I haven't done the comparison. But I believe it is difficult to generalize the result anyway. Keep in mind, the more accurate model may suffer the problem of over-fitting. 

.

with all the caveats that apply, since SEM and BN are very different approaches

http://homepages.inf.ed.ac.uk/amos/publications/StorkeyEtAl2007LearningStructuralEquationModelsForFMRI.pdf

http://real.mtak.hu/25494/

http://web.yonsei.ac.kr/dslab/Journal/EJOR%202008.pdf

.

What does the "results" in question refer to?

To my experience, the machine almost always finds a more probable model.

However, this may not be the way one wants to compare the models (and it usually is not what human modeller had in mind.).

i am not sure whether my answer could help. do you mean:

1- to compare BN to SEM? or

2- comparing the results  of BN to SEM?

because point 2 is almost meaningless; this is due to the fact that the answer may strongly depends on the problem in our hands.

in case of point 1 (compare SEM to BN) that's sound as a PhD project ?! because you will need to go back to the logic beyond them. in other words, you should read one of the many works of Pearl (e.g. The Causal Foundations of Structural Equation Modeling, UCLA, TECHNICAL REPORT R-370, 2012) . note also you should be aware of the works which focus on the causal meaning of the Bayesian Network (traditional metrics of BN focus on dependency rather than causality). 

i guess you may want to share us more info about the objective of your research.

regards;