It sounds like you're mostly interested in specular reflection probes here, which is typically a completely different system than the “irradiance” probes that you mentioned in your OP. Specular probes are almost always sparse and hand-placed, with a few exceptions (the automated probe generation process for Quantum Break comes to mind). Probes for diffuse (and sometimes low-frequency specular) are typically much more dense than specular probe cubemaps since they consume significantly less space: like I mentioned previously for diffuse you can get away with a dozen floats or so that can be compressed even further. Thus you trade angular frequency for spatial density, or vice versa. For diffuse probes it makes a lot more dense to use grids or other data structures since they tend to be so dense, and so you need some structure where's it's fast to find the relevant probe(s) for a given pixel/surface. For specular probes you can't store nearly as many and so they're almost always sparse and unstructured. Often there's some kind of offline process for “assigning” specular probes to static meshes, and other tricks (dynamic ambient/specular occlusion, proxy occluder volumes, influence volumes, hybrid screen-space reflections or even ray-traced reflections, etc.) to make sparse probes look halfway-decent. Ultimately though it can be really tough since they're so sparse spatially.