Can LLMs Faithfully Solve Ontology Engineering Tasks? A Mechanistic Interpretability Study

Bohui Zhang; Albert Meroño-Peñuela; Elena Simperl

cs.AIOct 28, 2026

Can LLMs Faithfully Solve Ontology Engineering Tasks? A Mechanistic Interpretability Study

Authors: Bohui Zhang, Albert Meroño-Peñuela, Elena Simperl

Organizations: King’s College London, United Kingdom · Technical University of Munich, Germany

Abstract

Ontological knowledge underpins large language model (LLM) performance on knowledge-intensive tasks. However, existing evaluations of LLMs' ontological understanding largely rely on behavioral correctness, leaving the underlying mechanisms unexplained. Existing mechanistic interpretability work has largely focused on instance-level factual knowledge, leaving ontological knowledge under-explored. To bridge this gap, we introduce MIOE, a causal-abstraction benchmark that evaluates interchange interventions for two ontology engineering tasks: entity typing and subsumption prediction. MIOE tests whether learned internal intervention targets can change an ontology-level target answer while preserving class-specific auxiliary information. Across Qwen3-8B and Qwen3-4B, we evaluate SAE-, DBM-, and DAS-based methods, including multi-task and multi-layer variants. Distributed subspace methods give the strongest results and generalize across prompt settings, but the induced changes remain only partially separated from auxiliary class-specific information. Multi-task objectives and contiguous layer-window interventions improve this balance, while class-level analyses reveal substantial variation across ontology classes. These results suggest that ontological predictions are mechanistically accessible in LLMs, although current methods do not yet isolate a cleanly disentangled causal mechanism for ontological knowledge.

Figures & tables

Axiom Type	# Classes	# Subjects	# Attributes (avg./class)	Avg. # Attribute Values (per subj. / per attr.)	# Prompt Templates (avg./class)
instance of	54	1,099	376 (7.0)	12.0 / 2.0	4,288 (79.4)
subclass of	39	1,113	160 (4.1)	51.8 / 17.1	2,095 (53.7)

Table 1: Dataset statistics by axiom type, including counts of domain classes, subjects, auxiliary attributes and values, and prompt templates. For auxiliary attributes, we report the average number of values per subject and per instantiated attribute.

Method	instance of								subclass of
	Dimension	Layer	Entity			Context			Dimension	Layer	Entity			Context
	Dimension	Layer	C	I	D	C	I	D	Dimension	Layer	C	I	D	C	I	D
SAE	8192	13	0.472	0.352	0.412	0.399	0.344	0.371	16384	12	0.584	0.210	0.397	0.624	0.233	0.429
DBM	4096	10	0.441	0.409	0.425	0.385	0.407	0.396	4096	10	0.528	0.334	0.431	0.526	0.359	0.443
MDBM	4096	10	0.418	0.469	0.443	0.344	0.472	0.408	4096	13	0.491	0.389	0.440	0.490	0.409	0.450
DAS	1024	17	0.589	0.579	0.584	0.582	0.512	0.547	1024	22	0.509	0.483	0.496	0.625	0.445	0.535

Table 2: Best single-layer intervention results for Qwen3-8B. C/I/D denote Cause/Iso/Disentangle. For each axiom type and method, we report the layer/configuration with the strongest Entity Disentangle score. Dimension denotes subspace rank for DAS/MDAS, ambient mask dimension for DBM/MDBM, and latent width for SAE. Qwen3-4B results are reported in the supplementary materials.

Figure 1: Representative layer profiles for Qwen3-8B on instance of . The top panel shows single-layer MDAS with rank 2048, and the bottom panel shows four-layer MDAS with

256\times 4

contiguous windows. Both panels report Cause, Iso, and Disentangle.

Figure 2: Cause–Iso scores for selected Qwen3-8B instance of configurations with different intervention dimensions. Percent labels indicate the dimensionality of

F_{\tau}

under each method’s parameterization; for multi-layer interventions, total rank is reported relative to a single-layer residual-stream width.

$k\times w$	Method	Window	Cause	Iso	Disentangle ( $\Delta$ )
instance of
$512\times 2$	DAS	23–24	0.560	0.726	0.643 (+0.059)
$512\times 2$	MDAS	22–23	0.529	0.730	0.629 (+0.038)
$256\times 4$	DAS	24–27	0.548	0.772	0.660 (+0.077)
$256\times 4$	MDAS	22–25	0.590	0.741	0.666 (+0.074)
$512\times 4$	DAS	24–27	0.577	0.756	0.667 ( +0.102 )

Table 3: Multi-layer DAS/MDAS results for Qwen3-8B. The

k\times w

column denotes a rank-

k

subspace at each layer in a contiguous

w

-layer window, and Window gives the contiguous layer span of the reported best configuration.

