How are goodness of fit statistics calculated in curve fit module when using robust fitting of nonlinear data?

I am using the curve fit tool to fit an exponential function. When robust fitting is turned off the SSE and RMSE goodness of fit calculations agree with my fit in Excel. SSE = sum of all squared differences (fit - data). That sum is what I typically minimize in Excel solver. RMSE is the square root of SSE/N when N is the total number of data points.

Now my question is what happens when you turn robust fitting on in MATLAB, using LAR or Bisquare methods to reduce the effect of outliers. MATLAB results for SSE and RMSE go down and do not agree with the same calculation methods. I assume they are weighting the data to reflect the reduced influence of the outlier data in the robust calculations. But this means the SSE and RMSE have a different, potentially confusing, meaning when doing robust fits. I would have expected the RMSE and SSE to go up when turning on robust fitting since you can see that the residuals are in fact larger in particular spots, since we are not trying as hard to fit that bad data.

I don't see this documented anywhere. Anyone have insight into what these goodness of fit statistics mean in the robust scenarios?