Scalable integration of logic programming and neural networks

Develop scalable frameworks that integrate traditional logic programming with neural networks to support complex reasoning tasks in Neuro-Symbolic AI.

Background

The survey highlights systems that blend probabilistic logic, neural components, and logic programming for reasoning. The authors explicitly mention that scaling such integrated frameworks for complex tasks remains open.

References

Open research questions remain around how Neuro-Symbolic AI can develop scalable frameworks that integrate traditional logic programming with neural networks for complex reasoning tasks, incorporate commonsense knowledge and advanced language understanding to enhance multi-hop reasoning capabilities, combine symbolic logic with neural networks to ensure reliable and trustworthy decision-making and integrate meta-cognitive abilities to enable self-monitoring and adjustment of reasoning strategies for clearer, more understandable explanations.

— Neuro-Symbolic AI in 2024: A Systematic Review  (2501.05435 - Colelough et al., 9 Jan 2025) in Section 4.4 Logic and Reasoning

Our experiments use $|\Delta| = 2$; the compiled SDD grows empirically as $\approx|\Delta|{2.9}$ with a $40\times$ compile-time jump at $|\Delta| = 4$ (\Cref{app:treewidth}). Saturation and grounding are polynomial and the compilation is exact, but we do not demonstrate ontology-scale ABoxes such as SNOMED~CT or the Gene Ontology; that regime requires lifted WMC and is the principal open problem.

— Moose: Latent concept learning with reasoning-shortcut awareness in $\mathcal{EL}^{++}$  (2608.12961 - Mashkova et al., 13 Aug 2026) in Section 5, Limitations

Their treatment leaves three regimes open: the EL profile with role chains and role hierarchies, partial supervision in which a subset of the ground atoms is latent, and an RS analysis of the resulting predictor.

— Moose: Latent concept learning with reasoning-shortcut awareness in $\mathcal{EL}^{++}$  (2608.12961 - Mashkova et al., 13 Aug 2026) in Appendix, Section “Related work: compilation, refinement, and embedding approaches”

In multi-rule scenarios, which we leave to future work, the above can be substituted by the satisfiability of the knowledge base.

— Think-Verify-Revise: Neuro-Symbolic Visual Reasoning with Vision-Language Models and Dynamic Logic Tensor Networks  (2609.05388 - Afshari et al., 4 Sep 2026) in Section 3, subsection “Implementation details,” paragraph “Training”

Developing architectures that preserve symbolic guarantees at such scales remains an open problem.

— Neuro-symbolic AI for Industrial Configuration  (2609.29947 - Valerio et al., 24 Sep 2026) in Section Discussion

Several directions remain open for future work. First, grasp currently operates over symbolic predicates defined over noisy structured relational data. Extending the framework to support neural predicates would allow grasp to operate over unstructured inputs such as images and text.

— Learning Probabilistic Logic Programs with Functional Gradient Guided Language Models  (2610.12303 - Mathur et al., 8 Oct 2026) in Section 5, Conclusion