TARDIS Framework Closes Synthetic-Real Gap in Tabular Data Generation
A novel framework for refinement during inference, named TARDIS (Tabular generation through Refinement, Distillation, and Inference-time Sampling), has been developed to bridge the gap between synthetic and real tabular data generation. Unlike earlier methods that concentrated on enhancements during training, TARDIS utilizes a fixed pre-trained backbone and employs a Tree-structured Parzen Estimator to guide score-level searches in reverse diffusion. Additionally, the framework features post-hoc sample selectors and an optional step for soft-label distillation. This innovative approach achieves cutting-edge results in the utility of synthetic tabular data, even outperforming real data in certain instances. The findings were shared on arXiv with the identifier 2605.06261.
Key facts
- TARDIS is an inference-time refinement framework for tabular data generation.
- It operates on a frozen pre-trained backbone.
- Uses Tree-structured Parzen Estimator search over score-level guidance.
- Includes post-hoc sample selectors and optional soft-label distillation.
- Closes the synthetic-real gap in tabular diffusion models.
- Published on arXiv with ID 2605.06261.
- Achieves state-of-the-art synthetic tabular data utility.
- Surpasses real-data utility in some cases.
Entities
Institutions
- arXiv