Inferential frameworks that propagate synthesis uncertainty

Develop general statistical frameworks for downstream estimation and inference that explicitly model and propagate uncertainty introduced by synthetic data generation, ensuring valid uncertainty quantification when synthetic observations are used alongside or in place of real data.

Background

Synthetic data are produced by estimated generative models and thus introduce additional sources of variability and bias beyond classical sampling noise. Most existing approaches that treat synthetic observations as equivalent to real data systematically misestimate uncertainty.

The paper calls for general inferential frameworks that explicitly account for this synthesis‑induced uncertainty, noting that such methods remain largely undeveloped despite growing reliance on synthetic data in downstream analysis.

References

Developing general inferential frameworks that explicitly account for synthesis uncertainty in downstream estimation and inference, therefore, remains an important open problem.

— Harnessing Synthetic Data from Generative AI for Statistical Inference  (2603.05396 - Abdel-Azim et al., 5 Mar 2026) in Section 4, Uncertainty Propagation from Data Synthesis

Deriving an asymptotically valid analytic variance estimator remains an open question for future research, because uncertainty passes through several estimated components, including target-state simulation and classifier-based weighting.

— From Cumulative Weights to Marginal Density Ratios: Per-Protocol Estimation in Sequential Target Trial Emulation  (2608.20976 - Ke et al., 21 Aug 2026) in Section Discussion

Several directions remain open. First, the synthetic pre-training framework could be validated across multiple tokamak devices to assess cross-machine transferability. Then, integrating uncertainty quantification into the reconstruction pipeline would enable confidence-aware equilibrium inference suitable for real-time control applications.

— Real-time Tokamak Equilibrium Reconstruction Under Limited Experimental Data via Physics-Grounded Synthetic Pre-training  (2610.09674 - He et al., 7 Oct 2026) in Section 6, Conclusion