While the advent of the Fourth Industrial Revolution (4IR) technologies, like cloud computing, machine learning, and artificial intelligence have brought convenience and productivity improvements, they have also introduced new challenges in training and education that require the reskilling of existing employees and the building of a new workforce. Exacerbated by the already existing workforce shortages, this mammoth workforce reskilling and building effort aims to build a high-tech workforce capable of operating and maintaining these 4IR systems; requiring a higher student retention and persistence. This increase in student retention and persistence will be especially critical when training the workforce originating from marginalized communities like Underrepresented Minorities (URM), where challenges arise due to lack of access to high-quality education throughout the trainee's formative years (pre/middle/high schools), creating a cyclic set of knowledge dependencies that are difficult to meet. To address these challenges, this research presents Generative AI-based Personalized Tutor for Industrial 4.0 (gAI-PT4I4), a framework that focuses on personalization of 4IR experiential learning, using sentiment analysis to gauge student's knowledge comprehension, while using a combination of generative AI and finite automaton to personalize the content to the students' learning needs. The framework administers experiential learning, using low-fidelity Digital Twins that enable virtual reality-based (VR) training exercises focusing on 4IR training. The VR environment, integrates a generative AI teaching assistant called the Interactive Tutor, that guides the student through the training exercises, with audio and text communications.

Large language models (LLMs) exhibit impressive in-context learning (ICL) capability, enabling them to perform new tasks using only a few demonstrations in the prompt. Two different mechanisms have been proposed to explain ICL: induction heads that find and copy relevant tokens, and function vector (FV) heads whose activations compute a latent encoding of the ICL task. To better understand which of the two distinct mechanisms drives ICL, we study and compare induction heads and FV heads in 12 language models. Through detailed ablations, we discover that few-shot ICL performance depends primarily on FV heads, especially in larger models. In addition, we uncover that FV and induction heads are connected: many FV heads start as induction heads during training before transitioning to the FV mechanism. This leads us to speculate that induction facilitates learning the more complex FV mechanism that ultimately drives ICL.

This paper presents a novel perspective on human-computer interaction (HCI), framing it as a dynamic interplay between human and computational agents within a networked system. Going beyond traditional interface-based approaches, we emphasize the importance of coordination and communication among heterogeneous agents with different capabilities, roles, and goals. A key distinction is made between multi-agent systems (MAS) and Centaurian systems, which represent two different paradigms of human-AI collaboration. MAS maintain agent autonomy, with structured protocols enabling cooperation, while Centau-rian systems deeply integrate human and AI capabilities, creating unified decision-making entities. To formalize these interactions, we introduce a framework for communication spaces, structured into surface, observation, and computation layers, ensuring seamless integration between MAS and Centaurian architectures, where colored Petri nets effectively represent structured Cen-taurian systems and high-level reconfigurable networks address the dynamic nature of MAS. Our research has practical applications in autonomous robotics, human-in-the-loop decision making, and AI-driven cognitive architectures, and provides a foundation for next-generation hybrid intelligence systems that balance structured coordination with emergent behavior. Keywords: multi-agent systems centaurian systems communication spaces satellite and swarm robots large action models (LAMs). 1 Introduction Agentic AI systems--capable of iterative planning, autonomous task decomposition, and continuous learning--are rapidly reshaping the landscape of human-computer interaction (HCI). Recent advances in Large Language Models (LLMs) and advanced conversational agents have revitalized the field of multi-agent systems, whose roots in Artificial Intelligence predate the current rise of generative AI. Historically, multi-agent systems relied on agents with relatively constrained capabilities; however, the emergence of powerful, conversationally Corresponding author: uwe.borghoff@unibw.de

Due to the widespread use of LLMs and the rising critical ethical and safety concerns, LLM unlearning methods have been developed to remove harmful knowledge and undesirable capabilities. In this context, evaluations are mostly based on single-value metrics such as QA accuracy. However, these metrics often fail to capture the nuanced retention of harmful knowledge components, making it difficult to assess the true effectiveness of unlearning. To address this issue, we propose UNCD (UNlearning evaluation via Cognitive Diagnosis), a novel framework that leverages Cognitive Diagnosis Modeling for fine-grained evaluation of LLM unlearning. Our dedicated benchmark, UNCD-Cyber, provides a detailed assessment of the removal of dangerous capabilities. Moreover, we introduce UNCD-Agent, which refines unlearning by diagnosing knowledge remnants and generating targeted unlearning data. Extensive experiments across eight unlearning methods and two base models demonstrate that UNCD not only enhances evaluation but also effectively facilitates the removal of harmful LLM abilities.

We study the problem of reinforcement learning in infinite-horizon discounted linear Markov decision processes (MDPs), and propose the first computationally efficient algorithm achieving near-optimal regret guarantees in this setting. Our main idea is to combine two classic techniques for optimistic exploration: additive exploration bonuses applied to the reward function, and artificial transitions made to an absorbing state with maximal return. We show that, combined with a regularized approximate dynamic-programming scheme, the resulting algorithm achieves a regret of order $\tilde{\mathcal{O}} (\sqrt{d^3 (1 - \gamma)^{- 7 / 2} T})$, where $T$ is the total number of sample transitions, $\gamma \in (0,1)$ is the discount factor, and $d$ is the feature dimensionality. The results continue to hold against adversarial reward sequences, enabling application of our method to the problem of imitation learning in linear MDPs, where we achieve state-of-the-art results.

