Misaligned Behaviors and Dangerous Capabilities in AI Safety: A Conceptual Cross-Source Analysis
Alismail, Ahmad; Woldai, Betiel; Lanquillon, Carsten; Schacht, Sigurd (2026)
AI Transparency Conference (AITC) 2026.
Open Access Peer Reviewed
From Recognition to Refusal: Mapping Safety Formation Zones Across Model Architectures
Lanquillon, Carsten; Schacht, Sigurd (2026)
AI Transparency Conference (AITC) 2026.
Open Access Peer Reviewed
Rethinking Red Teaming: There Is More To Agents Than LLMs
Wacker, Thomas; Lanquillon, Carsten; Schacht, Sigurd (2026)
AI Transparency Conference (AITC) 2026.
Open Access Peer Reviewed
Architecture as Governance: Why LLM Agents Need Structural Compliance Enforcement
Höpfner, Steffen; Schacht, Sigurd (2026)
AI Transparency Conference (AITC) 2026.
Peer Reviewed
eDIF: A European Deep Inference Fabric for Remote Interpretability of LLM
Heithoff, Irma; Guggenberger, Marc; Kalogiannis, Sandra; Susanne, Mayer; Maag, Fabian; Schacht, Sigurd; Lanquillon, Carsten (2025)
Heithoff, Irma; Guggenberger, Marc; Kalogiannis, Sandra; Susanne, Mayer; Maag, Fabian...
arXiv, 2508.10553.
DOI: 10.48550/arXiv.2508.10553
Open Access
Mapping Moral Reasoning Circuits: A Mechanistic Analysis of Ethical Decision-Making in Large Language Models
Schacht, Sigurd; Lanquillon, Carsten (2025)
In: Degen, H., Ntoa, S. (eds) Artificial Intelligence in HCI. HCII 2025. Lecture Notes in Computer Science, Springer, Cham 15820, 97-116.
DOI: 10.1007/978-3-031-93415-5_6
Peer Reviewed
This paper systematically investigates how large language models (LLMs) encode moral reasoning across six moral dimensions: care, fairness, loyalty, authority, sanctity, and liberty. We propose a novel interpretability pipeline that combines differential activation analysis, automated neuron description, and ablation experiments to identify specialized neurons aligned with each moral dimension. Our curated dataset of 240 validated moral and immoral statement pairs guides this exploration and reveals that certain neurons consistently exhibit increased activation in response to morally aligned statements. Notably, the care and sanctity dimensions show the largest sets of specialized neurons, whereas fairness and loyalty show fewer. We further demonstrate that ablating these neurons can causally modulate ethical decision-making, supporting the presence of discrete sub-circuits that influence moral outputs. Our findings not only advance the theoretical understanding of moral reasoning in LLMs, but also highlight avenues for targeted interventions and alignment.
Mechanistic Exploration of the Architectural Impact of DPO Fine-Tuning on Ethical Alignment in LLMs
Maag, Fabian; Woldai, Betiel; Schacht, Sigurd (2025)
In: Degen, H., Ntoa, S. (eds) Artificial Intelligence in HCI. HCII 2025. Lecture Notes in Computer Science, Springer, Cham 15820.
DOI: 10.1007/978-3-031-93415-5_3
Peer Reviewed
Eine Zeitreise mit der MariusKI – Generative KI macht den Ansbacher Hofastronomen erlebbar
Leich, Pierre; Schacht, Sigurd; Woldai, Betiel (2025)
Regiomontanusbote 38 (4), 10-13.
Open Access
Deception in LLMs: Self-Preservation and Autonomous Goals in Large Language Models
Kamath Barkur, Sudarshan; Schacht, Sigurd; Scholl, Johannes (2025)
arXiv, 2501.16513.
DOI: 10.48550/arXiv.2501.16513
Open Access
Inference Optimizations for Large Language Models: Effects, Challenges, and Practical Considerations
Donisch, Leo; Schacht, Sigurd; Lanquillon, Carsten (2024)
Arxiv.
DOI: 10.48550/arXiv.2408.03130
Open Access
Evaluierung von LLMs und Autonomen Agenten
Kamath Barkur, Sudarshan; Sitapara, Pratik; Leuschner, Sven; Schacht, Sigurd (2024)
In: Gollisch, S., Gröner, P. (eds): Ansbacher Kaleidoskop 2024, Festschrift zum 60. Geburtstag von Prof. Dr. Ute Ambrosius und Prof. Dr. Barbara Hedderich, Shaker Verlag, Düren, 35 - 54.
KI-basierte Sprachmodelle in der Lehre: Question-Generation-Modelle zur Messung des Lernfortschrittes von Studierenden
Woldai, Betiel; Schacht, Sigurd; Kamath Barkur, Sudarshan (2024)
Neues Handbuch Hochschullehre - Sonderausgabe zur TURN23.
Open Access
Systematic Evaluation of Different Approaches on Embedding Search
Aperdannier, Roman; Köppel, Melanie; Unger, Tamina; Schacht, Sigurd; Kamath Barkur, Sudarshan (2024)
Aperdannier, Roman; Köppel, Melanie; Unger, Tamina; Schacht, Sigurd...
In: Arai, K. (eds) Advances in Information and Communication. FICC 2024. Lecture Notes in Networks and Systems, Springer, Cham 920, 526–536.
DOI: 10.1007/978-3-031-53963-3_36
Peer Reviewed
Magenta: Metrics and Evaluation Framework for Generative Agents based on LLMs
Kamath Barkur, Sudarshan; Schacht, Sigurd (2024)
AHFE International, Intelligent Human Systems Integration: Integrating People and Intelligent Systems 119, 144–153.
