Sun, Oct 4 (Room: Torremolinos C) |
Session 1: Foundations of Multi-Level Modeling
9:15 – 10:30
|
|
Chair: Pierre Maier
|
| 9:15 |
Welcome and Introduction |
| 9:20 |
Modeled Field Semantics: Making Field Behavior Explicit in Multi-Level Modeling
Gergely Mezei, Ferenc Attila Somogyi and Gergely Gembela
Abstract:
Multi-level modeling provides established mechanisms for describing how fields
characterize elements across multiple abstraction steps.
Potency-based approaches are especially useful when the intended depth
of characterization is known and can be expressed directly in the modeling
language. This paper presents a different solution, referred to as Modeled
Field Semantics (MFS). Instead of defining field behavior only as a built-in
language mechanism, MFS represents certain aspects of field behavior explicitly
in the model.
The motivation for MFS comes from cases where field behavior is not limited to assigning a value at a predetermined
abstraction depth. A value may be interpreted differently by classifier-objects and
instance-objects, and the differences may also become part of the field structure
itself, for example when a general field later continues as several distinct fields
representing different roles. In MFS, such decisions are not treated only as
consequences of predefined field semantics, but can be expressed in the model
through explicit field relationships, assignment rules, and value access operations.
MFS treats selected aspects of field behavior as part of the modeled semantics. Field lineage, value-assignment rules,
getter and setter behavior, and their configuration can be represented explicitly
instead of being fully pre-encoded as one fixed language semantics. The result is a
way to describe field semantics in which structural behavior and value interpretation
remain explicit, adjustable, and subject to validation within the model.
|
| 9:55 |
Deep Reactions: Realizing and Evaluating Consistency Preservation for Multi-Level, Multi-View Models
Annika Kienle, Arne Lange, Thomas Weber, Lars König, Nathan Hagel and Erik Burger
Abstract:
In view-based development, consistency plays a central role across concrete models.
Consistency can be achieved automatically through consistency-preservation rules defined, for instance, in a change-driven language such as the Reactions language of the Vitruvius framework.
This language, however, assumes the classical two-level architecture with a strict separation of type and instance level and can therefore only handle multi-level models through their generic linguistic structure.
This forces domain questions to be encoded as not statically typed string operations on infrastructure metaclasses.
Based on a previously published vision, this paper presents the Deep Reactions language.
It lets a stable region of a deep model serve as an ontological metamodel, over which ontological types and features can be referenced directly and deep characterization through offspring-aware triggers can be exploited.
Building on the stable code generation of the original Reactions language, a pre-processing step first transforms Dep Reactions into regular Reactions which ensures compatibility with existing implementations.
Our evaluation shows that the number of referenced metamodels and the token count can be reduced while removing 83% of statically unvalidated type references.
|
| 10:30 |
Coffee Break |
Session 2: Practical Considerations of Multi-Level Modeling
11:00 – 12:45
|
|
Chair: Gergely Mezei
|
| 11:00 |
Multi-Level Modeling as the Execution Substrate of a No-Code Platform: An Industrial Experience with Metada
Petr Smolík
Abstract:
We report on an industrial realization of multi-level modeling
(MLM): Metada, a commercial enterprise no-code application platform
whose runtime interprets models directly rather than generating
code. Its foundation is a two-layered meta-architecture —
RBSM, a linguistic meta-metamodel offering one deep instantiation
relationship with per-property propagation controls, and Meta4, an
ontological metamodel of concepts and typed properties built on
RBSM. We claim no new multi-level concepts: RBSM’s primitives
align with the orthogonal classification architecture, characterization
potency, and the durability/mutability vitality properties
established in prior work. Our contribution is an engineering and
experience one — how these concepts behave once they become
the substrate of a persisted, multi-tenant business-application runtime,
integrated end-to-end with storage, an operational constraint
layer, and model-driven user interfaces. We use a published MLM
challenge domain as a worked, executable, test-covered example,
contrast our engineering decisions with five research approaches
(DeepTelos, DMLA, FMMLx, DLM, Melanee), and draw on three production
and reference models — an ArchiMate metamodel reused
by folder inheritance, a multi-tenant ticketing platform, and a 6000-
product banking catalog that instantiates down to a client-holding
level (a three-level chain). We distill lessons, most notably that deep
values are inherited by derivation, caching, and search-indexing
rather than duplication, and that integrity is best expressed as
platform-native operations rather than a declarative constraint
language.
