As Land Surface Models are increasingly applied at higher spatial-temporal resolutions, understanding how land heterogeneity influences the small-scale partitioning of precipitation into runoff is crucial for advancing Earth System Models. In this study, we use the HydroBlocks modeling framework to investigate runoff-streamflow relationships over the CONUS. Our overall objective was to understand via uncalibrated LSM experiments, how the representation of land heterogeneity impacts runoff generation at an effective 90-meter resolution and how this in turn changes simulated streamflow accuracy in natural basins from the CAMELS dataset. We conducted experiments by both increasing the level of heterogeneity and homogenizing key aspects of land heterogeneity, such as precipitation, land cover, and soil properties. The results reveal that basins with high spatial runoff variability—particularly in the Northwest, Northeast, and Southeast, which typically receive more precipitation—generally improved their streamflow accuracy when heterogeneity was increased. However, simply increasing model granularity does not guarantee satisfactory accuracy in many basins, especially in the Central North and South, where a high positive bias in meteorological inputs leads to unrealistic rainfall events. Clustering the input heterogeneity metrics in these transitional climate zones revealed that basins can respond differently to land surface homogenization: while homogenizing soil properties may improve model accuracy in some basins, homogenizing land cover can degrade performance. Our comparison with original CAMELS attributes shows that runoff spatial properties can be positively or negatively correlated with basin characteristics, suggesting that these spatial properties are fundamental to understanding why low-complexity models sometimes outperform more detailed approaches in certain regions.