Although the general idea of considering unlabeled
data to guide the choice of hypothesis is an
interesting idea, the method the authors use is
not very well founded theoretically. For example,
the paper says "we have found that the multiplicative
objective generally performs much better [compared
to the additive objective]" but I could not find an
explanation of why they considered it so. So,
I thought that the tests they did to check their
algorithm (only four tests for polynomials, for
example) was not enough, even considering that
they compared it with many other alternate
algorithms like REG, ADJ, etc.

Further work on the paper could therefore involve
exploring a more solid theoretical foundation for
the results found, and also testing with real-life
data. In general, the approach is very promising,
though.