Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents

Thanh Luong Tuan; Abhijit Sanyal

cs.AIApr 1, 2026

Ontology-Constrained Neural Reasoning in Enterprise Agentic Systems: A Neurosymbolic Architecture for Domain-Grounded AI Agents

Authors: Thanh Luong Tuan, Abhijit Sanyal

Organizations: Golden Gate University, San Francisco · Foundation AgenticOS (FAOS) · Associate Director, Data, Digital & IT, Novartis Healthcare Pvt. Ltd., Hyderabad, India

Abstract

Enterprise adoption of Large Language Models (LLMs) is constrained by hallucination, domain drift, and the inability to enforce regulatory compliance at the reasoning level. We present a neurosymbolic architecture implemented within the Foundation AgenticOS (FAOS) platform that addresses these limitations through ontology-constrained neural reasoning. We introduce a three-layer ontological framework--Role, Domain, and Interaction ontologies--grounding LLM-based enterprise agents. We formalize asymmetric neurosymbolic coupling: current enterprise systems constrain agent inputs (context assembly, tool discovery, governance thresholds) but not outputs, and we propose mechanisms extending this coupling to output-side validation (response checking, reasoning verification, compliance enforcement). A controlled experiment (1,800 runs across five industries and three LLMs: Claude Sonnet 4, Qwen 2.5 72B, Gemma 4 26B) finds ontology-coupled agents significantly outperform ungrounded agents on Metric Accuracy (p < .001) and Role Consistency (p < .001) across all three models with large effect sizes (Kendall's W = .46-.64). Improvements are greatest where LLM parametric knowledge is weakest--particularly in Vietnam-localized domains, where ontology lift is 2x that of English domains. Contributions: (1) a formal three-layer enterprise ontology model; (2) a taxonomy of neurosymbolic coupling patterns; (3) ontology-constrained tool discovery via SQL-pushdown scoring; (4) a proposed framework for output-side ontological validation; (5) empirical evidence for the inverse parametric knowledge effect--ontological grounding value is inversely proportional to LLM training-data coverage of the domain; (6) cross-model replication establishing model-independence; (7) a production system serving 22 industry verticals with 650+ agents.

Figures & tables

Level	Name	Description
L0	Ungrounded	LLM operates without ontological context
L1	Context-Injected	Ontology provides prompt context (input-side)
L2	Discovery-Constrained	Ontology filters available tools and skills
L3	Process-Gated	Ontology enforces approval gates and escalation during execution
L4	Output-Validated	Ontology validates and constrains LLM outputs post-generation
L5	Closed-Loop	Full bidirectional coupling: ontology constrains inputs, guides reasoning, validates outputs, and evolves from agent experience

Table 1: Neurosymbolic Coupling Maturity Model

Metric	$n$	$\chi^{2}(3)$	$p$	$W$	C1 vs. C3	Verdict
TF	20	9.42	.024	.157	$p_{\text{corr}}=1.0$	Omnibus only †
MA	15	20.71	$<$ .001	.460	$p_{\text{corr}}<.001$ ***	Significant
RC	15	14.30	.003	.318	$p_{\text{corr}}=.32$	Omnibus only ‡
RS	15	27.64	$<$ .001	.614	$p_{\text{corr}}=.003$ **	Significant
† Omnibus significant but pairwise C1 $\to$ C3 not significant on any model (see text).
‡ Pairwise C1 $\to$ C3 significant on Qwen 2.5 72B ( $p_{\text{corr}}=.019$ ); see Table 6 .

Table 2: Friedman test [ Friedman, 1937 ] results (

n

= tasks per metric,

k=4

conditions, 5 industries).

W

= Kendall’s coefficient of concordance (small:

{<}.06

; medium:

.06

–

.14

; large:

{>}.14

). Post-hoc: Wilcoxon signed-rank with Bonferroni correction (

m=3

selected pairs). Task counts: TF

n=20

(15 terminology + 5 cross-cutting); MA, RC, RS each

n=15

(10 category-specific + 5 cross-cutting scored on all metrics).

Figure 1: Four-metric profile by grounding condition (5 industries, 600 runs, primary model). C2 (RAG) and C3 (Ontology) both expand the profile relative to C1 (Ungrounded), with the largest gains on MA and RS. C4 (Ontology+Process) closely tracks C3.

