Andreas Soularidis

Mr. Andreas Soularidis

orcid

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 February 2025 in Applied Sciences
Verified
Journal Article
Open access
PDF file
DDDimitrios Doumanas
Dimitrios Doumanas
ASAndreas Soularidis
Andreas Soularidis
Dimitris SpiliotopoulosDimitris Spiliotopoulos
Dimitris Spiliotopoulos
Costas VassilakisCostas Vassilakis
Costas Vassilakis
Konstantinos KotisKonstantinos Kotis
Konstantinos Kotis
Ontology engineering (OE) plays a critical role in modeling and managing structured knowledge across various domains. This study examines the performance of fine-tuned large language models (LLMs), specifically GPT-4 and Mistral 7B, in efficiently automating OE tasks. Foundational OE textbooks are used as the basis for dataset creation and for feeding the LLMs. The methodology involved segmenting texts into manageable chapters, generating question–answer pairs, and translating visual elements into description logic to curate fine-tuned datasets in JSONL format. This research aims to enhance the models’ abilities to generate domain-specific ontologies, with hypotheses asserting that fine-tuned LLMs would outperform base models, and that domain-specific datasets would significantly improve their performance. Comparative experiments revealed that GPT-4 demonstrated superior accuracy and adherence to ontology syntax, albeit with higher computational costs. Conversely, Mistral 7B excelled in speed and cost efficiency but struggled with domain-specific tasks, often generating outputs that lacked syntactical precision and relevance. The presented results highlight the necessity of integrating domain-specific datasets to improve contextual understanding and practical utility in specialized applications, such as Search and Rescue (SAR) missions in wildfire incidents. Both models, despite their limitations, exhibited potential in understanding OE principles. However, their performance underscored the importance of aligning training data with domain-specific knowledge to emulate human expertise effectively. This study, based on and extending our previous work on the topic, concludes that fine-tuned LLMs with targeted datasets enhance their utility in OE, offering insights into improving future models for domain-specific applications. The findings advocate further exploration of hybrid solutions to balance accuracy and efficiency.

28 January 2024 in Electronics
Verified
Journal Article
Open access
PDF file
ASAndreas Soularidis
Andreas Soularidis
Konstantinos KotisKonstantinos Kotis
Konstantinos Kotis
George VourosGeorge Vouros
George Vouros
Natural disasters such as earthquakes, floods, and forest fires involve critical situations in which human lives and infrastructures are in jeopardy. People are often injured and/or trapped without being able to be assisted by first responders on time. Moreover, in most cases, the harsh environment jeopardizes first responders by significantly increasing the difficulty of their mission. In such scenarios, time is crucial and often of vital importance. First responders must have a clear and complete view of the current situation every few seconds/minutes to efficiently and timely tackle emerging challenges, ensuring the safety of both victims and personnel. Advances in related technology including robots, drones, and Internet of Things (IoT)-enabled equipment have increased their usability and importance in life- and time-critical decision support systems such as the ones designed and developed for Search and Rescue (SAR) missions. Such systems depend on efficiency in their ability to integrate large volumes of heterogeneous and streaming data and reason with this data in (near) real time. In addition, real-time critical data integration and reasoning need to be performed on edge devices that reside near the missions, instead of using cloud infrastructure. The aim of this paper is twofold: (a) to review technologies and approaches related to real-time semantic data integration and reasoning on IoT-enabled collaborative entities and edge devices in life- and time-critical decision support systems, with a focus on systems designed for SAR missions and (b) to identify open issues and challenges focusing on the specific topic. In addition, this paper proposes a novel approach that will go beyond the state-of-the-art in efficiently recognizing time-critical high-level events, supporting commanders and first responders with meaningful and life-critical insights about the current and predicted state of the environment in which they operate.

19 December 2025 in The Knowledge Engineering Review
Verified
Journal Article
Open access
PDF file
ASAndreas Soularidis
Andreas Soularidis
DDDimitrios Doumanas
Dimitrios Doumanas
Konstantinos KotisKonstantinos Kotis
Konstantinos Kotis
GVGeorge A. Vouros
George A. Vouros
Motivated by the astonishing capabilities of large language models (LLMs) in text-generation, reasoning, and simulation of complex human behaviors, in this paper, we propose a novel multi-component LLM-based framework, namely LLM4ACOE, that fully automates the collaborative ontology engineering (COE) process using role-playing simulation of LLM agents and retrieval augmented generation (RAG) technology. The proposed solution enhances the LLM-powered role-playing simulation with RAG ‘feeding’ the LLM with three different types of external knowledge. This knowledge corresponds to the knowledge required by each of the COE roles (agents), using a component-based framework, as follows: (a) domain-specific data-centric documents, (b) OWL documentation, and (c) ReAct guidelines. The aforementioned components are evaluated in combination, with the aim of investigating their impact on the quality of generated ontologies. The aim of this work is twofold, (a) to identify the capacity of LLM-based agents to generate acceptable (by human-experts) ontologies through agentic collaborative ontology engineering (ACOE) role-playing simulation, at specific levels of acceptance (accuracy, validity, and expressiveness of ontologies) without human intervention and (b) to investigate whether and/or to what extent the selected RAG components affect the quality of the generated ontologies. The evaluation of this novel approach is performed using ChatGPT-o in the domain of search and rescue (SAR) missions. To assess the generated ontologies, quantitative and qualitative measures are employed, focusing on coverage, expressiveness, structure, and human involvement.

