If I understand @junpenglao correctly, the way you used to provide the multimodal posterior was to provide biased prior and a very different ‘observed’ data to… confuse the inference?
With that in mind, I’ve used almost exclusively all Gaussian priors in my multidimensional linear model. And as @ricardoV94 suspected, at the end of my (very long) sampling there were some message shouting at me about the sampling was not complete. Upon closer inspection most of the posterior was also flat across multiple magnitudes, which is (correct me if I were wrong) signaling that data do not provide enough evidence for the inference process. Removing some of the redundant features resulted in 10x faster sampling as well as tighter posteriors.
Many thanks for the replies!