I check the noise projection from RCAV's error signal. It is still lower than the beat frequency. So the limiting noise source is not from RCAV loop.
From previous entry, I measured the slope of the error signal for RCAV loop. The value is 0.275 MHz/V. So I can calculate the frequency noise level from RCAV loop by measureing the noise at the error point, then multiplying the noise by the slope of the error signal.
The noise level is about 2*10^-2 Hz/rtHz which is lower than the current beat noise (~0.1 Hz/rtHz).
Since the beat signal contains all noise from RCAV ACAV, and PLL loop, and I verified that the noise contributions from RCAV and ACAV loop are smaller than the measured noise. The limiting noise must come from the readout part.
So the next thing to look into is the noise in PLL loop.