EXC 3057:  Reasonable Artificial Intelligence

• Subject Area: Computer Science
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 533677015

Over the past decade, deep learning (DL) has driven groundbreaking advances in artificial intelligence (AI). However, current DL-based AI systems are unreasonable in many ways: (1) They are developed and deployed in unreasonable ways, requiring overly large models, vast amounts of data and computational power, and extensive infrastructure. This has led to a monopoly held by a few large companies with the necessary resources. (2) They struggle with reasoning, handling unfamiliar situations, and nuanced context. They lack commonsense understanding and abstraction capabilities. (3) They do not improve continuously, learn through interactions, or adapt quickly. Instead, they require frequent retraining, resulting in high eco- nomic and environmental costs. Such unreasonable learning makes them brittle. Given these challenges, it is not surprising that an AI that can beat the Go world champion can also be easily tricked into defeat, or when ChatGPT, despite being able to write a long coherent text on complex topics, can still exhibit a lack of simple common sense. To mitigate the potential harms of such unreasonable AI, it is crucial to address these fundamental issues rather than relying on brute-force scaling of resources or temporary fixes. To this end, the proposed Cluster of Excellence introduces a new generation of AI, termed Reasonable Artificial Intelligence (RAI). RAI aims to develop AI that learns in more reasonable ways. This will involve models being trained in a modular fashion, allowing them to continuously improve and build abstract world knowledge. They will have an inherent ability to reason, interact, and adapt to their surroundings. RAI will serve as a collaborative platform, bringing together researchers from different disciplines. It is structured into four research labs, each dedicated to an essential research topic: (1) The Systemic AI (SAI) Lab focuses on novel software and systems methods to create sophisticated AI models composed of reusable, adaptive, and interchangeable building blocks, ensuring efficient integration into existing large-scale software systems. (2) The Observational (OAI) AI Lab rethinks AI algorithms from the ground up and lays new algorithmic foundations for a modular design on all levels: input (modalities), output (tasks), knowledge, and reasoning (fusion and routing of information). (3) The main goal of the Active AI (AAI) Lab is to equip AI systems with lifelong adaptation abilities and skills to enable them to interact with the world purposefully and deal with unknown situations. (4) The Challenging AI with Cognitive Science (CAI) Lab develops challenges and benchmarks to monitor and evaluate the progress of RAI. PIs from different labs will work in multidisciplinary teams, crossing traditional boundaries and fostering diversity. This approach will create a new learning-centered computing paradigm that will revolutionize AI development and application.

• DFG Programme: Clusters of Excellence (ExStra)
• Applicant Institution: Technische Universität Darmstadt
• Participating Institution: Eberhard Karls Universität Tübingen; Goethe-Universität Frankfurt am Main; Julius-Maximilians-Universität Würzburg; Universität Bremen; Universität des Saarlandes
• Spokespersons: Professor Dr. Kristian Kersting; Professorin Dr.-Ing. Mira Mezini; Professor Dr. Marcus Rohrbach
• Participating Researchers: Professor Dr. Carsten Binnig; Professorin Georgia Chalvatzaki, Ph.D.; Professor Carlo D' Eramo, Ph.D.; Professor Dr. Jan Gugenheimer; Professorin Dr. Iryna Gurevych; Dr. Mohammad Emtiyaz Khan; Professor Dr. Heinz Koeppl; Professorin Dr.-Ing. Hilde Kuehne; Professor Dr. Martin Mundt; Professor Jan Reinhard Peters, Ph.D.; Professorin Dr. Anna Rohrbach; Professorin Dr. Gemma Roig; Professor Stefan Roth, Ph.D.; Professor Dr. Constantin Rothkopf; Dr. Simone Schaub-Meyer; Professor Dr.-Ing. Justus Thies; Professorin Dr. Isabel Valera; Professorin Dr. Angela Yu