Since caustics are the noisiest part of the image once turned on, naturally you would want to set some setting to number crunch caustics more than the rest of the image.
As I understand, caustics are calculated with the every pass.
So If you have noise from caustics, you should lower AA vs GI to get more Passes done in the same render time, just like with DOF and MoBlur.
I tested it and it seems to be the case. Both images had a 10 minute limit. GIvsAA = 4 managed 76 Passes and GIvsAA = 32 managed 31 Passes. GIvsAA 4 was cleaner than 32.
So noise from caustics = lower GIvsAA.
Is my notion correct? Makes me recall that Dynamic GIvsAA on the roadmap...