RMS gradient

 Hi everyoone,
 Just a quick question,
 When performing geomtry optimisations via GAMESS
 (semi-empirical/ab-initio), the RMS gradient gradually decreases as
 the calculation progresses, however even though the equilibrium
 geometry is eventually located, the RMS gradient every so often
 'jumps up'.
 Can anyone enlighten me as to why this occurs?
 Much appreciated