CRF autoencoders (just now presented at NIPS) seem pretty neat. The central idea is to add an autoencoding layer to the typical latent-variable chain-structured conditional random field. This imposes the constraint that the model be able to accurately reproduce the input features given the latent labels, which translates to a slightly different objective when learning. It seems like a nice trick, to insist that the latent labels actually be able to generate the features.
However, that said, it does seem like a step back towards markov random fields. CRFs work so well because they are discriminitive rather than generative - no effort is wasted modeling the uncertainty over the entire problem space, and only the conditional distribution of the labels is important. CRF autoencoders seem to move more towards a generative approach, because the ability of the labels to generate the features is a factor in the objective. If a CRF autoencoder performs well on a task, how well would a markov random field work?
edit: I just spoke with author / presenter Waleed Ammar (nice guy). He pointed out that learning the corresponding MRF would be very slow, which makes sense.