Distribution shifts from sim2real gaps and controller–abstraction coupling

Address distribution shifts in statistical verification and control caused by (i) simulation-to-reality gaps and (ii) couplings between controllers and statistical abstractions that induce dependence and shift between calibration and deployment data; develop methods to quantify, compensate for, or guarantee safety under such shifts.

Background

Conformal prediction relies on exchangeability between calibration and test data. In practice, calibration often comes from simulators, leading to sim2real gaps, and control policies may affect the data used to build abstractions, creating couplings that break independence.

The authors emphasize dealing with these shifts as core open problems to enable reliable formal verification and control in realistic deployments.

References

Some of the open problems here, as already discussed in this article, are to deal with distribution shifts caused by: (1) sim2real gaps that arise in practice, and (2) couplings between the controllers and statistical abstractions and hence in dealing with distribution shifts.

— Formal Verification and Control with Conformal Prediction  (2409.00536 - Lindemann et al., 2024) in Section 7, Open Problems and Future Directions

By \cref{def:scope} a certificate holds over one mission distribution and one set of model versions. Nothing here says how fast it decays as a deployment drifts, which is the quantity an operator actually needs to know when deciding how often to recertify.

— Agent Behavioral Contracts II: Certifying Compositional Reliability Without Assuming Independence  (2608.12895 - Bhardwaj et al., 13 Aug 2026) in Section 7, “Future Work,” paragraph “Certificate decay”; see also Section 6, “Limitations”

The evaluation is in simulation. The sim-to-real gap for language-model control of this kind has not been quantified in any published study, and is not addressed here.

— Large reasoning models for abnormal situation management in safety-critical industrial processes  (2608.19819 - Alhazmi, 20 Aug 2026) in Discussion, paragraph beginning “Several limitations bound these claims.”
  1. Or architect for it. Where the sampling process is under your control, exchangeability can be manufactured, at a price paid in coordination or adaptivity. Quantifying that price across settings is open.
— The Exceedance Design Effect: Effective Sample Size for Thresholds under Clustering  (2608.21262 - Noonan, 21 Aug 2026) in Section 8, item 8, p. 29

The open question, which we return to in Section~\ref{sec:challenges}, is whether internalized control remains verifiable and reliable under distribution shift.

— Complex Problem Solving in Large Language Models: A Statistical Control Survey and Diagnostic Framework  (2609.20973 - Cai et al., 17 Sep 2026) in Section 2.3, Scope and Limitations of This Survey; revisited in Section 5.6

While the occasional failures of the learned policies are a limitation of the method, several ways of fixing it remain open for future work, such as shadow backup copies, the addition of history, or, most directly, removing the information asymmetry by adopting a more powerful input policy.

— Predictive Zonotope Reduction: Precise Runtime Monitoring under Uncertainty  (2609.03699 - Krsmanovic et al., 3 Sep 2026) in Section 4, subsection "Results and Analysis," paragraph "Learned Policies can learn the control behavior with minimal inputs"

Our method is established under the i.i.d. assumption, whereas practical applications may encounter distribution shifts. We leave the investigation of this extended setting to future work.

— Occupancy-based Quantile Risk Control  (2609.03104 - Shi et al., 2 Sep 2026) in Section 6, Conclusion, paragraph "Limitation" (Section \ref{sec:conclusion})

Finally, the subspace is estimated from non-speech calibration data. Although it transfers unchanged from ESC-50 to held-out ESC-50, UrbanSound8K, and FSD50K, broader calibration-to-deployment shifts remain to be evaluated.

— Reducing Hallucinated Transcripts in Whisper via Hallucination Space Projection  (2609.04561 - Abbasihafshejani et al., 3 Sep 2026) in Limitations section