First off what does grain size mean? It refers to two different things. One is the size of the carbides, and the other is the size of the bits that make up the main body of the steel. Steel is a crystal, however it isn't uniform along its entire length, but rather on a small scale (~one micron), there are groups of atoms that are lined up in a certain way, homogeneous orientations. These little groups are the grains.
When a steel is heated up past the critical point the ferrite + carbides (usual state) will transform to austenite. This is a transformation from one crystal structure (body centered cubic) to another (face centered cubic). The austenite structure has a more wide open distribution and thus can dissolve much more alloy, and this is what happens when the transformation takes place. The longer you keep the steel in the elevated temperature the more the carbided alloy bits (grains) dissolve into the austenite. Note the austenite grains will grow as time passes which is why you have to be careful about soak times and temperatures.
Anyway, if you look at the steel in this state you will see carbides (grains) disappearing, as well as big hunks of austenite that are clearly in groups (grains) which are expanding (grain growth). It looks like graph paper, with highly irregular spacing and a bunch of big marker dabs. When you cool the steel slowly the process reverses. The austenite transforms back into ferrite and the alloys all carbide out in hunks. Depending on the speed and where you hold it you can normalize, fully anneal, produce Pearlite, Bainite, Troositite, Sorbite etc. .
However if you do the cooling very fast the transformation from austenite to ferrite is so rapid the alloys don't have enough time to carbide out into nice little clumps (normalize) or long strings (pearlite), but instead there is a very explosive transformation to martenite which is yet another crystal structure (body centered tetragonal). This is not a stable state, it will quickly decay. This is why you temper which locks it into ferrite plus carbides, very similar to normalized steel, except the carbides are much smaller (order of magnitudes). Ok, what does all of that have to do with grain size? Well when you change crystal structure it doesn't happen all at once in a nice uniform manner. It happens very explosively at certain very specific spots (grain boundaries and points of crystal dislocation). These new crystals "grow" until they smack into other ones. Thus the more you start, the smaller the resulting overall grain (average size), when every thing is done. From a basic viewpoint this is one of the aspects of how deep cryo refines grain size (it also starts the tempering process). As well, it also produces finer grained (due to their nature) higher wear resistant carbides.
This by the way is also why forging can cause a finer grain structure. It doesn't actually refine the grain of the steel while its being hammered on. Note during the hammering, the steel is not even in the state that it will be when its ground into a knife. However when you transform the steel by heating and quenching, because it was hammered on, there will be a greater number of crystal transformation points and thus a lower grain size. In technical terms the plastic deformation increased the dislocation density in the steel and thus increased the nucleation sites for the phase transition.
This is also why multiple quenching can produce a finer grain size, when you transform from martensite to austenite, as opposed to ferrite to austenite, there are more transformation points and thus a finer grain structure. When you quench again, the finer austenite grain produces finer grained martensite. This process can be repeated again, as the finer the grain of the initial state, the finer the grain of the produced state, and so it is an iterative process.
once cooling occurs and the steel has become either martensitic or some form of
In some steels you can have ~25% of austenite retained if you don't go to sub-zero temperatures. Since cryo promotes the increase of further martensite transformation, this will directly decrease the growth of the grains already formed (and those forming). This effect of grain refinement by containment, is enhanced by the increase in dispersal of the carbided out alloy segregates which will act to do the same thing, which by their very nature are also finer grained.
When the crysal structure is transformed yet again through tempering, the greater percentage of martensite transforming and increased carbide precipitation and dispersal will insure a lower grain size in the final result. As well, the very low temp of deep cryo (liquid N2) may inhibit grain growth directly. I have only seen this referenced in a vague manner, usually noted as keeping the steel in a stable state before further experimentation has to be done (at a later time).
So in short, it doesn't actually shrink the grains already formed, but prevents them from growing, causes the formation of very fine grained carbides, and causes a finer grain to result when the steel is tempered due to the previous factors and the greater percentage of martensite transforming.