You're right in thinking that a method that can handle nonlinear
shapes is more likely to make sense of skewed predictors. But still
it can be helpful to have your errors distributed relatively evenly
throughout a response surface. That's one aspect of fit that can
improve by transforming a predictor. Another is, as you suggest, that
the measure of fit, e.g. R^2 can improve.
But you give a good reason for not transforming predictors. So you
need to weigh the benefits against the drawbacks, given your study
objectives and audience.
-Bruce McCune