SYNTHONY: A Stress-Aware, Intent-Conditioned Agent for Deep Tabular Generative Models Selection

Mar 31, 2026·
Hochan Son
Hochan Son
,
Xiaofeng Lin
,
Jason Ni
,
Guang Cheng
· 1 min read
Abstract
Deep generative models for tabular data perform inconsistently across datasets. We propose “stress profiling,” a method to characterize dataset difficulty along four dimensions, and introduce SYNTHONY — a framework that matches stress profiles against a calibrated capability registry of synthesizer families. Across 7 datasets, 10 synthesizers, and 3 user intents, our stress-based approach substantially outperforms random selection and zero-shot language-model baselines at identifying the best-fit synthesizer.
Type
Publication
ICLR 2026 Workshop on Deep Learning for Tabular Data (DeLTA)
Status
Peer-reviewed
publications

SYNTHONY treats synthesizer selection as a matching problem: it profiles a dataset’s “stress” — privacy sensitivity, class imbalance, cardinality, and sample size — and matches that profile against a calibrated capability registry of synthesizer families (GAN, VAE, Diffusion, DP-based, Bayesian, Normalizing Flow). It supports rule-based, LLM-based, and hybrid ranking with explainable tie-breaking and benchmark-driven comparison.

Generative AI Synthetic Data LLM Agents Tabular Data
Hochan Son
Authors
Hochan Son (he/him)
Database Architect · SRE & DevOps Engineer · ML/GenAI Practitioner
Hochan Son is a distributed systems architect and site reliability engineer who builds data infrastructure, ML pipelines, and high-scale distributed systems, working the gap between GenAI/LLM research and the systems that run it in production. His graduate research at the UCLA Trustworthy AI Lab spans synthetic tabular data generation and agentic, intent-conditioned model selection. His background spans Adtech, entertainment, and enterprise — from MySpace and Hallmark Labs to Branch.io and ADP.