We theoretically justify the recent empirical finding of Teh et al. (2025) that a transformer pretrained on synthetically generated data achieves strong performance on empirical Bayes (EB) problems. We take an indirect approach to this question: rather than analyzing the model architecture or training dynamics, we ask why a pretrained Bayes estimator, trained under a prespecified training distribution, can adapt to arbitrary test distributions. Focusing on Poisson EB problems, we identify the existence of universal priors such that training under these priors yields a near-optimal regret bound of $\widetilde{O}(\frac{1}{n})$ uniformly over all test distributions. Our analysis leverages the classical phenomenon of posterior contraction in Bayesian statistics, showing that the pretrained Bayes estimator adapts to unknown test distributions precisely through posterior contraction. This perspective also explains the phenomenon of length generalization, in which the test sequence length exceeds the training length, as the model performs Bayesian inference using a fractional posterior.

We consider the problem of empirical Bayes estimation for (multivariate) Poisson means. Existing solutions that have been shown theoretically optimal for minimizing the regret (excess risk over the Bayesian oracle that knows the prior) have several shortcomings. For example, the classical Robbins estimator does not retain the monotonicity property of the Bayes estimator and performs poorly under moderate sample size. Estimators based on the minimum distance and non-parametric maximum likelihood (NPMLE) methods correct these issues, but are computationally expensive with complexity growing exponentially with dimension. Extending the approach of Barbehenn andZhao (2022), in this work we construct monotone estimators based on empirical risk minimization (ERM) that retain similar theoretical guarantees and can be computed much more efficiently. Adapting the idea of offset Rademacher complexity Liang et al. (2015) to the non-standard loss and function class in empirical Bayes, we show that the shape-constrained ERM estimator attains the minimax regret within constant factors in one dimension and within logarithmic factors in multiple dimensions.

Consider a pair of input distributions which after passing through a Poisson channel become ϵ-close in total variation. We show that they must necessarily then be ϵ0.5+o(1)-close after passing through a Gaussian channel as well. In the opposite direction, we show that distributions inducing ϵ-close outputs over the Gaussian channel must induce ϵ1+o(1)-close outputs over the Poisson. This quantifies a well-known intuition that "smoothing" induced by Poissonization and Gaussian convolution are similar. As an application, we improve a recent upper bound of Han-Miao-Shen’2021 for estimating mixing distribution of a Poisson mixture in Gaussian optimal transport distance from n−0.1+o(1)to n−0.25+o(1).