With the growing network bandwidth demands of video streaming, it has become increasingly common to utilize multipath for Quality of Experience (QoE) improvement. The recent ABR algorithm Chorus presented at MobiCom estimates the equivalent bandwidth of all paths to guide the bitrate decision and cross-layer coordination, achieving significant QoE improvement. However, the equivalent bandwidth prediction is based on the slowest path and non-representative historical bandwidth samples. Therefore, the equivalent bandwidth is always underestimated, and thus Chorus always selects bitrate conservatively, degrading the QoE. To improve the accuracy of the equivalent bandwidth prediction, the Crossformer-based Bandwidth Prediction (CBP) is developed. Specifically, CBP obtains fine-grained historical bandwidth samples across multipath via the synchronized sampling mechanism and linear interpolation. Moreover, CBP incorporates stable segmentation processing and path-aware attention fusion into the Crossformer Model architecture for bandwidth prediction. In this way, the temporal stability of bandwidth in each path and the cross-path dependencies are well modeled for bandwidth prediction. Evaluations with public datasets show that CBP helps Chorus significantly improve the prediction accuracy, i.e., reducing the mean squared error by \(14.2\%\sim 42.8\%\) , thus improving the QoE by 11.4%, 21.5%, 5.9%, and 27.3% on the FCC, Norway, Oboe, and 4G datasets, respectively.