Neuro-Symbolic Artificial Intelligence: A Task-Directed Survey in the Black-Box Models Era

Neuro-Symbolic Artificial Intelligence: A Task-Directed Survey in the Black-Box Models Era discusses the integration of symbolic computing with neural networks to infer or exploit behavioral schemas, which has been considered a potential pathway toward human-level intelligence. The survey highlights that Neuro-Symbolic (NeSy) methods aim to enhance explainability and reasoning capabilities by combining the data-driven strengths of neural networks with the structured, logic-based processing of symbolic systems. However, practical implementation in real-world scenarios is hindered by limited semantic generalizability and the difficulty of applying pre-defined patterns and rules to complex domains such as Natural Language Processing and Computer Vision. Despite the remarkable performance of connectionist models since the AI breakthrough in 2017, questions remain about the competitiveness of NeSy approaches, particularly in tasks requiring robust generalization and transparency. The authors advocate for task-directed NeSy methodologies as a means to improve compositional reasoning and reduce dependence on large training datasets. Reproducibility details and further analysis of the surveyed works are available at https://github.com/disi-unibo-nlp/task-oriented-neuro-symbolic.git

This survey provides a comprehensive analysis of Neuro-Symbolic AI (NeSy), specifically examining its role in inferring behavioral schemas within the current landscape dominated by opaque "black-box" deep learning models. Unlike traditional surveys that categorize approaches solely based on architectural integration, this work adopts a task-directed perspective, organizing the review around the specific functional goals that NeSy systems aim to achieve. The authors explore how hybrid systems can leverage the pattern recognition capabilities of neural networks alongside the reasoning and representational strengths of symbolic AI to extract structured, interpretable behavioral patterns from complex data.

A core contribution of the paper is its critical assessment of the trade-offs involved in NeSy integration, particularly regarding semantic generalizability. While NeSy approaches show promise in overcoming the data-hungry and uninterpretable nature of pure neural models, the survey highlights significant limitations when applying these systems to highly complex domains. The authors argue that current methods often struggle to maintain semantic consistency and generalization when faced with the variability inherent in real-world environments, revealing a gap between rigid symbolic constraints and the fluidity of sub-symbolic representations.

This research is vital for the advancement of trustworthy AI, as it addresses the growing need for explainable and logically consistent systems in an era defined by large, inscrutable foundation models. By mapping the landscape of task-directed NeSy research and identifying the specific bottlenecks in semantic generalization, the paper provides a roadmap for future development. It underscores that successful integration requires not just architectural hybridization, but a deeper understanding of how behavioral schemas can be learned and transferred across domains without sacrificing the robustness provided by symbolic logic.

# Neuro-Symbolic AI: Bridging Black-Box Models in Complex Domains

This paper presents a task-directed survey of neuro-symbolic artificial intelligence (NeSy), focusing on its role in inferring behavioral schemas from data while addressing the limitations of semantic generalizability in complex domains. Neuro-symbolic approaches combine neural networks (for pattern recognition and perception) with symbolic reasoning (for logical inference and structure), aiming to mitigate the interpretability and robustness issues inherent in purely black-box deep learning models. The survey evaluates how NeSy methods can extract and refine behavioral representations—such as plans, goals, or causal relationships—from observational or interactive data, particularly in domains where raw neural models struggle with compositionality, explainability, or transfer learning.

A key contribution of this work is its critical examination of Neuro-Symbolic AI’s (NeSy) scalability and generalization challenges in real-world, high-dimensional environments. While NeSy frameworks excel at integrating symbolic constraints with neural learning (e.g., via neural theorem provers, differentiable inductive logic programming, or neural-symbolic architectures like DeepProbLog), the paper highlights persistent gaps in semantic alignment—where learned representations fail to generalize beyond training distributions or align with human-interpretable logic. The authors stress the need for task-specific benchmarks and hybrid evaluation metrics that balance neural performance with symbolic fidelity, emphasizing domains like robotics, reinforcement learning, and scientific discovery where interpretable, structured reasoning remains critical. By synthesizing recent advances and open problems, the survey underscores NeSy’s potential as a middle-ground solution in the era of increasingly opaque deep learning systems, while also exposing the technical and theoretical hurdles that must be overcome for widespread adoption.