At high enough true strain degrees (starting from about 1.5) in COLD deformation  the apparent "grain" becomes confusing  because of fragmentation effect i.e. gradual formation of fresh high-angle boundaries. While we can filter off this effect, the visible grain shape follows macroscopic deformation (Taylor's principle). In this case the boundary area evolution depends on BOTH the deformation mode AND the initial grain shape (Eq.31 in the paper "shape-dependent compatibility..." at my RG page. For metals the volume change should be excluded there.. ).

As to special effect of torsion, unlike other  modes you mention, this deformation is always non-uniform (increasing from the axis to periphery of your sample) and you should separately consider different domains of the sample. Owing to work hardening, the plastic strains eventually penetrate the axial part, however IN AVERAGE the deformation effects develops slower than in other (quasi-uniform) modes.

 Apart from the above reasoning, your plot still shows a too drastic peculiarity of torsion that suggests an apparent violation of Taylor's principle. This could happen, in particular, if the grain boundary sliding notably contributed to the macroscopic deformation. The latter effect depends on homologous temperature of course (ratio of deformation temperature to the melting point  of your metals). The only alternative explanation is that in torsion (that is LOCALLY simple shearing) the fragmentation much faster attain the STABLE LEAST size of fragments (taken for "grains"). This version wants proofs however...