Recent Research: Ontologies for Agentic AI, AI Agents, and LLMs

A 2025–2026 Literature Brief

Scope and How to Use This Brief

This brief surveys active research (mostly late 2025 through mid-2026) on using ontologies and knowledge graphs to ground LLM-based agents. This tracks the research and industry activity validating (and complicating) that model. A caution worth stating up front: several entries are arXiv preprints from 2026 and have not yet completed peer review. They’re included because they’re the most current signal in a fast-moving field, but treat their specific numbers as directional rather than settled.

1. Neurosymbolic Grounding for Enterprise Agents

The most architecturally complete recent work is a 2026 preprint proposing a three-layer ontological framework — Role, Domain, and Interaction ontologies — to constrain LLM reasoning in enterprise agents [1]. Two details make it worth reading closely rather than skimming the abstract:

• It’s validated against a production system reportedly serving 22 industry verticals with 650+ specialized agents, which is a rare case of ontology-grounding research with real deployment scale behind it rather than a lab benchmark alone.
• It proposes a six-level maturity model (L0–L5) running from ungrounded LLM deployment through full closed-loop neurosymbolic integration, which is a genuinely useful framework for benchmarking where your own platform sits today.

The paper also documents an unintuitive finding worth flagging for your architecture decisions: injecting ontological context can sometimes displace useful knowledge the model already has parametrically — a “context interference” effect independently corroborated by two other 2026 studies on resisting contextual interference in RAG and on performance degradation from context length alone [2, 3]. The practical implication is that “always inject the full ontology” is not obviously optimal — selective, confidence-aware injection may outperform blanket grounding, which the same research group formalizes as a context-efficiency optimization problem in a companion paper [4], and separately as a quantum-inspired model of constructive/destructive knowledge interference [5].

A third companion paper proposes ontology-powered simulation for formal safety certification of enterprise agents, tested across roughly 1,800 scenarios spanning five industries [6] — relevant if part of your platform work involves a compliance argument for why an agent’s action should be trusted.

On the industry side, the closest parallel is Skan AI’s Agentic Ontology of Work (AOW), released as a v1.0 whitepaper in early 2026 [7]. It defines eight canonical entities — Agents, Skills, Intents, Contexts, Policies, Memory, Confidence, Outcomes — explicitly modeled on the way Service-Oriented Architecture once standardized service description and discovery. It’s scoped to workflow orchestration and governance (who did what, under what policy, with what assurance level) rather than to constraining an LLM’s internal reasoning, which is the meaningful distinction from [1]. Worth noting for Kubernetes/GitOps governance work specifically: AOW is explicitly designed to be implementable over JSON-LD, RDF/OWL, or a property-graph knowledge graph rather than being a proprietary format.

2. Hallucination Mitigation via Ontology-Grounded Knowledge Graphs (Clinical Focus)

This is the theme most directly transferable to clinical AI platforms, and it has the strongest empirical numbers in this brief.

A 2026 Journal of Biomedical Informatics study built a domain-specific RDF/OWL ontology and knowledge graph from multi-institution hospital data, then integrated it with an LLM via a GraphRAG-style pipeline for clinical question answering [8]. Against a baseline of ChatGPT-4 (37% accuracy) and DeepSeek-R1 (52% accuracy), the ontology-grounded framework reached 98% accuracy, with hallucination rate falling from roughly 63% and 48% respectively down to 1.7%. That’s one of the cleanest before/after comparisons currently published for this pattern.

A separate 2025 paper, Grounding LLM Reasoning with Knowledge Graphs, reports a 26.5% improvement over chain-of-thought baselines on the GRBench benchmark, achieved by explicitly linking each individual reasoning step to a piece of graph-structured data rather than grounding only the final answer [9] — a useful evaluation pattern if you want to instrument your Supervisor-Critic loop with step-level rather than output-level verification. Referenced within [1], a related line of work — OG-RAG, ontology-grounded retrieval-augmented generation — reports a 55% improvement in fact recall from ontology-structured retrieval [10], reinforcing that the gain shows up specifically on relational, multi-hop questions rather than on simple lookups.