\Delta

reports the Disentangle improvement over the single-layer result with the same method and total dimension.

Figure 3: Class-level analysis of Qwen3-8B multi-layer MDAS for instance of . Left and middle: Cause–Iso scatter plots for base-role and source-role classes. Right: largest base–source role gaps, measured as base-role minus source-role Disentangle. Colors denote stronger base-role performance or stronger source-role performance .

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Qwen3-4B	Qwen3-8B
Layers / hidden size	36 / 2,560	36 / 4,096
Prompt / answer length	48 / at most 3 tokens	48 / at most 3 tokens
Site / position	block output / last subject subtoken	block output / last subject subtoken
Supervised train / eval batch	128 / 128	128 / 128
Epochs	6	6
Random seed	42	42

Appendix

Table 4: Experimental settings. Training pair counts are the examples passed to the supervised methods after dataset construction.

	instance of				subclass of
Model	Target Micro	Target Macro	Attr. Micro	Attr. Macro	Target Micro	Target Macro	Attr. Micro	Attr. Macro
Qwen3-4B	92.4	92.1	59.7	58.7	89.2	87.4	69.2	68.8
Qwen3-8B	94.7	94.3	64.9	63.8	93.0	91.8	69.1	68.3

Appendix

Table 5: Base LLM performance on ontology-oriented knowledge engineering queries. Target Micro/Macro denotes micro- and class-macro accuracy on target axiom queries; Attr. Micro/Macro reports the corresponding accuracy on auxiliary attribute queries.

Panel of Figure 4 — Figure 4: Class-level Cause–Iso scatter plots for the best MDAS instance-of settings. Each row shows base classes on the left and source classes on the right.

Panel of Figure 5 — Figure 5: Class-level Cause–Iso scatter plots for the best MDAS subclass-of settings. Each row shows base classes on the left and source classes on the right.

Panel of Figure 6 — Figure 6: Class-level Cause–Iso scatter plots for the best Qwen3-8B multi-layer MDAS settings. Each row corresponds to one axiom type, with base classes on the left and source classes on the right.

Model	Role	Cause	Iso	Disent.	$\boldsymbol{\Delta}$
instance of
Qwen3-8B	Base	0.539	0.674	0.606	+0.135
Qwen3-8B	Source	0.507	0.674	0.590	+0.167
Qwen3-4B	Base	0.521	0.613	0.567	+0.091
Qwen3-4B	Source	0.493	0.612	0.552	+0.119
subclass of

Appendix

Table 6: Mean scores over ontology classes in the base and source roles.

\Delta

reports

\mathrm{Iso}-\mathrm{Cause}

, with positive values indicating that changing the target answer is harder than preserving auxiliary answers and negative values indicating the reverse.

Axiom	Base cross-model $\rho$	Source cross-model $\rho$	8B base–source $\rho$	4B base–source $\rho$
instance of	0.831	0.810	0.273	0.267
subclass of	0.752	0.671	-0.082	0.059

Appendix

Table 7: Rank stability of class-level Disentangle scores. Each value is Spearman’s

\rho

over ontology classes. Base and Source cross-model

\rho

compare Qwen3-8B and Qwen3-4B class rankings within the same role. The 8B and 4B base–source

\rho

columns compare base-role and source-role rankings within the same model.

Panel of Figure 7 — Figure 7: Cross-model class-rank correlation for task-macro Disentangle scores. Each point denotes one high-level class, with coordinates given by its Qwen3-8B and Qwen3-4B ranks for the same axiom and role. Score $r$ is the Pearson correlation between the two models’ class-level Disentangle scores; rank $r$ is the Pearson correlation between their ranks, equivalently Spearman’s $\rho$ . Color gives the arithmetic mean of the two Disentangle scores. Point area increases with class-level evaluation support, using the larger support count across the two runs.

Panel of Figure 8 — Figure 8: Largest base–source class role gaps for the best MDAS settings. Rows correspond to axiom types and columns to model sizes. The horizontal axis reports base-role Disentangle minus source-role Disentangle. Colors denote stronger base-role performance or stronger source-role performance .

