relationship between amp_LO and IIP3

when I simu a bipolar double banlanced mixer, the simu results represent IIP3 would be improved as the decreasing of amp_LO. In a small variable range this phenomena is reasonable, for example the amp_LO varies from -10 to 10dBm. we know in order to guarantee the efficiency of the swithing transistors, amp_LO≥2~3Vt, which is about 100mV,but when the amp_LO is -30dBm, the IIP3 is better than the result when amp_LO is -20dBm.
Why?

It's much more easy to help you if you upload a circuit diagram and plots.

it is a typical bipolar double-balanced Gilbert mixer. I simu many tines and record data. IIP3 was decreasing as amp_LO varied from -30dBm to 10dBm .so strange! when amp_LO was -30dBm the IIP3 was bigger than 9dBm!