Source-linked AI summary
Neuro-Symbolic AI in 2024: A Systematic Review
Brandon C. Colelough, William Regli
TL;DR
Neuro-Symbolic AI emerged from integrating symbolic and sub-symbolic approaches, but research coverage remains uneven, especially for Meta-Cognition, explainability, and trustworthiness. This systematic review applies PRISMA screening to literature from 2020–2024 and identifies concentrated activity in learning and inference alongside gaps in these underexplored areas.
Problem
Research integrating Symbolic AI and Sub-Symbolic AI lacks balanced coverage of explainability, trustworthiness, and Meta-Cognition, despite their relevance to reliable and adaptive AI systems.
Method
The review used PRISMA methodology to screen peer-reviewed English-language literature published from 2020 to 2024 across five databases, retaining studies with relevant public codebases.
Results
The review found concentrated research activity in learning and inference, logic and reasoning, and knowledge representation, while explainability, trustworthiness, and Meta-Cognition remained comparatively sparse.
Takeaways & Limitations
The review identifies interdisciplinary opportunities to connect explainability and trustworthiness with other Neuro-Symbolic AI research areas and to develop Meta-Cognition further.
Takeaways & Limitations
The System 1/System 2 framing used for Neuro-Symbolic AI is a useful orientation but remains too simplistic to represent the field’s full systems-level cognitive structure.
Abstract
from arXiv · showhide
Background: The field of Artificial Intelligence has undergone cyclical periods of growth and decline, known as AI summers and winters. Currently, we are in the third AI summer, characterized by significant advancements and commercialization, particularly in the integration of Symbolic AI and Sub-Symbolic AI, leading to the emergence of Neuro-Symbolic AI. Methods: The review followed the PRISMA methodology, utilizing databases such as IEEE Explore, Google Scholar, arXiv, ACM, and SpringerLink. The inclusion criteria targeted peer-reviewed papers published between 2020 and 2024. Papers were screened for relevance to Neuro-Symbolic AI, with further inclusion based on the availability of associated codebases to ensure reproducibility. Results: From an initial pool of 1,428 papers, 167 met the inclusion criteria and were analyzed in detail. The majority of research efforts are concentrated in the areas of learning and inference (63%), logic and reasoning (35%), and knowledge representation (44%). Explainability and trustworthiness are less represented (28%), with Meta-Cognition being the least explored area (5%). The review identifies significant interdisciplinary opportunities, particularly in integrating explainability and trustworthiness with other research areas. Conclusion: Neuro-Symbolic AI research has seen rapid growth since 2020, with concentrated efforts in learning and inference. Significant gaps remain in explainability, trustworthiness, and Meta-Cognition. Addressing these gaps through interdisciplinary research will be crucial for advancing the field towards more intelligent, reliable, and context-aware AI systems.
Discussion:
The review finds strong integration across learning and inference, logic and reasoning, and knowledge representation, while explainability, trustworthiness, and Meta-Cognition remain less developed.
- Research is well integrated across learning and inference, logic and reasoning, and knowledge representation.
- Explainability and trustworthiness remain a notable research gap for reliable AI deployment.
- Meta-Cognition is sparsely represented, limiting the development of systems that self-monitor, evaluate, and adjust their processes.
1. Introduction
The paper situates Neuro-Symbolic AI within the current AI summer as a hybrid of symbolic and sub-symbolic approaches. It reviews this integration as a route toward systems combining learned representations with reasoning capabilities.
- Neuro-Symbolic AI emerged from integrating Symbolic AI and Sub-Symbolic AI during the current AI summer.
- Symbolic AI formalizes implicit human knowledge into declarative representations based on rules and logic.
- Sub-Symbolic AI learns implicit data representations without requiring rules or symbolic representations as inputs.
- The paper frames Neuro-Symbolic AI as combining fast, intuitive neural processing with slow, deliberate symbolic reasoning.
- The adopted definition describes Neuro-Symbolic AI as a composite framework merging symbolic AI and neural networks to create a hybrid model with reasoning capabilities.
2. Methodology
The paper organizes Neuro-Symbolic AI around five foundational research areas and uses a PRISMA-based review process to identify literature and gaps. Meta-Cognition is added to address limitations in existing taxonomies.
- Taxonomy: Five foundational areas structure the taxonomy: knowledge representation, learning and inference, explainability and trustworthiness, logic and reasoning, and Meta-Cognition.
- Knowledge Representation: Knowledge representation integrates symbolic and neural representations, including commonsense and domain-specific knowledge graphs.
- Learning and Inference: Learning and inference combines learning and reasoning through end-to-end differentiable reasoning and dynamic multi-source knowledge reasoning.
- Explainability and Trustworthiness: Explainability and trustworthiness focuses on interpretable models and reasoning processes to support trust and reliability.
- Logic and Reasoning: Logic and reasoning integrates logic-based methods with neural networks, including logical and probabilistic reasoning and syntax and semantics.
- Meta-Cognition: Meta-Cognition is defined as monitoring, evaluating, and adjusting a system’s reasoning and learning processes.
- Meta-Cognition: The taxonomy adds Meta-Cognition because existing taxonomies do not capture self-awareness, adaptive learning, reflective reasoning, self-regulation, and introspective monitoring.