Other clinically-focused work worth knowing:

• CliCARE grounds LLM-based decision support in clinical guidelines over longitudinal cancer EHR data, motivated explicitly by standard RAG’s failure to preserve the relational structure needed for oncology decision-making [11].
• A January 2026 pipeline for clinical knowledge graph construction combines multi-agent prompting, entropy-based uncertainty scoring, ontology-aligned RDF/OWL schema generation, and multi-LLM consensus validation, applied to pancreatic and breast cancer cohorts without requiring gold-standard annotation [12] — relevant if your Bronze-to-Silver entity resolution step needs an unsupervised bootstrapping strategy.
• A systematic review of agentic AI in radiology found that multi-agent, role-based architectures combined with RAG measurably reduce hallucination rates and improve diagnostic accuracy, while noting the approaches remain computationally expensive and short on comprehensive clinical validation [13].
• A 2026 MDPI paper proposes agentic hallucination risk scoring for medical LLMs, computing multiple uncertainty signals from an initial response and routing high-risk cases to a knowledge-injection step rather than treating every output uniformly [14] — a pattern that maps cleanly onto a critic agent with a confidence threshold.
• ESWC is running a dedicated 2026 workshop (LLM4KGOE) on LLM-driven knowledge graph and ontology engineering, with an explicit research track on hallucination benchmarking and calibrated abstention for uncertain graph links [15] — worth watching for emerging shared benchmarks in this exact space.

3. GraphRAG and Graph-Augmented Agent Reasoning

Microsoft Research’s original GraphRAG work remains the reference architecture and has since been folded into Microsoft Discovery, an agentic scientific-research platform on Azure [16]. The first systematic survey of the field, published in ACM Transactions on Information Systems in 2026, organizes the now-sprawling literature into three stages: graph-based indexing, graph-guided retrieval, and graph-enhanced generation [17] — a useful taxonomy if you’re evaluating vendor claims against a common framework rather than marketing language.

Two more recent systems worth tracking:

• MemGraphRAG (KDD 2026) addresses a specific weakness in earlier GraphRAG systems — isolated, chunk-level extraction that produces a noisy, globally inconsistent graph — by adding a persistent global memory layer during graph construction [18].
• A broader 2025 survey, Graphs Meet AI Agents, is the first systematic taxonomy specifically of how graph structures support agent planning, execution, memory, and multi-agent coordination, treated as distinct from the older “LLMs on graphs” literature [19]. A companion 2026 survey narrows in specifically on graph-based agent memory, distinguishing knowledge memory from experience memory and arguing graph structure is emerging as the dominant representation for both [20].

One practical caveat surfaced repeatedly in the applied literature and worth carrying into any build-vs-buy decision: knowledge graph construction is not a one-time cost. Industry commentary on the original Microsoft GraphRAG approach notes construction costs can run 10–100x higher than standard vector-RAG indexing depending on corpus size and extraction complexity, and unlike a vector index, updating a knowledge graph when source data changes is not a mechanical re-embed — a changed entity can ripple through downstream relationships [21].

4. Automated Ontology Construction with LLMs

This subfield has matured enough to have a recurring shared task: LLMs4OL, now in its second year (2025), benchmarks how well LLMs perform discrete ontology-learning subtasks — term typing, taxonomy discovery, and relation extraction — across a diverse set of participating models and methods [22]. A 2026 systematic literature review in the Semantic Web Journal surveys the full space of how LLMs are being adapted across the ontology-engineering lifecycle, from requirements elicitation through refinement [23].

Two specific contributions stand out for enterprise applicability:

• OntoScope (2026) proposes a divergent-convergent interaction framework where an LLM helps a human scope an ontology through structured dialogue rather than generating one unsupervised — a human-in-the-loop pattern that directly addresses the “don’t let an LLM validate its own ontology” risk [24].
• A 2026 arXiv paper on LLM-driven ontology construction is framed specifically around enterprise knowledge graphs rather than academic domain ontologies, making its pipeline choices more directly comparable to a production Bronze→Silver→Gold mapping exercise [25].
• A 2026 multi-agent approach to ontology generation from unstructured text assigns distinct LLM agents to extraction, schema alignment, and validation roles rather than using a single model end-to-end [26], echoing the same specialization pattern used in the clinical KG pipeline above [12].
• A related multi-agent system for mapping relational schemas to Schema.org-aligned knowledge graphs assigns separate LLM agents to mapping, extraction, and validation, and reports high mapping accuracy on the Spider cross-domain benchmark [27] — a reasonable reference architecture if part of your Silver-layer work involves aligning heterogeneous device-log schemas to a common vocabulary.

