates the structural modeling capacity of LLMs with the statistical faithfulness of PGMs, achieving high analytical utility, visual realism, and verifiable privacy guarantees. Experimental results demonstrate that data synthesized by Amalgam yields an average chi-squared test p-value of 91%, indicating strong statistical consistency, and attains a realism score of 3.8 out of 5—significantly outperforming the current state-of-the-art method (3.3) and approaching the realism of real data (4.7).

Abstract

To generate synthetic datasets, e.g., in domains such as healthcare, the literature proposes approaches of two main types: Probabilistic Graphical Models (PGMs) and Deep Learning models, such as LLMs. While PGMs produce synthetic data that can be used for advanced analytics, they do not support complex schemas and datasets. LLMs on the other hand, support complex schemas but produce skewed dataset distributions, which are less useful for advanced analytics. In this paper, we therefore present Amalgam, a hybrid LLM-PGM data synthesis algorithm supporting both advanced analytics, realism, and tangible privacy properties. We show that Amalgam synthesizes data with an average 91 % $χ^2 P$ value and scores 3.8/5 for realism using our proposed metric, where state-of-the-art is 3.3 and real data is 4.7.

Problem

• synthetic data generation
• probabilistic graphical models
• large language models
• data realism
• advanced analytics

Innovation

• Hybrid LLM-PGM
• Synthetic Data Generation
• Realism
• Advanced Analytics
• Privacy-Preserving