The demand for accurate prediction of flow separation in modern aircraft design is becoming increasingly urgent， yet the prediction results of the widely used eddy viscosity model are not satisfactory. With its solid theoretical foundation， the Reynolds stress model may obtain more reliable results； however， its performance advantages still need to be further evaluated and explored. In this paper， the shear stress transport model and stress baseline model are selected as the representatives of the eddy viscosity model and Reynolds stress model， respectively， and numerical simulations are carried out for two-dimensional hump， two-dimensional transonic bump and three-dimensional transonic ONERA M6 wing. Compared with the experimental values， the prediction results of the two models show that the flow separation point is ahead of time and the reattachment point lags behind， while the prediction error of the Stress BSL model is smaller， showing the advantage of separation flow prediction under strong reverse pressure gradient. It is found that the two models underestimate the Reynolds stress， resulting in a large separation zone. Specifically， the Bradshaw hypothesis introduced into the SST model limits the generation of turbulent kinetic energy， reduces the eddy viscosity coefficient calculated by the model， underestimates the Reynolds stress in the boundary layer under strong adverse pressure gradient， and leads to early flow separation. The smaller Reynolds stress prediction value at the upper edge of the separation zone is considered to be the main cause for the lag of flow reattachment. For the Reynolds stress model， the error mainly originates from the inaccurate modeling of the redistribution term of the Reynolds stress transport equation. Finally， aiming at the above causes， we recalibrate and preliminarily verify the key closure parameters of the SST model. The results show that the modified model performs better than the original model.