Panel of Figure 9 — Figure 9: Largest base–source class role gaps for the best Qwen3-8B multi-layer MDAS settings. The horizontal axis reports base-role Disentangle minus source-role Disentangle. Colors denote stronger base-role performance or stronger source-role performance .

Axiom	Role	Recurring high-scoring classes
instance of	Base	airport; city; country; currency; science award; software; stock exchange
instance of	Source	airport; city; country; currency; data set; disease; type of sport; university
subclass of	Base	animal organ; flower; measuring instrument; medical device; organization; theory; tree
subclass of	Source	activity; creative work; emotion; fish; flower; music

Appendix

Table 8: Representative recurring high-scoring classes from the class-level Disentangle rankings. Rankings are computed separately for each axiom type and base/source role on the Entity split, and listed classes are high-scoring cases for both Qwen3-8B and Qwen3-4B within the same role.

Axiom	Model	$\boldsymbol{N}$	Mean	Median [IQR]	$\boldsymbol{\Pr(\Delta\geq 0.2)}$	$\boldsymbol{\Pr(\Delta\geq 0.4)}$
instance of	Qwen3-8B	121	0.166	0.132 [0.076, 0.256]	29.8%	5.0%
instance of	Qwen3-4B	141	0.184	0.158 [0.069, 0.253]	39.0%	7.1%
subclass of	Qwen3-8B	91	0.204	0.179 [0.090, 0.287]	41.8%	12.1%
subclass of	Qwen3-4B	52	0.185	0.152 [0.075, 0.291]	36.5%	7.7%

Appendix

Table 9: Distribution of absolute reciprocal Disentangle gaps

\Delta_{A,B}

over eligible unordered class pairs for the best MDAS setting in each model–axiom combination.

N

is the number of eligible pairs, IQR denotes the interquartile range, and probability columns report the fraction of eligible pairs exceeding each threshold. Eligibility requires at least five Cause and five Iso examples in each direction.

Figure 10: Empirical survival distributions of absolute reciprocal Disentangle gaps for the best MDAS settings. At horizontal position

x

, each curve reports the fraction of eligible unordered class pairs satisfying

\Delta_{A,B}\geq x

. Curves are pair-macro: each eligible reciprocal pair contributes once, without weighting by its evaluation support.

Axiom	Method	Dimension	Layer	Entity			Context
Axiom	Method	Dimension	Layer	Cause	Iso	Disentangle	Cause	Iso	Disentangle
instance of	SAE	5120	14	0.408	0.375	0.391	0.350	0.369	0.359
	DBM	2560	10	0.458	0.354	0.406	0.400	0.350	0.375
	MDBM	2560	10	0.404	0.436	0.420	0.340	0.430	0.385
	DAS	1024	22	0.516	0.537	0.527	0.489	0.509	0.499
	MDAS	2048	16	0.520	0.592	0.556	0.471	0.639	0.555

Appendix

Table 10: Qwen3-4B single-layer intervention results under the same reporting rule as the Qwen3-8B comparison. For each axiom type and method, we report the layer/configuration with the strongest Entity Disentangle score and evaluate the same configuration on Context. Dimension denotes subspace rank for DAS/MDAS, ambient mask dimension for DBM/MDBM, and latent width for SAE.

Panel of Figure 11 — Figure 11: SAE single-layer profiles across model sizes, axiom types, and latent widths.

Panel of Figure 12 — Figure 12: Fixed-layer DBM/MDBM intervention profiles across model sizes and axiom types.

Panel of Figure 13 — Figure 13: Fixed-layer DAS/MDAS intervention profiles across model sizes and axiom types with 1024-dimensional subspaces.

Panel of Figure 14 — Figure 14: Fixed-layer DAS/MDAS intervention profiles across model sizes and axiom types with 2048-dimensional subspaces.

Panel of Figure 15 — Figure 15: Fixed-window multi-layer intervention profiles for Qwen3-8B on instance of . Each panel reports Cause, Iso, and Disentangle on 12 contiguous windows. The $a\times b$ notation denotes $a$ intervened dimensions per layer over a $b$ -layer window.

Panel of Figure 16 — Figure 16: Fixed-window multi-layer intervention profiles for Qwen3-8B on subclass of . Each panel reports Cause, Iso, and Disentangle on 12 contiguous windows. The $a\times b$ notation denotes $a$ intervened dimensions per layer over a $b$ -layer window.

