Title: Calibrating Flow Matching for Simulation-Based Inference in the Presence of Model Misspecification

Abstract: Simulation-based inference (SBI) is revolutionizing experimental sciences by facilitating parameter estimation for complex, non-linear models using simulated data. However, a significant obstacle remains: model misspecification. Within a Bayesian framework focused on posterior distribution targeting, inaccuracies can stem from the simulator, noise modeling, or prior assumptions. Since these components are merely approximations of reality, substantial discrepancies can result in posteriors that are either biased or excessively confident. To resolve this, we present Flow Matching Corrected Posterior Estimation (FMCPE), a methodology that utilizes the flow matching paradigm to enhance posterior estimators trained on simulations, employing a limited set of calibration samples. The process involves two steps: initially, a posterior approximator is trained on extensive simulated data; subsequently, flow matching is used to shift its predictions toward the true posterior, which is anchored by calibration observations. By leveraging the latter to direct the correction, the method operates without needing explicit details regarding the nature of the misspecification or the specific affected model components. This architecture allows FMCPE to merge the scalability inherent in SBI with resilience against distributional shifts. Our evaluations across both synthetic benchmarks and real-world datasets demonstrate that FMCPE effectively counteracts the impacts of misspecification. It achieves superior inference accuracy and uncertainty quantification relative to standard SBI baselines, all while maintaining computational efficiency.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC