Post

What we need is a kind of recurrent model that can also estimate the loss of the current answer that we can refine until the estimated loss is sufficiently low.