Where Perceptrons Meet Proof Trees: A Contemporary Synthesis Review of Neuro-Symbolic Reasoning Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This article offers a contemporary synthesis review of neuro-symbolic reasoning systems, the research program that seeks to combine the perceptual strength and statistical learning of neural networks with the compositionality, transparency, and verifiability of symbolic reasoning. The review reconstructs the two parent traditions and their characteristic failures, the brittleness of classical symbolic systems whose common-sense knowledge could never be completed, and the opacity of connectionist systems whose competence does not decompose into inspectable structure, and follows the recurring attempts to marry them: the neural-symbolic program of the 2000s that mapped logic onto networks, the differentiable logic wave of the late 2010s that made theorem proving and rule learning gradient-compatible, the neuro-symbolic architectures that solved compositional visual reasoning by dividing perception from inference, and the present interregnum of large language models, whose fluent generalization coexists with demonstrated failures at planning and compositional reasoning, and whose most credible deployments increasingly wrap the model in symbolic scaffolds for verification and control. Three synthetic claims are advanced. First, neuro-symbolic research is best read not as a merger of two equals but as a recurring negotiation of the division of labor between them, with each era redrawing the boundary according to which side holds the current advantage. Second, the large language model did not settle this negotiation; it industrialized one side of it while benchmarks exposed exactly the compositional failures the symbolic tradition was built to prevent, reviving external symbolic structure as the vehicle of reliability. Third, the field's weakest instrument remains evaluation, since no standard assessment separates reasoning from pattern matching. An agenda is proposed spanning reasoning-faithful benchmarks, verified neuro-symbolic stacks for safety-critical use, and the theory of when learned components can be safely embedded under logical control.