The analysis of land-atmosphere water and energy interactions requires reliable representation of soil moisture (SM) and evapotranspiration (ET) coupling strength (ρ). However, ρ in current land surface models (LSMs) is prone to bias and its error sources remain unclear. This study aims to identify error sources controlling LSM ρ biases. To this end, we first use a well-established algorithm to produce debiased reference ρ map based on multi-source remote sensing data. Then, ρ values from three LSMs (Variable Infiltration Capacity (VIC), Catchment Land Surface Model (CLSM), and Noah) in Global Land Data Assimilation System (GLDAS) are evaluated against our reference ρ map. Results show that all the three LSMs demonstrate strong ρ in the northern part of China (dominated by arid and semi-arid climates). In contrast, the biases of modeled ρ are marginal in the (relatively) humid southern part of China. Across the three LSMs, Noah and CLSM contain substantial positive biases in ρ representation, while VIC-based ρ estimates are significantly lower than the remotely sensed reference ρ values. Based on our analysis, it appears that the relative importance of transpiration and soil evaporation determines the overall magnitude of ρ and the inter-model ρ differences. However, the LSMs employ different schemes in land surface energy balance. Therefore, the exact process and/or parameter controlling the transpiration and soil evaporation partitioning may be model-specific.