Adaptive integration of synthetic and real data

Design adaptive integration strategies for combining synthetic and real datasets that balance robustness and efficiency by calibrating the contribution of synthetic data according to their reliability while preserving valid inference.

Background

Naively pooling synthetic and real observations can lead to bias when generative models are misspecified, whereas overly conservative approaches may forgo efficiency gains. Existing paradigms (synthetic data‑based vs. synthetic data‑assisted) embody different trade‑offs.

The paper highlights the need for adaptive methods that weight or calibrate synthetic information based on reliability, opening a path to guard against misspecification while still leveraging efficiency gains.

References

Moreover, a key open question in this direction is how to design adaptive integration strategies that balance robustness and efficiency, allowing synthetic data to contribute information where they are reliable while limiting their influence where they are not.

— Harnessing Synthetic Data from Generative AI for Statistical Inference  (2603.05396 - Abdel-Azim et al., 5 Mar 2026) in Section 4, Trade-offs among Validity, Robustness, and Efficiency When Integrating Synthetic and Real Data

Future work should focus on whether conflicting information in simulated and real data can also be reliably extracted via coupled tensor factorizations, for instance, through unshared factors.

— Knowledge-guided Pattern Discovery via Coupled Tensor Factorizations  (2608.13234 - Johannessen et al., 13 Aug 2026) in Section 5, Conclusion

CNEI \citep{Li2023} addresses noise robustness, but the broader problem of controlling how strongly observational information influences acquisition decisions under potential misspecification remains open. Methods that adaptively downweight observational priors as interventional data accumulate, analogous to Bayesian robustness techniques in other domains, would be valuable.

— Causal Bayesian Optimization: Foundations, Methods, and Applications  (2609.24112 - Huang et al., 21 Sep 2026) in Section 6.1, paragraph “Prior misspecification”

Preventing forgetting remains open.

— Attenuated in-context identification in time-series foundation models: diagnosis under counterfactual inputs and repair by synthetic forced-system fine-tuning  (2610.08118 - Won, 6 Oct 2026) in Appendix, Section “Additional results,” paragraph “Replay against forgetting”