@xntang88 could you share the dataset you used for your test by email?
The difference could be random (restarting the search in the old version might sometimes take longer too), as there were few changes to the core algorithm between 2.16.1 and 2.23.1, but it would be good to make a test to be sure.
I also found that the prediction curve on the solution plot for the new version (V3.0, like v2.23.1) is not smooth on the mode of either 'Show cross-validation error' or not, but when 'Show combined data (training + testing)' is selected, the prediction curve becomes smooth.
Could you send here before/after screenshots to show what you mean?