5. Multi-Agent Semantic Interoperability & Communication Protocols

A 2025 survey (with citations extending into 2026) examines four emerging agent interoperability protocols side by side: MCP, ACP (Agent Communication Protocol), A2A (Agent-to-Agent), and ANP (Agent Network Protocol) [28]. It places them at different architectural layers — MCP handles model-layer context translation and tool invocation via JSON-RPC; ACP defines message-level, RESTful, session-aware multimodal communication; A2A manages peer-to-peer task collaboration; and ANP targets decentralized, internet-scale agent discovery using W3C decentralized identifiers. Usefully, it also traces the lineage back to 1990s agent-communication languages like KQML, which already carried an explicit :ontology field in every message envelope — the “shared vocabulary” problem in multi-agent systems is not new, only newly urgent.

A 2026 follow-on proposal for a unified ACP makes the ontology connection explicit rather than implicit: it defines a semantic layer as the “ontology of intent,” translating high-level agent goals into standardized JSON-LD so that agents built on different internal representations can still interoperate [29]. That’s a directly actionable pattern for a Supervisor-Critic architecture spanning multiple vendor-built agents.

A broader historical synthesis argues that today’s agent protocol stack (MCP/A2A/ACP/ANP) is best understood as a continuation of 2000s-era Semantic Web and multi-agent-systems research rather than a clean break from it [30] — useful context if you’re making an architectural case internally for why ontology investment isn’t a detour from “real” agentic AI work but a return to a body of research the field temporarily set aside.

Two more specialized but relevant papers: a maritime-domain paper proposes a multilayered ontology-driven model specifically for semantic interoperability among heterogeneous autonomous agents [31], and an EACL 2026 thesis proposal explores a multi-agent system where perspective-specific agents extract knowledge in parallel into a shared “polyvocal” ontology that preserves rather than collapses conflicting stakeholder viewpoints [32] — potentially relevant if different clinical departments need to reason from different but simultaneously valid interpretations of the same underlying data.

6. Cross-Cutting Synthesis: What’s Converging

A few patterns show up independently across multiple research groups and application domains in this brief, which is a reasonable signal they’re real rather than artifacts of one team’s framing:

7. Recommended Reading Path

If you only read four of these:

1. [1] — closest architectural match to a Supervisor-Critic pattern grounded in formal ontology, with the most useful maturity model for benchmarking your own platform.
2. [8] — the strongest clinical-domain evidence base (98% vs. 37% accuracy), directly citable if you need to justify ontology investment to non-technical stakeholders.
3. [28] — essential background on where MCP sits relative to the rest of the emerging agent-protocol stack, and where ontology fits into that stack.
4. [7] — useful as an enterprise-governance vocabulary reference even where you disagree with specific entity choices, since it’s the most complete open attempt at standardizing agentic-work terminology to date.