|
| 11:35 |
Transforming a large Flat Model into a Multi-Level Model: A Retrospective Case Study
Ulrich Frank
Abstract:
This paper presents a case study comprising the analysis and transformation of a large
flat object model into a multi-level model. To support the transformation, a
corresponding method is introduced and applied. The resulting multi-level model is
evaluated against the original requirements as well as with regard to further aspects.
|
| 12:10 |
Property-Precedence Analysis for Model Deepening
Pierre Maier
Abstract:
Despite widely accepted prospects, multi-level modeling remains sparsely adopted in
industry. Model deepening –, i.e., the re-engineering of flat
representations into multi-level models – has the potential to facilitate the
industrial adoption of multi-level modeling by transforming models and representations
of information systems an organization already possesses. Prior research has focused
primarily on transformation operations; comparatively little attention has been devoted
to the preceding analysis problem of identifying beneficial model-deepening
transformations. Existing approaches typically rely on type-level heuristics such as
naming patterns, aggregation relationships, or association multiplicities.
Since these heuristics suggest multi-level hierarchies based on modeling choices and
conventions rather than the represented instance population, they cannot reliably
distinguish genuine classification levels from incidental modeling choices.
This paper proposes property-precedence analysis, an instance-driven approach to
conduct model-deepening analysis. Building upon Bunge's ontological notion of property
precedence, the proposed approach uncovers possible multi-level hierarchies by
analyzing the distribution of property values across model instances rather than
relying on type-level heuristics. The analysis is implemented in the prototype
ModelDeepener for the MLM language
FMMLx and demonstrated using two example models involving
associations and attributes. Several challenges in property-precedence analysis
for model deepening remain, such as handling conflicting precedence relations and
naming newly suggested classes. Further research should investigate to what degree
the identified challenges may be counteracted and how property-precedence analysis
may be complemented with type-level heuristics for model deepening.
|
| 12:45 |
Lunch |
Session 3: Support for Creating and Managing Multi-Level Models
14:30 – 15:45
|
|
Chair: Georg Grossmann
|
| 14:30 |
Assessing Capabilities of Large Language Models for Multi-Level Modeling
Arianna Fedeli, Ludovico Iovino, Maria Teresa Rossi and Adrian Rutle
Abstract:
Large Language Models (LLMs) are increasingly being explored in Model-Driven
Engineering as assistants for generating, transforming, and evolving modeling
artifacts from natural language descriptions. Existing work, however, has mainly
focused on traditional two-level modeling settings, while the applicability of
LLMs to multilevel modeling remains largely unexplored. Multilevel modeling
introduces specific challenges, including reasoning across multiple abstraction
levels, handling deep instantiation, preserving cross-level consistency, and
distinguishing between types, instances, and linguistic classifications.
This paper investigates the capabilities and limitations of LLMs in supporting multilevel modeling tasks from
natural language descriptions. We propose a scenario-based evaluation framework
that combines two representative case studies from the MULTI challenge series with
a conventional two-level modeling benchmark. The framework evaluates LLMs across six
representative MLM tasks, covering hierarchy navigation, additive and subtractive
model evolution, structural modifications, instance generation, and complete
hierarchy synthesis. The results show that LLMs can support exploratory MLM tasks,
but their outputs remain candidate artifacts that require expert validation,
especially when correctness depends on cross-level reasoning and the preservation
of consistency.
|
| 15:05 |
Managing Model Complexity: A Segment-Based Approach
Thomas Kühne and Pierre Maier
Abstract:
Multi-level modeling (MLM) is typically characterized as minimizing
accidental complexity. We observe that MLM levels also aid in managing
essential model complexity by separating and organizing model content
based on the domain classification level it represents. Building on this insight,
we generalize the notion of a level—and previously recognized pluralities of
elements, e.g., those contained in packages—to that of a model segment.