21 June 2024
Verified
Book Chapter
Closed access
PDF file
DDDimitrios Doumanas
Dimitrios Doumanas
ASAndreas Soularidis
Andreas Soularidis
Konstantinos KotisKonstantinos Kotis
Konstantinos Kotis
George VourosGeorge Vouros
George Vouros

15 September 2025 in AI
Verified
Journal Article
Open access
PDF file
DDDimitrios Doumanas
Dimitrios Doumanas
AKAlexandros Karakikes
Alexandros Karakikes
ASAndreas Soularidis
Andreas Soularidis
EMEfstathios Mainas
Efstathios Mainas
Konstantinos KotisKonstantinos Kotis
Konstantinos Kotis
The rapid proliferation of Large Language Models (LLMs) has democratized access to advanced generative capabilities while raising urgent concerns about misuse in sensitive security domains. Border security, in particular, represents a high-risk environment where malicious actors may exploit LLMs for document forgery, synthetic identity creation, logistics planning, or disinformation campaigns. Existing studies often highlight such risks in theory, yet few provide systematic empirical evidence of how state-of-the-art LLMs can be exploited. This paper introduces the Silent Adversary Framework (SAF), a structured pipeline that models the sequential stages by which obfuscated prompts can covertly bypass safeguards. We evaluate ten high-risk scenarios using five leading models—GPT-4o, Claude 3.7 Sonnet, Gemini 2.5 Flash, Grok 3, and Runway Gen-2—and assess outputs through three standardized metrics: Bypass Success Rate (BSR), Output Realism Score (ORS), and Operational Risk Level (ORL). Results reveal that, while all models exhibited some susceptibility, vulnerabilities were heterogeneous. Claude showed greater resistance in chemistry-related prompts, whereas GPT-4o and Gemini generated highly realistic outputs in identity fraud and logistics optimization tasks. Document forgery attempts produced only partially successful templates that lacked critical security features. These findings highlight the uneven distribution of risks across models and domains. By combining a reproducible adversarial framework with empirical testing, this study advances the evidence base on LLM misuse and provides actionable insights for policymakers and border security agencies, underscoring the need for stronger safeguards and oversight in the deployment of generative AI.

7 January 2026 in AI
Verified
Journal Article
Open access
PDF file
DDDimitrios Doumanas
Dimitrios Doumanas
ASAndreas Soularidis
Andreas Soularidis
Konstantinos KotisKonstantinos Kotis
Konstantinos Kotis
Military decision-making is inherently complex and highly critical, requiring commanders to assess multiple variables in real-time, anticipate second-order effects, and adapt strategies based on continuously evolving battlefield conditions. Traditional approaches rely on domain expertise, experience, and intuition, often supported by decision-support systems designed by military experts. With the rapid advancement of Large Language Models (LLMs) such as ChatGPT, Claude, and DeepSeek, a new research question emerges: can LLMs perform causal reasoning at a level that could meaningfully replace human decision-makers, or should they remain human-led decision-support tools in high-stakes environments? This paper explores the causal reasoning capabilities of LLMs for operational and strategic military decisions. Unlike conventional AI models that rely primarily on correlation-based predictions, LLMs are now able to engage in multi-perspective reasoning, intervention analysis, and scenario-based assessments. We introduce a structured empirical evaluation framework to assess LLM performance through 10 de-identified real-world-inspired battle scenarios, ensuring models reason over provided inputs rather than memorized data. Critically, LLM outputs are systematically compared against a human expert baseline, composed of military officers across multiple ranks and years of operational experience. The evaluation focuses on precision, recall, causal reasoning depth, adaptability, and decision soundness. Our findings provide a rigorous comparative assessment of whether carefully prompted LLMs can assist, complement, or approach expert-level performance in military planning. While fully autonomous AI-led command remains premature, the results suggest that LLMs can offer valuable support in complex decision processes when integrated as part of hybrid human-AI decision-support frameworks. Since our evaluation directly tests this capability, this paradigm shift raises fundamental question: Is there a possibility to fully replace high-ranking officers/commanders in leading critical military operations, or should AI-driven tools remain as decision-support systems enhancing human-driven battlefield strategies?