Coherence scores (u_mass) for LDA models very volatile when varying the number of topics

Question

Why does coherence vary so much as number of topics change?

I am using Gensim's coherence model to calculate u_mass coherence scores for a variety of Latent Dirichlet Allocation (LDA) topic models which vary the number of topics (k). My aim is to optimise k.

I understand that in general, the higher the u_mass coherence score the better, so k should be selected where coherence is maximised (according to sources here, here and here). However, I find when I plot the coherence scores the graph is highly volatile. This is also reflected in other examples I've found online:

My graph

Another example

Another example

Can someone explain why the coherence varies so much as k changes? If I understand it correctly - if k increases and the new topic has words which have low co-occurrence, the global coherence score falls significantly. If k changes again but the new words grouped by the topic have high co-occurrence, the global coherence score increases. This is why we see so much variation in coherence to changing k.

A small additional question; is k optimised where we see local peaks in coherence?

Any help or additional information is greatly appreciated. Thanks!