In addition to solution-based segments, we recognize domain-induced segments, such
as domain diversity allowing semantic information to be represented in
models that was previously lost to mapping different modeler
intents to the same modeling constructs. These new, semantically relevant, model
segments support better model navigability and aid model understanding by facilitating
the isolation of particular model aspects. Crucially, all model segments are subject to
being represented by proxies, thus enabling a significant reduction of first-contact
model complexity. We posit that introducing segments for managing model complexity has
the potential to lower the threshold of MLM adoption due to its ability to counteract
the often-held perception of MLM models being complex.
|
| 15:45 |
Coffee Break |
Session 4: Emerging Application Areas of Multi-Level Modeling
16:15 – 17:35
|
|
Chair: Ulrich Frank
|
| 16:15 |
Deep Instantiation Meets Dependent Typing – The Potential of Coding Potency in Potent Type Systems
Friedrich Steimann and Ralf Lämmel
Abstract:
Multi-level modelling (MLM) extends conventional modelling by
allowing classification hierarchies that span more than two levels,
enabling concepts such as deep instantiation and level-crossing relationships.
While several formal foundations for MLM have been proposed, the potential of
contemporary dependent type systems as a semantic basis has received comparatively
little attention. This paper therefore explores the idea of encoding multi-level
models in the type systems of languages such as Scala, Lean, and Haskell, which
feature dependent types, type families, and other advanced typing constructs.
Although we limit ourselves to a simple setup of deep instantiation here, our
findings suggest that this idea may indeed be valuable for exploring the design
space of MLM languages.
|
| 16:35 |
Multi-Level Modeling for Interactive Scene Description and Content Creation
Gergely Gembela, Gergely Mezei and Ferenc Attila Somogyi
Abstract:
Multi-level modeling (MLM) approaches and frameworks are reaching a stage where research
and development of both conceptual frameworks and tooling would greatly benefit from
feedback and experience from real-world use cases and industrial applications.
MLM-based solutions could address the ongoing challenge of maintaining flexibility,
full control, and semantic integrity throughout the development process of video games,
as players expect increasingly vast, unique, and highly detailed environments.
In this position paper, we propose the concept of a multi-level model-based framework: Multi-Layer Content Model (MLCM),
that utilizes MLM concepts to describe, store, and transport Digital Content
Creation (DCC) assets, serving as a universal interconnect between DCC tools,
game engines, and other DCC-related tools. MLCM also has potential to serve as a
data source for both traditional and AI-based Procedural Content Generation (PCG)
tools. Using MLCM, structural, behavioral, and generative descriptions are integrated
into a single model, supporting iterative development, modular extension, and
long-term maintainability. Beyond serving as a compact representation of generative
parameters and rules, MLCM facilitates verification of model elements and, potentially,
the resulting content, while also serving as a prime case study for the industrial
application of MLM.
|
| 16:55 |
Multi-Level Modeling Pressure in the FSL Ontology – When OWL Punning Is Not Enough
Ralf Lämmel and Friedrich Steimann
Abstract:
The ‘Foundations of Software Languages’ (FSL) ontology
integrates software languages, tools, artifacts, concepts, technological spaces,
and engineering activities. FSL features types of resources that must be treated
both as objects of discourse and as classifiers, for example, when modelling artifact
kinds or language concepts. FSL leverages OWL 2 punning as a pragmatic encoding for
such dual roles, thereby serving as a case study for a domain ontology that could
benefit from proper multi-level modelling support.
|
| 17:15 |
Going beyond structure – Can Dynamic Multi-Level Algebra Support Multi-Level Simulation?
Sebastian Weber, Thomas Weber, Arne Lange, Jörg Henß and Robert Heinrich
Abstract:
Multi-level simulation (MLS) combines simulation models that differ in abstraction,
resolution, timing, or computational cost. This makes MLS attractive for
complex systems, but it also creates recurring problems: simulation artefacts
must be selected, coupled, and exchanged without losing semantic compatibility.
