Gm has a peak and this is usually a little bit below VT. Which
direction you'd go, depends on where you started.
Increasing gm helps kill the backend offset contribution,
whether it improves matching significantly in the diff
pair itself I don't think is the issue. I rarely if ever see
mismatch data besides VT-match, and there only for
a single or very limited number of current densities.
So I wouldn't be very trusting of how well modeled an
operating point dependent mismatch statistics bundle
might be.