symbolic representations in real-world sensory data? How can neuro-symbolic systems improve generalization to novel tasks beyond training distributions? What strategies ensure transparent, in terpretable reasoning processes? How can neuro-symbolic models learn robustly from small or sparse datasets? These questions motivate our comprehensive analysis of architectures, integration patterns, and de ployment strategies that enable effective neural-symbolic synergy. Our main contributions are: (1) The first comprehensive survey on neuro-symbolic AI agents, reviewing 178 papers (2020–November ∗Corresponding author. Email address: muhammad.adil@ieee.org (M. Adil). 1 Present address: Dept. of Computer Science, Texas Southern University, Houston, TX, USA. Received 19 May 2025; Re ceived in revised form 27 November 2025; Accepted 12 January 2026 Computer Science Review 60 (2026) 100902 Available online 31 January 2026 1574-0137/© 2026 Elsevier Inc. All rights are reserved, including those for text and data mining, AI training, and similar technologies. S.B. Hakim, M. Adil, A. Velasquez et al. 2025) using rigorous PRISMA methodology; (2) A novel taxonomy encompassing architectural configurations (single vs. multi-agent, ver tical vs. horizontal) and integration strategies (sequential, parallel, end-to-end differentiable, unified representation); (3) Systematic per formance analysis demonstrating neuro-symbolic advantages across diverse domains with quantitative comparisons; (4) Identification of crit ical research gaps, particularly meta-cognition (5%) and explainability (28%); (5) Structured open challenges and future directions addressing benchmarking consistency, integration strategies, and meta-cognitive capabilities via frameworks like TRAP. 1.1 . Survey methodology We employed the PRISMA methodology [15] ensuring a compre hensive, and rigorous review. Fig. 1 illustrates our systematic selection process maintaining breadth and depth while capturing emerging neurosymbolic AI agent trends. 1.1.1 . Search strategy We queried IEEE Xplore, ACM Digital Library, Springer Link, arXiv, Google Scholar, and ScienceDirect, combining “neurosymbolic” OR “neuro-symbolic” with agent-related terms (“agent”, “multi-agent”, “au tonomous system”) and integration approaches (“reasoning”, “learning”, “knowledge representation”, “explainability”, “meta-cognition”). The initial search yielded 1428 papers from January 2020 to November 2025. 1.1.2 . Inclusion and exclusion criteria We focused on English-language papers including peer-reviewed journal articles, conference papers, book chapters, and highly-cited arXiv preprints containing relevant advances. While we prioritize peerreviewed publications, we sel ectively include influential arXiv preprints that demonstrate significant architectural innovations, have catalyzed follow-on research, or provide unique empirical insights not yet avail able in the peer-reviewed literature (e.g., Agent Q’s MCTS-guided web navigation [16], GITM’s text-based Minecraft agent [17]). We acknowledge that such preprints have not undergone formal peer review Fig. 1. PRISMA flow diagram for paper selection, identifying 1428 initial records, ultimately including 178 papers meeting all neuro-symbolic AI agent research criteria (2020–November 2025). and may contain limitations identified during review processes; notably, several included preprints have been submitted to top-tier AI venues (e.g., NeurIPS, ICLR, AAAI) with acceptance rates below 25%, have in corporated reviewer feedback in revised arXiv versions, or are currently under review in subsequent submission cycles. Papers must explicitly address both neuro-symbolic integration and agent archi tectures, pro viding empirical results or theoretical frameworks. We excluded papers focusing exclusively on neural or symbolic approaches without integra tion, those not addressing agent-based systems, surveys without original contributions, and works lacking sufficient technical detail. 1.1.3 . Selection process After removing 641 duplicates, 787 papers underwent title/abstract screening, yielding 392 candidates. Full-text review reduced this to 178 papers meeting all criteria, as shown in Fig. 1. 1.1.4 . Data extraction and analysis From each paper, we extracted bibliographic information, system architecture, primary research focus, application domains, evalua tion methods, limitations, reproducibility status, and identified failure modes. Papers were categorized according to our taxonomy and an alyzed quantitatively and qualitatively to identify trends, gaps, and limitations. 1.2 . Paper organization Section 2 provides background on AI agents, symbolic AI, neural approac hes, and neuro-symbolic integration. Section 3 presents our com prehensive taxonomy. Section 5 analyzes architectural components and prominent systems. Sections 4 through 11 present applications, method ologies, performance analysis, and evaluation methods. Section 10 explores meta-cognition. Section 12 presents the AlphaGeometry case study. Sections 13–15 discuss open challenges, future directions, and conclusions. 2 . Background and definitions This section familiarizes readers with the integration of neural and symbolic components within AI agent architectures, outlining key characteristics and limitations grounding their integration. AI agents are autonomous software entities perceiving environments, making decisions, and taking actions to achieve goals [8]. Key charac teristics include autonomy (operating without intervention), reactivity (responding to changes), proactivity (goal-directed behavior), and social ability (interacting with agents/humans) [18]. Agent evolut ion pro gressed from simple reflex agents acting on current percepts to learning agents adapting through experience. Recent advances have produced sophisticated agents based on large language models [6], multimodal systems [7,19], and embodied agents [20] capable of complex planning and reasoning, as illustrated in Fig. 2. 2.1 . Symbolic AI Symbolic AI represents knowledge through explicit symbols and rules [21], founded on the Physical Symbol System Hypothesis posit ing human cognition models through symbol manipulation [22,23]. Key characteristics include explicit knowledge representation, logical rea soning, transparency, and modularity. While symbolic systems excel in interpretability and structured reasoning, they struggle with uncer tainty, unstructured data, and adapting to novel situations. Traditional symbolic approaches continue offering advantages in constrained envi ronments prioritizing computational efficiency, formal verification, or explicit rule-based reas oning [24–26]. 