Unfortunately, the issue comes from a problem with the underlying design of our lighting cluster which is not something that we can really fix.

Currently, refraction works by using reflection probe proxies to apply parallax correction. Since we don't have an explicit probe linked to each object (due to lighting being done in screen space) we choose the first probe's proxy in the pixel's current lighting cluster.
The root of the problem here is that, for optimization purpose, the repartition of the clusters along the depth of the camera depends on the max depth of each tile (the further away is the background, the bigger the clusters). This will result in different probe list per cluster which is exactly what we see here. The blocky artefacts are actually clusters with different probe list and because of that, the proxy used for refraction is different which explains the discontinuities we can see here.

We don't have a satifsying solution here so as a workaround, users need to make sure that there are enough probes and that the background is regular when looking at the refracting object. This should avoid most of these issues.