I do not think having an exponent helps here, because the information is in the mantissa. Suppose you have system noise (sky noise + receiver noise) and strong RFI. In quantisation you want to keep the quantisation noise small with respect to the system noise. This then means that the mantissa needs to fit the RFI and preserve the level system noise. If you would reduce the mantissa width and increase the exponent, then the quantisation noise becomes dominant with respect to the system noise. The point with RFI is that you hope that if you separate the subbands into finer channel, that then some fine channels are RFI free. However, if you have introduced more quantisation noise due to using the exponent instead of a wider mantissa, then you introduced more white quantization noise in the whole subband, so then separating into finer channel does not yield clean fine channels anymore.
For subbands without RFI we need e.g. 4 bit complex subband samples, provided that they are equalized to their nominal level (i.e. compensated for frequency dependent gain of the recever chain).
For subbands with RFI we need enough extra bits to fit the RFI level, so that we can still hope to find finer channels without RFI within that subband.
This implies that we would need extra meta information that tells number of bits per subband. Alternative is to use one fixed width that kind of fits all subbands, e.g. 8 bit complex subband values.
Anyway, I think in each case we would transport fixed point (= integer) values, because transporting exponent bits does not add information.