Characterizing Knowledge Representation Boundaries in Large-Scale Neural-Symbolic Architectures

Abstract

Neural symbolic systems aim to integrate the perceptual generalization strengths of neural networks with the structural reasoning capabilities of symbolic logic. However, this study finds that the internal representations formed by large-scale neural components are inherently limited in their ability to preserve symbolic identity, compositional structure, and rule invariance across transformations. Through controlled evaluation of representational load, referential continuity, context perturbation, domain transfer, and embedding drift over scale, we show that neural representations remain context-dependent and correlation-driven, leading to systematic breakdowns when deeper logical abstraction or cross-domain consistency is required. These findings indicate that performance on symbolic tasks in familiar contexts does not imply stable knowledge representation. Therefore, achieving reliable neural symbolic reasoning requires architectures that incorporate explicit symbolic binding and structural grounding mechanisms, rather than relying solely on distributed neural encoding.