References

1. Luong Tuan, T. et al. (2026). Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents. arXiv preprint. https://arxiv.org/abs/2604.00555
2. Lin, C. et al. (2026). Resisting Contextual Interference in RAG via Parametric-Knowledge Reinforcement. ICLR 2026.
3. Du et al. (2025). Cited in [1] regarding performance degradation from context length.
4. Luong Tuan, T. (2026a). The Information-Theoretic Bounds of Ontological Context: Optimal Knowledge Injection for Enterprise LLM Agents. arXiv preprint (companion to [1]).
5. Luong Tuan, T. (2026b). Knowledge Interference in Enterprise LLM Agents: A Quantum-Inspired Framework for Context Engineering. arXiv preprint (companion to [1]).
6. Luong Tuan, T. & Sanyal, A. (2026). Toward Verifiable Enterprise AI Agents: Ontology-Powered Simulation and Formal Safety Certification. arXiv preprint (companion to [1]).
7. Skan AI (2026). Agentic Ontology of Work (AOW), Version 1.0. Whitepaper. https://www.skan.ai/whitepapers/agentic-ontology-of-work
8. Ali, M., Taha, Z., & Morsey, M. M. (2026). Ontology-grounded knowledge graphs for mitigating hallucinations in large language models for clinical question answering. Journal of Biomedical Informatics. https://www.sciencedirect.com/science/article/abs/pii/S1532046426000171
9. Amayuelas, A., Sain, J., Kaur, S., & Smiley, C. (2025). Grounding LLM Reasoning with Knowledge Graphs. arXiv:2502.13247.
10. Sharma et al. (2025). OG-RAG: Ontology-Grounded Retrieval-Augmented Generation. Cited in [1].
11. CliCARE: Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records. arXiv:2507.22533. https://arxiv.org/pdf/2507.22533
12. Clinical Knowledge Graph Construction and Evaluation with Multi-LLMs via Retrieval-Augmented Generation (2026). arXiv:2601.01844. https://arxiv.org/html/2601.01844
13. Agentic AI and Large Language Models in Radiology: Opportunities and Hallucination Challenges. PMC. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12729288/
14. Agentic Hallucination Risk Scoring for Medical LLMs via Uncertainty Quantification and Clinical Knowledge Injection. Algorithms (MDPI), 2026. https://doi.org/10.3390/a19040315
15. LLM4KGOE 2026: Workshop on LLM-driven Knowledge Graph and Ontology Engineering (ESWC 2026). https://koncordantlab.github.io/LLM4KGOE-ESWC/
16. Microsoft Research. Project GraphRAG. https://www.microsoft.com/en-us/research/project/graphrag/
17. Graph Retrieval-Augmented Generation: A Survey. ACM Transactions on Information Systems, 2026. https://dl.acm.org/doi/10.1145/3777378
18. MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation. KDD 2026. https://arxiv.org/html/2606.00610v1
19. Graphs Meet AI Agents: Taxonomy, Progress, and Future Opportunities. arXiv:2506.18019.
20. Graph-based Agent Memory: Taxonomy, Techniques, and Applications. arXiv:2602.05665.
21. Industry commentary on Microsoft GraphRAG construction costs and index-maintenance overhead. DEV Community, 2026.
22. Babaei Giglou, H., D’Souza, J., Mihindukulasooriya, N., & Auer, S. (2025). LLMs4OL 2025 Overview: The 2nd Large Language Models for Ontology Learning Challenge. Open Conference Proceedings, 6. https://www.tib-op.org/ojs/index.php/ocp/article/view/2913
23. Large Language Models for Ontology Engineering: A Systematic Literature Review. Semantic Web Journal, 2026. https://www.semantic-web-journal.net/system/files/swj3864.pdf
24. Zhao, Y., Meroño-Peñuela, A., & Simperl, E. (2026). OntoScope: Using a Divergent-Convergent Interaction Framework to Support LLM-based Ontology Scoping. Proceedings of the 31st International Conference on Intelligent User Interfaces, 67–84.
25. LLM-Driven Ontology Construction for Enterprise Knowledge Graphs. arXiv:2602.01276.
26. Towards Automated Ontology Generation from Unstructured Text: A Multi-Agent LLM Approach. arXiv:2604.23090.
27. A Multi-Agent System for Semantic Mapping of Relational Data to Knowledge Graphs. arXiv:2511.06455.
28. Ehtesham, A., Singh, A., Gupta, G. K., & Kumar, S. (2025). A Survey of Agent Interoperability Protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP). arXiv:2505.02279.
29. Beyond Context Sharing: A Unified Agent Communication Protocol (ACP) for Secure, Federated, and Autonomous Agent-to-Agent (A2A) Orchestration. arXiv:2602.15055.
30. From Multi-Agent Systems and the Semantic Web to Agentic AI: A Unified Narrative of the Web of Agents. arXiv:2507.10644.
31. Semantic Interoperability of Multi-Agent Systems in Autonomous Maritime Domains. MDPI Electronics, 2025. https://www.mdpi.com/2079-9292/14/13/2630
32. Thesis Proposal: A Multi-Agent System for Ontology-Based Perspective-Aware Knowledge Extraction. EACL 2026 Student Research Workshop. https://aclanthology.org/2026.eacl-srw.46.pdf