Theoretical computer science conferences usually have a review period of about two months. Programming committee members have potentially dozens of papers to review, alongside their regular day jobs. How thoroughly are papers usually reviewed? What kind of heuristics are often employed to review papers faster? Does correctness of every proof get verified in detail, or do reviewers put some trust into the authors getting it right?
It varies. As a first year PhD student I was asked to review a theory paper, even though I had almost no background in the field. The professor explicitly said I only had to read the first half of the paper, and I only had to tell him if it looked interesting and correct.
At the same time I'm sure many reviewers do a diligent job and don't outsource their reviewing to inexperienced graduate students.
If a PC member has dozens of papers to review, then either a) the conference is doing it wrong or b) the PC member is expected to outsource most of the reviewing work to external reviewers. A good heuristic for not over-burdening the reviewers is to expect each reviewer to handle no more than 3-6 papers. With those sort of numbers, it's entirely reasonable to expect that each paper will get reviewed in great detail.
In practice, however, a great degree of variability in the quality of reviewers. Some reviewers will do a very thorough job---my theoretical papers often get back detailed comments from at least one reviewer that makes it clear that they worked through all of the math. Everybody in academia also has "bad reviewer" stories about people who didn't take their job seriously. For example, I recently saw a reviewer provide a review which, in its entirety, said, "Accept as talk." Good conferences with responsible program chairs try to mitigate these effects, but there's always a good deal of noise in the process.