DeepSWIP: Quotient-WMC Counterfactuals for Neural Probabilistic Logic Programs

DeepSWIP introduces a novel counterfactual reasoning framework for DeepProbLog programs by combining neural perception with probabilistic logic through weighted model counting. The approach achieves 2.14× inference speedup while enabling causal intervention analysis, demonstrated through experiments on visual reasoning and fairness estimation tasks.

DeepSWIP addresses a fundamental limitation in neurosymbolic AI systems: the inability to perform rigorous counterfactual reasoning beyond associational inference. Traditional DeepProbLog systems excel at combining neural networks with symbolic logic but lack proper causal semantics for interventions. This work bridges that gap by introducing Single World Intervention Programs (SWIPs) adapted for neural probabilistic logic, enabling researchers to answer "what-if" questions with formal guarantees.

The technical contribution centers on neural materialization—converting learned neural predicates into explicit probabilistic choices—followed by weighted model counting over a transformed program. This allows counterfactual queries to be resolved through a single logical program rather than constructing separate models. The quotient-WMC formulation elegantly explains why standard plug-in estimators fail under causal interventions, particularly when neural calibration degrades.

For the AI research community, DeepSWIP enables more trustworthy AI systems by supporting causal reasoning alongside perception. The experimental validation across MPI3D visual reasoning and SUMO traffic fairness demonstrates practical applicability. The 2.14× speedup over baseline DeepTwin approaches suggests the method scales reasonably for real-world deployment. The AIPW estimator results indicate the framework properly handles distribution shift and fairness concerns in downstream applications.

Looking forward, adoption depends on integration with existing neurosymbolic frameworks and demonstration on larger-scale problems. The open-source release facilitates community adoption, while the theoretical guarantees under specific assumptions provide confidence for safety-critical applications. Researchers should monitor whether this approach influences broader neurosymbolic system design.

• →DeepSWIP achieves 2.14× faster inference while enabling causal counterfactual reasoning in neural probabilistic logic programs
• →The approach transforms neural predicates through materialization, reducing them to standard probabilistic choices for exact counterfactual computation
• →Experiments confirm the method handles neural calibration degradation and removes first-order bias in fairness estimation tasks
• →The quotient-WMC formulation explains intervention semantics and identifies failure modes in standard neural inference approaches
• →Open-source implementation available, lowering barriers for adoption in neurosymbolic AI research