Many microstructural changes in a structural material are engaged with the kinetic behaviour of grain boundaries (GBs). Thus, the correct description of the kinetic characteristics of these defects plays a crucial role in the development of physical-informed models of the material evolution. One of the possible ways to reveal the key details of defects properties is the classical atomistic simulation. This work is focused on temperature-induced order-order and order-disorder transitions of GBs and impact of these transitions on diffusion and mobility characteristics. The study was carried out for several metals (Ni, Fe, Nb, Mo and W) with use of new reliable interatomic potentials developed with machine-learning-like force-matching technique. The simulation revealed that atomic self-diffusion along tilt GBs at low temperature (T < 0.6Tm) is mostly controlled by migration of self-interstitial atoms. However, heating leads to a change in the GB diffusion mechanism to a more complex exchange process, not associated with a specific defect, but similar to diffusion in a liquid. This change is associated with disordering complexion transition of GBs, which also significantly affects the GB mobility.