Numerically robust land modeling is essential for making good predictions of climate change, yet sufficient attention is still due to the quality of numerical methods in land modules of Earth System Models (ESMs). Using CLM (Community Land Model) and ALM (ACME land model), we show through examples that the “ popular” sequential coupling approach (SCA) can lead to large predictive errors due to numerically generated pseudo physics. The first example is about the coupling between water movement in roots and soil for simulating the hydraulic root redistribution in CLM. The SCA obtained better simulation of evapotranspiration (ET) at the Blodget Forest site in California as compared to that from the tight coupling approach (TCA) and the two performed equally well for simulating ET at the Tapajos Forest site in Amazon. This result may mislead one to assert that the SCA should be preferred for calibration and prediction. However, SCA led to extreme time step size sensitivity, whereas TCA did not. In global simulations, SCA overestimated ET by as much as 3.5 mm d^-1 in some tropical sites, suggesting large land-atmosphere feedbacks could be triggered due to numerical error. The second example focused on the coupling between carbon and nitrogen dynamics in ALM. SCA assumed that the newly mineralized nitrogen is not available for plant and microbe uptake until the next numerical time step. In both site and global simulations, SCA resulted in larger nitrification rates and ecosystem nitrogen losses as compared to those from TCA that tightly couples nitrogen mineralizers and immobilizers. In particular, for long-term simulations driven by the RCP8.5 atmospheric CO₂ trajectory, SCA overestimated (compared to the tight coupling approach) global land-atmosphere CO₂ exchange by as much as 440 ppmv CO₂ equivalent over 300 years. Further considering uncertainty in initial conditions led to a difference as large as 890 ppmv CO₂ equivalent. These differences are as large as that across the CMIP5 model predictions, suggesting that land model developers need to pay as much attention to the numerical implementation of processes as they do to model structure and parametrizations.