RoBAS: Robust Bayes-Assisted Conformal Prediction reduces interval widths

A new paper by Ashouritaklimi, Cortinovis, and Caron, accepted to AISTATS 2026, tackles a key limitation of Bayes-assisted conformal prediction: while conformal methods guarantee valid coverage even under model misspecification, prediction set efficiency degrades when the Bayesian working model (BWM) poorly aligns with observed data. The authors propose RoBAS (Robust Bayes-Assisted Shrinkage), a framework that constructs robust nonconformity scores with two instantiations. The first uses a heavy-tailed BWM to down-weight extreme prior information; the second is a closed-form empirical Bayes shrinkage score. Both scores automatically adapt: when prior information is reliable, they exploit it to produce tighter prediction sets; when it is weak or inaccurate, they revert to a Distance-To-Average (DTA) baseline, ensuring robustness without losing conformal validity.

The researchers evaluated RoBAS on tabular and image regression tasks where training distribution may differ from calibration/test distributions (while calibration/test remain exchangeable). Results show that RoBAS scores are competitive with widely used nonconformity scores when no distribution shift occurs. Critically, under shifted settings, RoBAS substantially reduces interval widths—offering more informative predictions without sacrificing the distribution-free coverage guarantees that make conformal prediction attractive for safety-critical applications like medical diagnosis or autonomous systems. The framework provides a principled way to leverage Bayesian priors when they help, while automatically safeguarding against harmful prior misspecification.