ADRF6720-27 carrier frequency stability

Dear,

I have an ADRF6720-27 EVAL-Z board and an ADRF6820 EVAL-Z connected together through RFout -> RFin. When IQ inputs of ADRF 6720-27 are generating the sequence (-1,-1), (-1,1), (1,1), (1,-1), (1,1), (-1,1) cyclically, where 1 represents 3.2V on the positive input and 2.2V on the negative input, and -1 represents 2.2V on the positive input and 3.2V on the negative input, I obtain the good constellation at the output of ADRF6820 when both cards share the same local oscillator. The following figure shows the received unwrap angle in the Niquist plane.

 Amplitude is in radian and 1 point represent 1.4 microseconds.

However, when the two cards have their own local oscillator, the frequency difference at reception vary drastically when I or Q change. Instead of the former figure, I obtain

 while I was expecting something like 

The decreasing trend is due to the frequency difference between local oscillators (not the same RFIN), but a more or less alternating signal is superimposed on it. The period of this alternating signal is exactly the same as the duration of each IQ state. In order to understand theses variations I have increased by 10 the duration of each IQ state (notice that I removed the high pass filters at IQ ADRF6720-27 inputs). I obtain

Here, there is 3 IQ transitions. One sees that a transition is accompanied with a phase dumped oscillation, turning the phase by almost 10rad (almost 2 turns) ! I filtered out IQ inputs so the transition is more then 400ns. Since the signal is perfect when I use the same LO oscillator for ADRF6720-27 and ADRF6820, I suspect that the phase variation when IQ state changes is due to a destabilization of the ADRF6720-27 PLL.

My questions are : What is the problem? Can is be corrected ? If not, how to manage to recover my states ?

Thank you,

Stéphane