The main issue with verification is that we as humans are bad at it and seldom agree on subjective facts. Therefore we’re not good at producing valuable data at scale that then can be used to build verification methodology. A lot of research has worked to use proxy method (like pairwise comparison I.e is this better than the other) but the verification process is a very human thing and differs from one to the other so to answer your question, I see that humans must be in the loop and some part of the verification will have to be manual.
Love this perspective! Do you think we can truely trust these models, or will verification always be a manual process?
The main issue with verification is that we as humans are bad at it and seldom agree on subjective facts. Therefore we’re not good at producing valuable data at scale that then can be used to build verification methodology. A lot of research has worked to use proxy method (like pairwise comparison I.e is this better than the other) but the verification process is a very human thing and differs from one to the other so to answer your question, I see that humans must be in the loop and some part of the verification will have to be manual.