[PHI 3000] Coherentism and the regress problem

The regress problem is a skeptical argument that goes something like this:

1. A belief B is justified only if its chain of reasons C is infinite, circular, or stops at some point.
2. If C is infinite, then B is not justified, since C involves a vicious regress.
3. If C is circular, then B is not justified, since C involves a vicious circle.
4. If C stops at some point, then any such point is arbitrary, thus leaving every subsequent point in C dependent on a starting point that cannot justify its successor.
5. (Therefore) No belief is justified.

Coherentism is usually advertised as a solution to the regress problem (usually contrasted with foundationalism). For coherentists, justification is holistic, rather than linear, and it doesn't depend on having an inferential chain of reasons with an appropriate stopping point. The metaphors that are usually invoked to illustrate the difference between foundationalism and coherentism are those of an edifice of beliefs and a network of beliefs.

According to foundationalism, beliefs at the bottom of the edifice (foundation) justify beliefs at the top.

According to coherentism, beliefs in a network of beliefs are justified insofar as they cohere (e.g., logically consistent, inferentially connected, etc.) with other beliefs in the network.

Now, does coherentism really avoid the regress problem? No matter how the relation of coherence is cashed out, it seems that the following vicious regress occurs:

p is justified if and only if it coheres with q

q is justified if and only if it coheres with r

r is justified if and only if it coheres with s...

In terms of the illustrations above, the network of beliefs could go on to infinity with infinitely many concentric circles. How can coherentism avoid this kind of regress?