Figure 2: Mean scores by condition for each metric (5 industries, 600 runs). MA and RS show the clearest separation between C1 and grounded conditions. High variance reflects task-level heterogeneity in LLM parametric knowledge across English and Vietnamese domains.

Task	Metric	C1	C2	C3	C4	$\Delta$
Largest improvements (C3 $>$ C1):
FT-T10 (BNPL handoff)	MA	.27	.87	.93	1.00	+.66
IN-T5 (UW cycle time)	MA	.07	.00	.69	.69	+.62
BV-T10 (loan approval)	TF	.30	.90	.90	.90	+.60
HC-T4 (readmission/HAI)	MA	.42	1.00	1.00	1.00	+.58
IVN-T8 (bancassurance)	RS	.39	.71	.96	.91	+.57

Table 3: Selected per-task scores (tested metric, mean of 3 reps).

\Delta

= C3

-

C1.

Industry	C1	C2	C3	C4	$\Delta_{\text{C1}\to\text{C3}}$	Note
FinTech	.66	.94	.77	.85	+.12	Strong LLM baseline
Insurance	.71	.78	.79	.77	+.08	Inverse PKE on TF
Healthcare	.63	.92	.81	.84	+.17	Large MA gain
Banking VN	.47	.77	.76	.81	+.29	Largest $\Delta$
Insurance VN	.45	.72	.74	.75	+.28	Strong $\Delta$
Grand	.58	.83	.77	.80	+.19

Table 4: Mean scores by industry and condition (

n=120

per industry, balanced design). Grand means are simple averages across the five industries. Each cell averages all four metrics across conditions and repetitions for that industry.

Figure 3: C3 (Ontology) scores by industry and metric. Vietnamese industries (banking_vn, insurance_vn) show lower TF scores but strong RS, consistent with the inverse parametric knowledge effect.

Figure 4: Ontology improvement ( $\Delta_{\text{C1}\to\text{C3}}$ ) by industry and metric. Complements Figure 3 by visualizing change rather than absolute C3 scores: green cells indicate improvement; red cells indicate regression. Vietnamese industries show the largest positive deltas on MA and RS, while well-known English terminology (TF) sometimes regresses, consistent with the Inverse Parametric Knowledge Effect.

Capability	RAG	Fine-tuning	Ontology
Domain terminology	Partial	Strong	Strong
Regulatory constraints	Partial †	Implicit	Explicit
Role-aware reasoning	Indirect †	None	Native
Workflow enforcement	None	None	Native
Multi-industry reuse	Low	None	High
Formal verifiability	None	None	Yes

Table 5: Comparison of Grounding Approaches for Enterprise AI Agents

Metric	Model	$\chi^{2}(3)$	$p$	$W$	C1 vs. C3 ( $p_{\text{corr}}$ )
TF	Claude Sonnet 4	9.42	.024*	.157	1.000
	Qwen 2.5 72B	13.54	.004**	.226	.693
	Gemma 4 26B	10.41	.015*	.173	.088
MA	Claude Sonnet 4	20.71	$<$ .001***	.460	$<$ .001***
	Qwen 2.5 72B	22.52	$<$ .001***	.501	$<$ .001***
	Gemma 4 26B	25.14	$<$ .001***	.559	$<$ .001***

Table 6: Cross-model replication of Friedman test results. Columns report Friedman

\chi^{2}(3)

p

-value, Kendall’s

W

, and post-hoc C1 vs. C3 Bonferroni-corrected

p

-value. All three models show significant omnibus effects on MA and RS; the Inverse PKE replicates across all models.

Figure 5: Ontology lift (

\Delta_{\text{C1}\to\text{C3}}

) by industry across three generator models. Vietnamese industries (shaded region) consistently show larger improvement than English industries across all models. Open-source models (Qwen, Gemma) benefit more than Claude, supporting the Inverse PKE at both domain and model levels.

Figure 6: Semantic entropy change (

\Delta H

, C1

\to

C3) by metric and model. Negative values indicate entropy reduction (constructive grounding); positive values indicate entropy increase (destructive interference). 11 of 12 metric

\times

model combinations show entropy reduction. The single exception—MA on Claude—is an empirical signature of the Inverse PKE: Claude’s strong parametric metric knowledge is disrupted by ontological injection.

