At that point I think it would just be optimal to do a distance check between each instant element center and if the distance is smaller than the radius of half of the element then treat it as collided (if the object is reasonably square shape).
Maybe a sphere collisions would use an octree calculation somewhere in the background for optimizations but a distance check might be cheaper and if these are armies then the units will probably in close groups most of the time so octree optimizations probably wouldn’t happen too often (you would need to have 1 soldier split off from the group in some way).
You could also do a box approximation of the group based on the hism bounds extent, it would need some tests to see what is optimal.
I think hism to hism collisions probably don’t occur as they were made with the thought of being an optimized masses of polygons that are supposed to be mostly background assets that don’t interact with other assets of that type.
Instancing also occurs purely on GPU so much like particle GPU collisions their calculations are tougher to do. You need a CPU collision doppelganger at that point that handles the collisions.