Axiom	Model	Method	Dim.	Layer	Cause	Iso	Disentangle
instance of	Qwen3-4B	SAE	5120	10	0.458	0.292	0.375
		DBM	2560	9	0.463	0.323	0.393
		MDBM	2560	10	0.404	0.436	0.420
		DAS	2048	14	0.615	0.346	0.481
		MDAS	2048	13	0.524	0.559	0.541
	Qwen3-8B	SAE	16384	13	0.472	0.350	0.411

Appendix

Table 11: Highest-Cause single-layer intervention results on the Entity split, selecting the configuration with the highest Cause score for each model, axiom type, and method.

Class	# Entity	Attributes
academic conference	30	inception, main subject, event interval, organizer
academic journal	27	main subject, country of origin, publisher, ISSN, indexed in bibliographic review, inception, editor
airport	15	country, located in the administrative territorial entity, date of official opening, elevation above sea level, IATA airport code, ICAO airport code, airline hub
album	15	form of creative work, genre, performer, publication date, record label, producer, number of parts of this work, tracklist
algorithm	20	inception, named after, discoverer or inventor, computes solution to, best-case time complexity, average time complexity, best-case space complexity, worst-case time complexity, worst-case space complexity, average space complexity
animal	20	parent taxon, taxon rank, color, endemic to, main food source, gestation period, diel cycle

Appendix

Table 12: Classes in the instance of dataset, their number of entities, and associated attributes.

Class	# Concept	Attributes
academic discipline	17	field of work, has subclasses, is the study of, part of
activity	65	has subclasses, location, participant
algorithm	34	has subclasses, has use
animal organ	17	body system, function, has part(s)
astronomical object	26	has part(s), has subclasses, parent astronomical body
award	17	conferred by, field of work, has subclasses

Appendix

Table 13: Classes in the subclass of dataset, their number of subclasses, and associated attributes.

Mechanistic Interpretability

Mechanistic interpretability studies how neural networks implement behaviors through internal features, circuits, and causal information flow.

See all 394 papers

Agent-OM: Leveraging LLM Agents for Ontology Matching
Dec 1, 2023Zhangcheng Qiang, Weiqing Wang, Kerry TaylorSemantic Mapping
Australian National University, Canberra, ACT, Australia · Monash University, Melbourne, VIC, Australia
From Subsumption to Satisfiability: LLM-Assisted Active Learning for OWL Ontologies
Apr 17, 2026Haoruo Zhao, Wenshuo Tang, Duncan Guthrie +3Active Learning Ontology Learning
University of Glasgow
OntoLearner: A Modular Python Library for Ontology Learning with Large Language Models
Jul 2, 2026Hamed Babaei Giglou, Jennifer D'Souza, Andrei Aioanei +2Relation Extraction Ontology Learning
TIB – Leibniz Information Centre for Science and Technology, Hannover, Germany · 3IBM Research, New York, USA · 1TIB – Leibniz Information Centre for Science and Technology, Hannover, Germany +1

Dec 1, 2023J/K move · Enter open · S save

Agent-OM: Leveraging LLM Agents for Ontology Matching

Zhangcheng Qiang, Weiqing Wang, Kerry Taylor

Australian National University, Canberra, ACT, Australia · Monash University, Melbourne, VIC, Australia

Semantic Mapping

Open paper AskPDF arXiv

Ontology matching (OM) enables semantic interoperability between different ontologies and resolves their conceptual heterogeneity by aligning related entities. OM systems currently have two prevailing design paradigms: conventional knowledge-based expert systems and newer machine learning-based predictive systems. While large language models (LLMs) and LLM agents have revolutionised data engineering and have been applied creatively in many domains, their potential for OM remains underexplored. This study introduces a novel agent-powered LLM-based design paradigm for OM systems. With consideration of several specific challenges in leveraging LLM agents for OM, we propose a generic framework, namely Agent-OM (Agent for Ontology Matching), consisting of two Siamese agents for retrieval and matching, with a set of OM tools. Our framework is implemented in a proof-of-concept system. Evaluations of three Ontology Alignment Evaluation Initiative (OAEI) tracks over state-of-the-art OM systems show that our system can achieve results very close to the long-standing best performance on simple OM tasks and can significantly improve the performance on complex and few-shot OM tasks.

Abstract

Mechanistic Interpretability

Explore similar work

Agent-OM: Leveraging LLM Agents for Ontology Matching

From Subsumption to Satisfiability: LLM-Assisted Active Learning for OWL Ontologies

OntoLearner: A Modular Python Library for Ontology Learning with Large Language Models