Found a nice little collection of test scenes: http://graphics.cs.williams.edu/data/meshes.xml (there are also commonly used images for postprocessing effects and such)

And I think the Global Illumination using radiance hints paper is missing in the GI links. The technique seems to create good (although quite low res) results, is fully dynamic and I have hopes that it could actually be used for games: http://graphics.cs.aueb.gr/graphics/docs/papers/RadianceHintsPreprint.pdf
There is also the presentation: http://www.highperformancegraphics.org/previous/www_2011/media/Papers/HPG2011_Papers_Papaioannou.pdf

The only real alternative seems to be voxel cone tracing, with the big disadvantage, that it needs the precomputed voxel data (which one can update in realtime, but preferably only for dynamic things close to the camera, at least that is what unreal engine 4 does according to some paper, which is probably already in hummels post).

To support many lights I am currently quite a fan of clustered shading, which splits the viewfrustum in tiles and adds the depth also. The disadvantage is, that you get many small frustums for which has to be checked which light is visible, but there are some approaches to make it fast (my current brute force approach isn´t with about 25ms :P). The big advantage over tiled shading (the same without the depth part, but instead restricted to the actual visible lights) is, that the min max check is not needed for a good performance in complex scenes. The problem with it is, that it needs the scenes depth before the final rendering with lights AND it has to read the data on the cpu for the same frame, which is a serious cpu stall, especially with multithreaded opengl and double buffering, as it not only renders the depth there but also finishes rendering the lighting of the previous frame.
There is some nice presentation on clustered lighting on this page called "Practical clustered deffered and forward shading": http://advances.realtimerendering.com/s2013/