2.2 . Neural approaches Neural approaches derive representations implicitly from data, ex celling in pattern recognition and adaptation. Characterized by datadriven learning, distributed information representation across network Computer Science Review 60 (2026) 100902 2 Downloaded for personal academic use. All rights reserved. https://papernode.online/ S.B. Hakim, M. Adil, A. Velasquez et al. Fig. 2. Evolution of AI agent architectures from simple reflex agents to learning agents with increased reasoning capabilities. weights, adaptability to varying conditions, and powerful pattern recog nition [27,28]. Despite their strengths, neural systems face challenges in systematic reasoning, explicit knowledge representation, and decisionmaking interpretability, which are particularly problematic in domains requiring formal logical inference or transparent justification. 2.3 . Neuro-symbolic integration Neuro-symbolic AI merges neural and symbolic paradigm strengths [ 29], creating hybrid systems that combine neural network flexibil ity and pattern recognition with symbolic reasoning’s logical structure and interpretability. Integration is motivated by dual-process reasoning theories [30,31] distinguishing System 1 (intuitive) versus System 2 (de liberative) thinking [32], supported by recent meta-analyses [33] and computational models [34]. Integrated learning cycles combine neural pattern recognition with symbolic knowledge representation [29], while modular design patterns [35] provide computational frameworks for im plementing dual processes. Cognitive architectures like Soar [36,37] and ACT-R [38] demonstrate the realization of hybrid processing. 2.4 . Agentic AI: recent advances Recent research (2024–2025) extended traditional neuro-symbolic approaches developing agentic AI frameworks—autonomous systems proactively designing workflows and coordinating multiple agents for complex tasks. Innovations like narrative memory for long- term con textual reasoning [39,40] enable agents to maintain coherent under standing across temporally distant events. Refined chain-of-thought prompting improved agents’ capacity to utilize external tools and co ordinate workflows. These developments underscore neuro-symbolic integration’s critical role in enabling truly agentic behavior, as neither purely neural nor symbolic approaches easily achieve this combination of flexibility and structured persistence. 3 . Taxonomy of neurosymbolic AI agents Based on systematic literature analysis, we propose a comprehensive taxonomy considering both architectural configurations and integration approaches, as illustrated in Fig. 3. 3.1 . Architectural configurations Neuro-symbolic agent architectures are broadly categorized as single-agent or multi-agent systems. Single-agent systems integrate neural and symbolic components for independent task performance, representing 65% of reviewed papers. These are subdivided by integra tio n patterns: Sequential coupling operates modules in sequence, as in ReAct [41] and Reflexion [42], providing clear separation but risking information loss at boundaries [43]. Parallel coupling processes in puts concurrently, combining outputs for rapid pattern recognition and Table 1 Key capabilities across major neuro-symbolic agent frameworks. Framework Auton. Metacog. Tool Int. Symb. Reason. Neural Learn. Explainability Agent Q [16] ✓ ✘ ✓ ✓ ✓ ✓ GoalAct [48] ✓ ✘ ✓ ✓ ✓ ✘ AlphaGeometry [49] ✘ ✘ ✘ ✓ ✓ ✓ Reflexion [42] ✘ ✓ ✘ ✘ ✓ ✓ MetaGPT [50] ✓ ✘ ✓ ✘ ✓ ✓ TRAP [51] ✓ ✓ ✘ ✓ ✓ ✓ structured reasoning, handling uncertainty well but potentially intro ducing conflicts. End-to-end differentiable architectures embed symbolic operations within differentiable networks, facilitating gradient-based optimization, as in Logic Tensor Networks [44,45] and Neural Theorem Provers [46]. Zero-shot concept learning systems like Ze roC [47] enable novel concept recognition through symbolic graph integration with neural energy-based models. Table 1 summarizes key capabilities across major neuro-symbolic frameworks, showing varying support for autonomy, meta-cognition, tool integration, symbolic reasoning, neural learning, and explainability. Multi-agent architectures [52–54], comprising 35% of the literature, address complex tasks through collaboration [55]. Vertical architec tures implement hierarchical structures with leader agents coordinating subordinates, improving task efficiency [56]. Horizontal architectures operate as egalitarian systems with peer collaboration, exemplified by AgentVerse [56] and MetaGPT [50]. Hybrid architectures combine vertical and horizontal elements with dynamic leadership. Emerging dy namics emphasize dynamic task allocation and enhanced collaborative reasoning [57,58] adapting to changing conditions. 3.2 . Integration approaches Neuro-symbolic integration divides into four main approaches. Knowledge Representation Integration [59] leverages neural enhance ments to symbolic knowledge bases, organizing neural representations per symbolic schemas [60–64]. Recent advances in symbolic knowl edge distillation [65] enable the extraction of structured, interpretable symbolic knowledge from LLMs, complementing knowledge graph in tegration by transforming implicit neural representations into explicit symbolic forms that enhance transparency and reasoning capabilities. Learning and Inference Integration embeds symbolic reasoning within neural frameworks, employing differentiable reasoning [66–69], neuralguided symbolic search (AlphaGeometry [49]), neuro-symbolic program synthesis [70], and continual learning [71,72]. LINC [73] demonstrates significant improvements in logical reasoning by integrating formal logic with neural language understanding. Explainability and Trustworthiness Integration derives transparent symbolic explanations from neural deci sions [74,75], incorporating Computer Science Review 60 (2026) 100902 3 Downloaded for personal academic use. All rights reserved. https://papernode.online/
Download PDF: https://exemples.eu.org/7IzAwN