DOI: 10.54941/ahfe1004478
Open Access Peer Reviewed
PHOENIX: Open-Source Language Adaption for Direct Preference Optimization
Uhlig, Matthias; Schacht, Sigurd; Kamath Barkur, Sudarshan (2024)
Arxiv.
DOI: 10.48550/arXiv.2401.10580
Open Access
The Social Media Hate Speech Barometer: Making of
Sauer, Sebastian; Piazza, Alexander; Schacht, Sigurd (2023)
5th International Conference Business Meets Technology, Valencia, Spain, 143-162.
DOI: 10.4995/BMT2023.2023.16724
Open Access Peer Reviewed
Hate speech, particularly on social media channels, is a pressing cybersecurity concern and can even threaten the very foundations of societal stability. While there is a growing body of literature on how to detect and mitigate hate speech, applied researchers lack a state-of-the-art yet easily accessible infrastructure to build their own hate speech detection pipelines. We aim to provide an example of such an infrastructure that can serve as a template for other researchers. The infrastructure we present is based on the latest machine learning technologies available in the R environment: The Tidymodels framework and its extension Tidytext, plus the Targets project management approach, are the building blocks of our proposed infrastructure. In short, our data pipeline starts with downloading and preprocessing tweets, using various methods to convert text into numerical information. We then apply state-of-the-art supervised machine learning pipelines, drawing on a range of learning algorithms and incorporating new tuning capabilities. The focus of this paper is to explain the setup and rationale of the infrastructure. Our infrastructure is freely available on Github at https://github.com/sebastiansauer/hate-speech-barometer.
A Qualitative Evaluation of an AI-Based Study Progress Forecast
Kamath Barkur, Sudarshan; Fersch, Mascha-Lea; Henne, Sophie; Schacht, Sigurd; Woldai, Betiel (2023)
Kamath Barkur, Sudarshan; Fersch, Mascha-Lea; Henne, Sophie; Schacht, Sigurd...
Artificial Intelligence in Education Technologies: New Development and Innovative Practices. 190, 3-13.
DOI: 10.1007/978-981-99-7947-9_1
Peer Reviewed
This paper presents the development and evaluation of a first prototype of an ai-based study progress forecast. This service is integrated within a conversational agent and can be used by students to show them their current study progress. First, implications for the set-up of a forecast application from the literature are described. Based on the requirements identified in the literature and from the project itself, a lightweight formula was created that enables calculating the remaining study time. In order to assess preliminary feasibility and perception of the model prototype, a qualitative focus group discussion was conducted with five participants. Overall, the study progress forecast was well received by the participants, especially the offer itself as well as the promptness of the service were highlighted.
This is a preview of s
Ontology Definition for University Knowledge Graph
Woldai, Betiel; Kamath Barkur, Sudarshan; Henne, Sophie; Schacht, Sigurd; Schmid, Elena (2023)
Woldai, Betiel; Kamath Barkur, Sudarshan; Henne, Sophie; Schacht, Sigurd...
The Barcelona Conference on Education 2023: Official Conference Proceedings.
DOI: 10.22492/issn.2435-9467.2023.69
Open Access
Finding the required information to succeed in the organisation of everyday study life is not always easy for a student. Ontologies are an instrument to define a domain by illustrating its concepts and thereby presenting knowledge in a structured way. In this paper, our aim is to design an ontology that is suitable for the higher education environment of a German university to build a Knowledge Graph for a conversational AI. As a research context, the Ansbach University of Applied Science is used. The paper is organised into five sections. After a brief introduction in Section 1, Section 2 reviews previous work of conducted ontologies within the higher education environment, whereas Section 3 outlines the methodology for developing the ontology and presents the final result. The development procedure is thereby partly based on the ontology framework provided by Stanford University (Noy & McGuinness, 2001). The presented ontology, which delivers possible classes for the development, and transferability to other universities will then be discussed in Section 4. Finally, the conclusion and approaches for future work with ensuring a constant up-to-dateness of the classes are given in Section 5.
Der Einsatz von KI-basierten Sprachmodellen in der Lehre: Question-Generation Modelle als Instrument zur Messung des Lernfortschrittes von Studierenden
Woldai, Betiel; Schacht, Sigurd; Kamath Barkur, Sudarshan (2023)
Idee-Pitch-Workshop auf der TurnConference23 „Prototyp Zukunft – Lösungen für transformative Lehre teilen“, Köln.
Open Access Peer Reviewed
Knowledge Science - Grundlagen. Methoden der Künstlichen Intelligenz für die Wissensextraktion aus Texten
Lanquillon, Carsten; Schacht, Sigurd (2023)
Springer Vieweg Wiesbaden.
DOI: 10.1007/978-3-658-41689-8
Peer Reviewed
Knowledge Science beschäftigt sich mit Konzepten, Methoden und Prozessen zur systematischen Erzeugung, Extraktion, Speicherung und Bereitstellung von Wissen zur Lösung von Problemen und lässt sich somit dem Wissensmanagement zuordnen. Kognitive Assistenten sorgen dafür, das richtige Wissen zur richtigen Zeit in der richtigen Art und Weise seinen Anwendern und Anwenderinnen bereitzustellen. Damit dies gelingen kann, kommen inzwischen zahlreiche Methoden der Künstlichen Intelligenz (KI) zur Unterstützung unterschiedlicher Aufgaben des Wissensmanagements zum Einsatz.
Prof. Dr. Sigurd Schacht
Hochschule Ansbach
Fakultät Wirtschaft
Technologietransferzentrum Neustadt a.d. Aisch - Secure & Smart Data & Process Management
Residenzstr. 8
91522 Ansbach
sigurd.schacht[at]hs-ansbach.de
ORCID iD: 0000-0002-1161-4724