Keynote Lectures

LLMs for Automated Algorithm Design
Lars Kotthoff, University of St Andrews, United Kingdom

AI in Industry 5.0 : Towards Human-Centric Cyber-Physical-Cognitive Systems
Hervé Panetto, University of Lorraine, France

Advancing Explainable AI for Mental Health Monitoring
Gabriella Casalino, Universita degli Studi di Bari, Italy

Abstract
Automatically designing high-performing algorithms in diverse problem domains has been an active research area for a long time. Instead of laboriously designing algorithms manually, the vision is to efficiently and effectively create new algorithms that are tailored to a particular application, leverage the structure of a problem better than existing algorithms, or work under more restrictive constraints. The involvement of human experts is reduced to providing guidance, rather than doing all the heavy lifting. In this keynote, I will argue that LLMs are the missing piece to achieve this vision. They are surprisingly good at generating code, and combined with established and new techniques in automated algorithm design, have demonstrated that they can beat state-of-the-art expert-designed algorithms in several application domains. While LLMs are part of the solution, challenges still remain and interesting research questions are waiting to be tackled.

Abstract
While the shift toward Industry 5.0 introduces a human-centric philosophy to production, the technical evolution of industrial environments now hinges on the integration of cognition within AI-driven frameworks. Modern architectures are transitioning from standard Cyber-Physical Systems (CPS) toward Cyber-Physical-Cognitive systems. This evolution addresses a critical gap in existing models: the neglect of safety, trustworthiness, and human-centric reasoning. By leveraging Hybrid Artificial Intelligence—specifically through Active Learning, Forecasting, and Explainable AI (XAI)—these systems move beyond simple automation to engage in proactive decision-making. Through the use of simulated realities and real-time user feedback, these cognitive layers foster a genuine synergy between humans and machines. Ultimately, this abstract argues that the future of Industrial and Systems Engineering lies in embedding deep cognitive capabilities into AI, ensuring technology serves as a collaborative, transparent partner in sustainable production.

Abstract
Mental health conditions are complex and heterogeneous, often manifesting through subtle behavioral and physiological changes, including variations in speech patterns. Clinicians routinely exploit these signals during assessments, while recent advances in mobile and wearable technologies now enable continuous and unobtrusive data collection in everyday life. Despite this growing availability of data, AI-based approaches for mental health monitoring are still limited in clinical adoption. Key challenges include the lack of standardized machine learning pipelines, transparent evaluation strategies, and, most importantly, interpretable explanations that clarify how computational models relate to clinically meaningful constructs. In this talk, I will present my research on explainable AI methods for mental health monitoring, focusing on interpretable and human-centered models based on fuzzy logic and explainable acoustic biomarkers. The bipolar disorder domain will be discussed as a representative case study. The overarching goal is to contribute to narrowing the gap between data-driven models and clinical reasoning, supporting the development of AI-assisted tools that are intended to be more transparent, usable, and better aligned with psychiatric practice.