What problem does it solve?
Bayesian inference becomes difficult when every likelihood evaluation wraps an expensive forward model: a differential-equation solver, simulation, numerical integrator, waveform generator, or another substantial computation. A conventional sampler may call that model hundreds of thousands of times while treating each result as disposable.
GPry instead learns a probabilistic surrogate of the log-posterior. It uses the surrogate's predictive uncertainty to decide where another exact evaluation would be most informative, concentrating the computational budget around the posterior structure that ultimately matters.
How does it work?
A Gaussian process supplies both a local prediction and an uncertainty estimate. An acquisition rule combines those quantities to balance refinement of promising posterior regions against exploration of poorly known ones. GPry also identifies invalid or effectively negligible likelihood regions, checks convergence against posterior-level changes, and can use NORA to propose batches of acquisition points in parallel.
When is it useful?
GPry is aimed at deterministic, sufficiently smooth likelihoods in low- to moderate-dimensional problems where evaluations are expensive. Cosmological inference is one application, but the method is relevant to expensive inverse problems more broadly. It is not a universal replacement for MCMC or nested sampling: stochastic likelihoods, very high dimensions, and widely separated modes require additional care.