At the same time, the word “level” has different meanings in MLS and in multi-level
modeling (MLM). Treating simulation resolution levels as ordinary MLM classification
levels therefore risks confusing refinement, instantiation, specialization, and
simulator coupling. This position paper examines Dynamic Multi-Level Algebra (DMLA)
as a modeling basis for MLS compatibility. We argue that DMLA should be investigated
not as a direct encoding of MLS levels, but as a level-blind, refinement-based
modeling formalism for explicitly representing simulation artefacts, simulator
capabilities, compatibility constraints, and validation operations. An ECU software
architecture example illustrates how DMLA-style validation can expose missing inputs,
semantic mismatches, and unjustified simulator switches. From selected MLS and
DMLA literature, we condense recurring challenges into six modeling needs and
identify where DMLA can address them and where further work is required.
|
| 17:35 |
Closing Day 1 |
Mon, Oct 5 (Room: Mediterraneo A) |
Session 5: Keynote
9:15 – 10:30
|
|
Chair: Arne Lange
|
| 9:15 |
Keynote: Models That Must Live: Metamodeling Challenges in MBSE Tool Practice
Clemens Reichmann
Abstract:
Industrial MBSE tools are subject to a permanent tension: simplicity versus expressiveness, short-term
modelability versus long-term consistency. This keynote offers a practitioner's perspective on
metamodel development, drawing on experience with PREEvision, a development environment for automotive
E/E systems. It addresses questions that arise repeatedly in day-to-day tooling work: What drives the
complexity of a metamodel? How can modeling layers be kept consistent when partial models are versioned
and evolved independently? The keynote discusses where the classical two-level paradigm reaches its
limits – and what impulses multi-level modeling can offer for industrial practice.
|
| 10:30 |
Coffee Break |
Session 6: MULTI Vector Industry Challenge and Closing Discussion
11:00 – 12:45
|
|
Chair: Arne Lange
|
| 11:00 |
MULTI Vector Industry Challenge
Arne Lange, Clemens Reichmann, Thomas Kühne, Pierre Maier and Thomas Weber
Abstract:
The MULTI workshop series has set a number of multi-level modeling challenges,
each designed to allow competing multi-level modeling approaches to demonstrate
their capabilities and/or to tease out their limitations. The challenges,
therefore, have been serving a three-fold purpose: First, they have allowed
technologies to demonstrate their abilities. Second, they have pointed out where
technologies still fall short of providing optimal modeling support.
Third, they have provided a basis for comparing competing technologies,
often revealing the trade-offs implied by certain design choices.
The MULTI Vector Industry Challenge described in this paper is the fifth installment
in this series. It is unique among the existing challenges in that it has
been developed in collaboration with the industry partner
Vector
and is directly inspired
by a respective real-world modeling challenge.
|
| 11:15 |
Poster Session MULTI Vector Industry Challenge
- Contractual Type-Square Pattern and Its Lack of Necessity in DLMA
[Poster]
Áron Furmann, Gergely Gembela, Ferenc Attila Somogyi, Gergely Mezei and Dániel Palatinszky
- A PAMoLa Solution to the MULTI 2026 Vector Industry Challenge
[Poster]
Shilpi Gupta, Mohammad Sadeghi, Monalisha Ojha and Colin Atkinson
- Ensuring Consistency via Deep Composition Semantics
[Poster]
Thomas Kühne and Arne Lange
- An Algorithmic Solution to the Vector Challenge
[Poster]
Arne Lange and Romain Pascual
- Native Deep Instantiation for Automotive Connector Catalogs on the Metada
No-Code Platform
[Poster]
Petr Smolík
|
| 12:00 |
Future of MULTI: Closing Discussion |
Industrial MBSE tools are subject to a permanent tension: simplicity versus expressiveness, short-term
modelability versus long-term consistency. This keynote offers a practitioner's perspective on
metamodel development, drawing on experience with PREEvision, a development environment for automotive
E/E systems. It addresses questions that arise repeatedly in day-to-day tooling work: What drives the
complexity of a metamodel? How can modeling layers be kept consistent when partial models are versioned
and evolved independently? The keynote discusses where the classical two-level paradigm reaches its
limits – and what impulses multi-level modeling can offer for industrial practice.
