Bayesian optimal experimental design is a principled way to choose experiments by estimating how much information they are expected to yield. However, the framework relies on the statistical model of the data-generating process being correctly specified. When that assumption is violated, both the resulting inferences and the estimates of information gain can become unreliable.

The authors introduce Generalised Bayesian Optimal Experimental Design (GBOED), which adapts generalised Bayesian—or Gibbs—inference to the design setting. Instead of using the likelihood in the Bayesian update, GBOED substitutes a suitable loss function. From a generalisation of Bayesian decision theory, they derive a new acquisition function called the Gibbs expected information gain.

In experiments, GBOED demonstrated enhanced robustness to outliers and to incorrect assumptions about the outcome noise distribution compared with standard approaches. This suggests that generalised Bayesian methods can make experimental design more dependable in realistic settings where models are imperfect.