Reference-frame-independent quantum key distribution (RFI-QKD) can generate secure keys between two remote peers with an unknown and slowly drifted reference frame. It adopts a statistical quantity C independent of the rotation of the reference frame to estimate the eavesdropper’s knowledge. However, in practical scenarios, the estimation of C should be completed with data collected in a short period of time, since the variation of the reference frame during the run will smear its correlations. In this paper, we propose to improve the performance of RFI-QKD with a new, to the best of our knowledge, statistical quantity R, which is less sensitive to statistical fluctuations and more robust to the variation of the reference frame, and obtain an analytical bound to estimate the eavesdropper’s knowledge. Simulation results show that the performance of RFI-QKD in the non-asymptotic case can be significantly promoted, which is very promising in practical QKD systems. Furthermore, we apply the new statistical quantity R to the RFI measurement-device-independent QKD (RFI-MDI-QKD) protocol and demonstrate that significant improvement can also be obtained in RFI-MDI-QKD.