Hello, this could be done via one-sample MIS and something like the balance or power heuristic, where the PDFs returned by the BSDF represent a combined probability density. Because of the one-sample MIS, no matter your choice of lobe when sampling you would be returning a combined probability density.
There is a caveat here in that arnold would not know which category of BSDF was chosen unless you tell it via the lobe masks and lobe index in the output (which is a codepath that is not very well documented and is under-utilized), so this sort of thing works best with matched categories of BSDF (ie, two glossy reflection lobes).
