Invariant-guided denoising without target supervision

Determine how to steer tabular diffusion-model denoising trajectories using invariance signals when reliable target-environment supervision is unavailable.

Background

The paper argues that conventional generative augmentation maximizes fidelity to a static source distribution, which can amplify correlations that do not generalize under covariate shift. Although guided diffusion has been studied in image domains, the paper identifies a gap in adapting such guidance to tabular data when the target environment provides no reliable supervision.

The unresolved problem is to design a principled steering mechanism that uses invariant or otherwise stable signals to guide the denoising trajectory toward task-relevant samples without relying on target-domain gradients. The proposed IGDPR framework addresses this challenge empirically through invariant potentials, but the cited passage presents the broader question as an open one.

References

Recent work on guided diffusion explores classifier or energy guidance for image domains, but tabular adaptations rarely incorporate invariance signals, leaving an open question on how to steer the denoising trajectory when target supervision is unavailable.

— Rethinking Data Augmentation under Covariate Shift: Invariant-Guided Diffusion and Prototype Reweighting  (2610.00873 - Cao et al., 1 Oct 2026) in Section 2.1, subsection “Generative Data Augmentation”