Reconstructions of prehistoric vegetation composition help establish natural baselines, variability, and trajectories of forest dynamics before and during the emergence of intensive anthropogenic land use. Pollen–vegetation models (PVMs) enable such reconstructions from fossil pollen assemblages using process-based representations of taxon-specific pollen production and dispersal. However, several PVMs and variants now exist, and the sensitivity of vegetation inferences to PVM selection, variant, and calibration domain is poorly understood. Here, we compare the reconstructions, parameter estimates, and structure of a Bayesian hierarchical PVM, STEPPS, both to observations and to REVEALS, a widely used PVM, for the pre–Euro-American settlement-era vegetation in the northeastern United States (NEUS). We also compare NEUS-based STEPPS parameter estimates to those for the upper midwestern United States (UMW). Both PVMs predict the observed macroscale patterns of vegetation composition in the NEUS; however, reconstructions of minor taxa are less accurate and predictions for some taxa differ between PVMs. These differences can be attributed to intermodel differences in structure and parameter estimates. Estimates of pollen productivity from STEPPS broadly agree with estimates produced for use in REVEALS, while comparison between pollen dispersal parameter estimates shows no significant relationship. STEPPS parameter estimates are similar between the UMW and NEUS, suggesting that STEPPS parameter estimates are transferable between floristically similar regions and scales.