That something has gone substantially wrong in one of your groups? Or not: you don't actually state what "vastly differing" represents in numerical terms.

This is sort of the point of using reference genes: we cannot be certain that RNA extraction, RNA integrity and cDNA synthesis efficiency will be identical under all circumstances, not to mention all the potential error from pipetting during all those processes. Consequently, we use reference genes to normalise these variables.

Your reference genes do not need to give near-identical values every time: I mean, it's NICE if they do, because it's a fine validation of your pipetting etc, but ultimately, if your reference genes always gave near identical values, you wouldn't really need to use reference genes at all. You use reference genes because you CANNOT make this latter assumption. If your reference genes show marked differences, this can just be normal behaviour: if one sample happens to have ten-fold lower total cDNA than another because...reasons, reference genes will handle this and allow that sample to be brought in line with the others with respect to your GOIs. If one sample happens to have 100-1000 fold lower total cDNA than another because...weirder reasons, then reference genes will again handle this, and allow you to most likely discard this sample because it is obviously suspect.

If your reference genes indicate you have substantial and systematic differences between control and treated groups, then either something went wrong in one or other of those groups, or your reference genes are not actually functioning as such, and you should find better reference genes.

To advice further I would need to know

A) how you determined these were suitable reference genes

B) how many samples you have in each group

C) what your typical Cq values are (ballpark is fine)

D) what you consider to be 'vastly differing'

E) what is the corresponding behaviour of your GOIs.

Is there something different about the starting materials in the 3rd occasion?

Is it a different type of tissue, different timepjoint, different genetic strain, etc.?

It's entirely biologically possible for samples to have very different levels of gene expression.

Are your inputs and dilutions the same? I.e. for cDNA and PCR dilution? Assuming you have the same inputs, and the expression levels of the reference gene is stable across your groups (which has to be shown empirically), then I would suspect differences in input. How did you come to this conclusion, did you do group averages?