Upon further reading of the documents, I noticed I made some mistakes. It looks like "Split/Second" just uses Lighting Volumes for diffuse Global Illumination. From what I can tell, their dynamic specular lighting is supplied by local lights, while the cubemap reflections are forward rendered for any objects that need metallic reflections.
To offer some clarification, a true Irradiance Volume would have an array of blurred environment cube maps for each 3D grid cell. These maps would represent rough-to-smooth specular lighting, which materials would sample via a gloss value. Lighting Volumes are an approximation of this, having only one extremely low-res cube map per grid cell to represent rough specular lighting. So if I want my materials to have dynamic specular lighting consistent with the Lighting Volume's contribution, I must use a constant gloss value for all my materials.
Lighting Volumes are basically 3D texture irradiance volumes, very similar to Crytek's Light Propogation Volumes. As such, they carry many of the same benefits, including easy compatibility with deferred lighting. However, they are fundamentally different in that they are precomputed for offline storage, and use a variation of Valve's ambient cube basis rather than spherical harmonics. Nonetheless, they are apparently quite effective and have already shipped in a few titles, most notably F.E.A.R 2 and Split/Second.
Despite my general apathy toward precomputed lighting, I am intrigued by this technique. Granted, it will never be comptetitive with lightmaps in terms of resolution, lighting quality, etc, but that isn't the point. My deferred lighting renderer already handles high-resolution direct lighting well enough, and doesn't need any assistance in that area. Instead, these will handle high-quality ambient lighting, a task which deferred lighting is not well-suited to handle.
Now this system does raise a few concerns, such as how I should handle specular lighting. As I understand it, F.E.A.R.2 achieved a glossy specular by just reusing the volume data with a reflection vector rather than a surface normal. Whereas Split/Second appears to be using a separate cubemap, so that they can get high-res metallic reflections. Since each approach has different merits, I will need to try them both to see which one will best meet my needs.
Of course this system also raises questions of what else I can do with the volume data, which warrant further investigation. For example, I can use the volume to provide a reasonable source of environment lighting for translucent objects/particles. I'm also wondering if I should take it as far as Crytek, and try to do colored fog using the volume. Needless to say, there are certainly plenty of fun possibilities with this system that demand to be explored.
One thing that is likely to be a problem is that, much like most other forms of IBL, I need to use world-space inputs when sampling the data from one of these volumes. So with my current renderer that means I not only have to unpack my normals, but also transform both the normals and the recovered positions from view-space to world-space. Something that will likely be a fullscreen pass, since Lighting Volumes will be needed for each part of the scene.
Yet another reason why I am strongly considering dumping packed normals + gloss in favor of raw world-space normals.