Ontology-Based Reasoning

See all 15 papers

Aug 24, 2026

Toward Effective and Reliable LLM Agents via Dynamic Ontology

Large language model (LLM) agents rely heavily on knowledge encoded in model parameters or presented as unstructured context. In domain-specific tasks, this leaves important semantic connections implicit. This often results in incomplete evidence use and brittle multi-step decisions. Ontologies offer a way to externalize domain concepts and relations as machine-interpretable structures, but constructing task-usable ontologies traditionally requires substantial effort from domain experts and is difficult to scale. Automatic construction is also challenging: an ontology that appears semantically plausible may not contain the relational structures needed for actual decision making. We present OaK, an ontology-as-a-kernel framework that dynamically constructs and refines task-oriented ontologies for LLM agents. Given task requirements and training data, OaK constructs an ontology and its knowledge graph, generates task-adaptation functions for graph reasoning, and uses judge feedback to iteratively refine both. By making relevant concepts and relations explicit, the ontology grounds knowledge retrieval and multi-step decision making. We evaluate OaK on TravelPlanner, CRMArenaPro, and ToolQA. Results show that OaK improves standard LLM agents, strengthens evidence grounding, and boosts the reliability of multi-step reasoning.

Xiaohui Zhang, Zequn Sun, Chengyuan Yang +3

State Key Laboratory for Novel Software Technology, Nanjing University, China · School of Computer Science, Nanjing University of Information Science and Technology, China · School of Intelligence Science and Technology, Nanjing University, China +1

LLM Agents LLM Agent Reliability LLM Reasoning with Graphs KG Construction

Ask PDF

Jun 3, 2026

Provably Auditable and Safe LLM Agents from Human-Authored Ontologies

We introduce the LLM agent architecture Agentic Redux, intended for use with nontrivial problem domains that require linear auditability. Using the typed lambda calculus, we prove that, run on appropriate domains, Agentic Redux executions are semantically guaranteed to be correct, with all decisions recorded in an append-only ledger. We present two production-grade appropriate domains, in healthcare billing compliance, and security vulnerability disclosure. Working code for Agentic Redux run on both domains is available in a supporting code repository. We also introduce Ontology-First Agent Design, a methodology for creation of agent frameworks on a problem domain, in which a human expert ontologizes the problem domain with Basic Formal Ontology, and then assigns an LLM to derive roles that agents and humans-in-the-loop can fill, in order to work the problems in the domain.

Aaron Sterling

Thistleseeds

AI Agent Auditing LLM Auditing LLM Agent Safety LLM Agent Reliability

Ask PDF

Mar 11, 2026

Unifying Ontology Construction and Semantic Alignment for Deterministic Enterprise Reasoning at Scale

While enterprises amass vast quantities of data, much of it remains chaotic and effectively dormant, preventing decision-making based on comprehensive information. Existing neuro-symbolic approaches rely on disjoint pipelines and struggle with error propagation. We introduce the large ontology model (LOM), a unified framework that seamlessly integrates ontology construction, semantic alignment, and logical reasoning into a single end-to-end architecture. LOM employs a construct-align-reason (CAR) pipeline, leveraging its unified architecture across all three stages: it first autonomously constructs a domain-specific ontological universe from raw data, then aligns neural generation with this structural reality using a graph-aware encoder and reinforcement learning, and finally executes deterministic reasoning over the constructed topology, node attributes and relation types. We evaluate LOM on a comprehensive benchmark constructed from diverse real-world enterprise datasets. Experimental results demonstrate that LOM-4B achieves 88.8% accuracy in ontology completion and 94% in complex graph reasoning tasks, significantly outperforming state-of-the-art LLMs. These findings validate that autonomous logical construction is essential for achieving deterministic, enterprise-grade intelligence.

Hongyin Zhu

Yonyou AI Lab · Yonyou Network Technology Co., Ltd.

Semantic Alignment Ontology Learning Neuro-Symbolic Reasoning

Ask PDF

Abstract

Ontology-Based Reasoning

Explore similar work

Toward Effective and Reliable LLM Agents via Dynamic Ontology

Provably Auditable and Safe LLM Agents from Human-Authored Ontologies

Unifying Ontology Construction and Semantic Alignment for Deterministic Enterprise Reasoning at Scale