Estimating the construction year of buildings is of great importance for sustainability. Sustainable buildings minimize energy consumption and are a key part of responsible and sustainable urban planning and development to effectively combat climate change. By using Artificial Intelligence (AI) and recently proposed Transformer models, we are able to estimate the construction epoch of buildings from a multi-modal dataset. In this paper, we introduce a new benchmark multi-modal dataset, i.e. the Map your City Dataset (MyCD), containing top-view Very High Resolution (VHR) images, Earth Observation (EO) multi-spectral data from the Copernicus Sentinel-2 satellite constellation, and street-view images in many different cities in Europe, co-localized with respect to the building under study and labelled with the construction epoch. We assess EO generalization performance on new/ previously unseen cities that have been held-out from training and appear only during inference. In this work, we present the community-based data challenge we organized based on MyCD. The ESA AI4EO Challenge MapYourCity was opened in 2024 for 4 months. Here, we present the Top-4 performing models, and the main evaluation results. During inference, the performance of the models using both all three input modalities and only the two top-view modalities, i.e. without the street-view images, is examined. The evaluation results show that the models are effective and can achieve good performance on this difficult real-world task of estimating the age of buildings, even on previously unseen cities, as well as even using only the two top-view modalities (i.e. VHR and Sentinel-2) during inference.

Large Vision-Language Models (LVLMs) have recently gained attention due to their distinctive performance and broad applicability. While it has been previously shown that their efficacy in usage scenarios involving non-Western contexts falls short, existing studies are limited in scope, covering just a narrow range of cultures, focusing exclusively on a small number of cultural aspects, or evaluating a limited selection of models on a single task only. Towards globally inclusive LVLM research, we introduce GIMMICK, an extensive multimodal benchmark designed to assess a broad spectrum of cultural knowledge across 144 countries representing six global macro-regions. GIMMICK comprises six tasks built upon three new datasets that span 728 unique cultural events or facets on which we evaluated 20 LVLMs and 11 LLMs, including five proprietary and 26 open-weight models of all sizes. We systematically examine (1) regional cultural biases, (2) the influence of model size, (3) input modalities, and (4) external cues. Our analyses reveal strong biases toward Western cultures across models and tasks and highlight strong correlations between model size and performance, as well as the effectiveness of multimodal input and external geographic cues. We further find that models have more knowledge of tangible than intangible aspects (e.g., food vs. rituals) and that they excel in recognizing broad cultural origins but struggle with a more nuanced understanding.

As large language models (LLMs) converge towards similar capabilities, the key to advancing their performance lies in identifying and incorporating valuable new information sources. However, evaluating which text collections are worth the substantial investment required for digitization, preprocessing, and integration into LLM systems remains a significant challenge. We present a novel approach to this challenge: an automated pipeline that evaluates the potential information gain from text collections without requiring model training or fine-tuning. Our method generates multiple choice questions (MCQs) from texts and measures an LLM's performance both with and without access to the source material. The performance gap between these conditions serves as a proxy for the collection's information potential. We validate our approach using three strategically selected datasets: EPFL PhD manuscripts (likely containing novel specialized knowledge), Wikipedia articles (presumably part of training data), and a synthetic baseline dataset. Our results demonstrate that this method effectively identifies collections containing valuable novel information, providing a practical tool for prioritizing data acquisition and integration efforts.

We present PeerQA, a real-world, scientific, document-level Question Answering (QA) dataset. PeerQA questions have been sourced from peer reviews, which contain questions that reviewers raised while thoroughly examining the scientific article. Answers have been annotated by the original authors of each paper. The dataset contains 579 QA pairs from 208 academic articles, with a majority from ML and NLP, as well as a subset of other scientific communities like Geoscience and Public Health. PeerQA supports three critical tasks for developing practical QA systems: Evidence retrieval, unanswerable question classification, and answer generation. We provide a detailed analysis of the collected dataset and conduct experiments establishing baseline systems for all three tasks. Our experiments and analyses reveal the need for decontextualization in document-level retrieval, where we find that even simple decontextualization approaches consistently improve retrieval performance across architectures. On answer generation, PeerQA serves as a challenging benchmark for long-context modeling, as the papers have an average size of 12k tokens. Our code and data is available at https://github.com/UKPLab/peerqa.

Large Language Models (LLMs) enhance their problem-solving capability by leveraging both parametric and external knowledge. Beyond leveraging external knowledge to improve response accuracy, they require key capabilities for reliable knowledge-handling: resolving conflicts between knowledge sources, avoiding distraction from uninformative external knowledge, and abstaining when sufficient knowledge is unavailable. Prior studies have examined these scenarios in isolation or with limited scope. To systematically evaluate these capabilities, we introduce a comprehensive framework for analyzing knowledge-handling based on two key dimensions: the presence of parametric knowledge and the informativeness of external knowledge. Through analysis, we identify biases in knowledge utilization and examine how the ability to handle one scenario impacts performance in others. Furthermore, we demonstrate that training on data constructed based on the knowledge-handling scenarios improves LLMs' reliability in integrating and utilizing knowledge.