The standard reference for the technique the engine is built on. It turns a probabilistic question into a counting problem and compiles the model once, so exact answers come back where textbook methods run out of time and memory.
Every standard method for exact reasoning in a Bayesian network shared one ceiling. Cost grew exponentially with a structural property of the network called treewidth, in the best case as well as the worst. As real problems grew, that ceiling pushed most of the field toward approximate answers instead.
The paper reframes the question as weighted model counting. Write the network as a logical formula with a weight on each variable, and the probability you want becomes the total weight of the formula's solutions. That reframing exposes structure the classic methods ignore, such as parameters that are simply zero and dependencies that only hold in some contexts.
The paper organises the approach along three choices, how the network is encoded, whether the counting is done by search or by compiling the formula once, and how much of that local structure the encoding captures. Its experiments answered queries on networks the standard algorithms could not, including networks whose treewidth put them out of reach entirely, and showed compiled models answering orders of magnitude faster than search.
It is the standard reference for the technique PiLogic's engine is built on, and its first author is PiLogic's co-founder and CTO. The engine, now in its fifth generation, takes the same approach. It also names the two ideas that run through everything else on this page, compile once and exploit the structure real systems have.
The advantage depends on a model having local structure to exploit, and the paper says so plainly. Real engineered systems usually do. The paper restricts itself to exact inference and exact model counting.