- Literature Review Approach: The review follows PRISMA across five databases and selects literature from 2020 to 2024 using peer review, relevance, and public-codebase criteria.
3. Results
The review narrowed a broad 2020–2024 literature search to a smaller reproducible corpus, with research concentrated in five focal areas and limited overlap among them.
- Literature selection: 1,428 papers were extracted from five databases covering Neuro-Symbolic AI publications from 2020–2024.The search terms combined “neurosymbolic” with terms for the five foundational research areas.
- Literature selection: 167 papers had associated code-base repositories and were gathered for detailed reading before nine were removed, leaving 158 included papers.Meta-Cognition papers were exempted from the public-codebase requirement because no codebases could be found for that category.
- Research focal areas: 63% (n=99) of included entries addressed Learning and Inference, compared with 44% (n=70) for Knowledge Representation, 35% (n=55) for Logic and Reasoning, 28% (n=44) for Explainability and Trustworthiness, and 5% (n=8) for Meta-Cognition.Entries could be sub-categorised across the five focal research areas.
- Research intersections: 3.33% (n=9) represented the intersection of Explainability and Trustworthiness with Logic and Reasoning.Only one entry intersected all four main research focal areas, excluding Meta-Cognition.
4. Discussion & Open Questions
The reviewed literature shows substantial activity across knowledge, learning, reasoning, and language-oriented research, while identifying open questions around context-aware inference, explainability, incremental learning, and meta-cognition.
- Knowledge Representation: Knowledge Representation research advances semantic grounding, complex-relationship representation, commonsense knowledge, and data efficacy.Examples include commonsense and event-based representations, knowledge graphs, personalized knowledge, and NeuroQL.
- Learning and Inference: Learning and Inference work combines Neuro-Symbolic integration, advanced problem solving and decision making, and semantic enhancement for model trustworthiness.Pseudo-Semantic Loss integrates logic within an autoregressive model’s loss function, while Logic Tensor Networks target logical consistency, toxicity reduction, and prediction accuracy.
- Learning and Inference: Open questions include incremental learning, context-aware inference, fine-grained explainability for complex inference chains, and meta-cognitive monitoring and optimization.These questions concern systems that evolve with new experiences and adjust reasoning or learning in dynamic environments.
- Explainability and Trustworthiness: Explainability and Trustworthiness research develops NLP techniques, logical reasoning, and language understanding and summarization through structured, probabilistic, and semantic methods.Examples include Braid, Structure-Aware Abstractive Conversation, semantic-level confounder correction, and FactPEGASUS.
- Intersections: AlphaGeometry is the only reported project at the intersection of all four main focal areas, combining a neural language model with a symbolic deduction engine.The review describes sparse integration of Explainability and Trustworthiness with the other three areas as an opportunity for interdisciplinary work.
- Meta-Cognition: Meta-Cognition research includes reinforcement learning, cognitive architectures with large language models, and systems aligned with the Common Model of Cognition.These approaches approximate meta-cognitive capabilities and support robustness, interoperability, or cognitively guided few-shot learning.
5. Meta-Cognition in Neuro-Symbolic AI
The paper frames Neuro-Symbolic AI’s system 1/system 2 analogy as a useful orientation, but too simplistic to represent cognition’s full systems-level complexity. It argues that advancing the field requires systems with comparable system-1/system-2 characteristics.
- The system 1/system 2 representation helps orient Neuro-Symbolic AI toward integrating neural and symbolic processes.
- The current human-level cognitive-processing analogy is too simplistic to capture the systems-level breakdown needed to guide research investment.
- Cognition is described as a deeply complex, layered, and densely connected system of systems that acts with system-1/system-2 characteristics.
- The field’s goal is to create hybrid AI models with reasoning capabilities and system-1/system-2-like behavior.
6. Conclusion
The review finds rapid Neuro-Symbolic AI growth since 2020, with research concentrated in learning and inference, logic and reasoning, and knowledge representation. It also identifies sparse work in explainability, trustworthiness, and Meta-Cognition, alongside opportunities for interdisciplinary research.
- Research activity surged from 2020 onward as interest grew in integrating symbolic and sub-symbolic approaches to enhance AI reasoning.
- The review defines Meta-Cognition, examines 2020–2024 literature themes, and identifies current Neuro-Symbolic AI research gaps.
- Most research efforts focus on learning and inference, with substantial work also addressing logic and reasoning and knowledge representation.
- Explainability and trustworthiness remain relatively sparse despite their importance for transparency and reliability in real-world AI applications.
- The four main research areas are generally well integrated, while their intersections present significant opportunities for interdisciplinary work.
University of Maryland College Park
The supplied passages identify the University of Maryland College Park as a department affiliation and provide the paper’s title, authors, publication metadata, and workshop context.
- The supplied affiliation text identifies the Department of Computer Science.
- The paper is titled “Neuro-Symbolic AI in 2024: A Systematic Review.”
- The document lists arXiv identifier 2501.05435v2, dated 5 April 2025.
- The listed authors are Brandon Colelough and William Regli.
- The work fulfilled a Master of Science scholarly-paper requirement and was presented at the Logical Foundations of Neuro-Symbolic AI 2024 workshop.