This is the official codebase corresponding to the above-titled paper, which was accepted at ICLR 2026 and authored by Tommy Rochussen and Vincent Fortuin.
Any questions, reach out to me (Tommy) at 'rochussen' + 'tn' + '@' + 'gmail' + '.' + 'com'.
N.B. the primary model is referred to as the BNNP in the paper. At an earlier point in the project, it was called the BDNP. Some code within the experiments folder of the codebase still reflects this. Sorry about that.