Distributed d-DNNF compiler scales reasoning beyond single-node limits

Knowledge compilation transforms complex logical formulas into tractable forms like d-DNNF, but real-world problems often exceed single-node memory and time. The new distributed framework, dkc, tackles this by partitioning the search space into independent subproblems using a Cube-and-Conquer strategy. This avoids the high communication overhead typical of work-stealing architectures for circuit tasks. Alongside dkc, the authors introduce dreasoner, a distributed reasoning engine that executes core inference tasks—model counting, direct access, and uniform sampling—on the distributed d-DNNF structure, even under variable conditioning.

Experimental benchmarks confirm that dkc and dreasoner scale effectively, allowing compilation and querying of complex formulas that state-of-the-art sequential compilers cannot handle. This work opens the door to applying formal verification, probabilistic reasoning, and AI planning to much larger problem instances. The authors are from Northeast Normal University and Université d'Artois, and the paper is available on arXiv (2607.13642).