Go to slides 11, 35, 39, 40, 52, 53 and 55 in your first link
Go to page 16 and read brown box especially the part talking about epipolar(which plays into the third link), 24, 25, 42 in your second link
The third link is a different approach, but you “Render the scene from the camera and from the light source.” So it sounds like deferred+forward rendering and you’re still ray marching. It gives decent results, but you can’t get results like what I linked on the first page (see page 16 of your second link)
Any time there is “ray marching” going on, it’s still similar to voxelizing because the ray travels X distance, pauses, reads a sample of the surrounding area, performs a bunch of math and/or scatters rays out, travels X distance again, pauses again, reads that new data, performs more math with it and/or scatters rays out, and so on; until either the ray runs out of “energy” or it hits a solid surface that kills off the ray. Voxelizing is just a means of “LODing” the system down; to make things more efficient. Helps a lot for things that are further off in the distance that don’t need as much color/shadow accuracy. So up close, the system might use 50cm cubes, further away, it might use 100cm cubes and really far away, it might use 200cm cubes or ignore the rays all together.
So in order to pull off good volumetric lighting, you’re going to need fog/clouds/particles/etc that can react realistically to the lighting, which means you’re going to have to raymarch and/or voxelize the scene. Basically, it’s all a part of the same package.
You are mixing terms here. These volumetric techniques does use voxels but only for participating media. They don’t have to go all scene geometry and voxelize that but just approximate density of non solid objects. That can be some combination of height fog, fog volumes, noise and voxelisation of particles. VXGI is doing actual voxelisation and that is quite expensive but not any of these volumetric techniques need that.