Prior-data fitted networks (PFNs) are a class of transformers pretrained on simulated datasets to approximate Bayesian posterior predictive distributions. Their standard theoretical treatment assumes a separable task prior, a simplification that makes the approximation tractable.

The new paper, posted on arXiv, focuses on what happens when that separability assumption does not hold. It characterizes the 'separability gap'—the discrepancy between the ideal Bayesian behavior and what a PFN actually produces under non-separable priors.

By quantifying this gap, the work helps clarify the limits of PFN approximations and points to where the standard theory may need refinement. The abstract is brief, but the title and framing suggest a formal analysis of this failure mode.