This paper introduces a robust pipeline enabling Large Language Models (LLMs) to perform precise Table Question Answering (Table-QA) on complex statistical tables. Processing complex, visually oriented table formats remains a major challenge for large language models, as they exhibit inherent multi-dimensional and hierarchical structures. Despite their visual complexity, these formats adhere to strict structural conventions. Our approach leverages these regularities—such as specific cell border styles and hierarchical indentations—to transform these files into a unified JSON structure using a deterministic extraction method. Instead of generating direct answers, the LLM is fine-tuned to produce Python scripts that programmatically extract information from the generated JSON files. To achieve this, we generated a synthetic dataset comprising approximately 13,000 high-quality question-code pairs. Our fine-tuned Microsoft Phi-4 model achieves a functional success score of 88.5
The Ride-Hailing Problem is an online optimization challenge that coordinates a fleet of vehicles to serve requests revealed over time. The problem is NP-hard, and many solution approaches use stochastic or machine-learning based algorithms, which perform well but are difficult to interpret. We study the trade-offs between performance and explainability. Our results show that greedy assignment consistently outperforms more complex methods, indicating that explainability need not come at the cost of efficiency. We also find that the effectiveness of repositioning strategies strongly depends on demand patterns, highlighting the value of context-aware policies.
Professional ethics boards in fields such as engineering must identify applicable obligations when multiple provisions of the same ethical code generate competing requirements. In prior work we defined a nine-component case representation for professional ethics analysis. This paper presents the ontological implementation of that representation, grounded in the Basic Formal Ontology (BFO) and classified through the Information Artifact Ontology (IAO). Type-level disjointness is verified by the Pellet OWL-DL reasoner. Three object properties (competesWith, prevailsOver, defeasibleUnder) expose the competition structure between obligations as asserted, SPARQL-queryable relationships between typed individuals. OntServe, an ontology server, instantiates the representation as a knowledge graph of case ontologies extracted from a professional board of ethics archive.
This paper introduces Warehouse Logistics 2D (WaLo2D), an open-source environment designed for testing multi-agent reinforcement learning algorithms in automated warehouse logistics. WaLo2D is a configurable, low-compute simulation that models simple 2D warehouse tasks. This enables researchers and developers to quickly evaluate cooperative strategies for package storage, pick-up, and redistribution centre management. While more realistic 3D simulations are essential for final deployment, they often slow down early policy training due to the time-consuming manual crafting of details and high hardware requirements. In contrast, overly simplified models, such as discrete grid-based navigation, may not generalise well. WaLo2D bridges this gap by providing an accessible testbed that accelerates the initial evaluation of algorithms and tuning of reward functions, which is a critical advantage given the high number of repetitions required in multi-agent reinforcement learning experiments. It supports logistics scenarios, such as package routing, storage management and truck loading. The project is available at https://github.com/KIForAN/WaLo2D .
Linear recurrent neural networks (LRNNs) provide a simple yet powerful framework for modeling time series. Unlike traditional recurrent neural networks, LRNNs use linear activation functions and can be trained without backpropagation by solving linear equation systems. This paper summarizes their theoretical properties, learning procedure, and ability to approximate arbitrary time-dependent functions. A key contribution is a method for reducing network size using eigenvalue analysis, enabling both efficient learning and architecture optimization, which is nowadays an important topic. We demonstrate the properties and advantages of LRNNs by several case studies, among them multiple superimposed oscillators (MSO), robotic soccer (RoboCup), and stock price prediction. LRNNs outperform the previous state-of-the-art for the MSO task with a minimal number of units.
Effective planning and scheduling of industrial processes are vital across numerous industries, including the daily operations of coal mining logistics. The Stockyard Planning Problem (SPP) aims to address these tasks in stockyards efficiently and cost-effectively. This research examines a realistic scenario involving an opencast mine linked to a stockyard, where bulk materials of varying quality are stored on stockpiles and transported by train. We approach the SPP in this setting using constraint programming and develop a model that accounts for the time-dependent nature of the problem: materials can be stored on stockpiles only for a limited duration before export, and transitions between machine operating states introduce blocking periods. The primary objective of the SPP in this scenario is to facilitate seamless operations of both opencast mining and stockyard activities. It is crucial to ensure continuous material flow over conveyor belts to prevent stoppages in mining operations while meeting export requirements.
Within the last years, generative AI usage rose dramatically. Using such tools is problematic in areas like academia where integrity matters most. Hence ways must be found to reliably determine authorship. Our system addresses this issue by involving four classifiers and gradient boosting. Its classification pipeline processes texts sentence-wise and applies two tests afterwards: A Gauss test evaluates whether the proportion of flagged sentences is within an expectation, the second test checks for consecutive flagged sentences. Here the false positive rate can be controlled with custom cutoffs. This leads to false positive rates of less than 1
Passive Acoustic Monitoring (PAM) is increasingly used for biodiversity monitoring, but training sound event detectors for PAM remains difficult because strong temporal annotations are expensive to obtain. Archival sound libraries provide weakly labelled focal recordings at scale, yet these recordings differ substantially from PAM soundscapes. Focal recordings typically contain one dominant species under relatively clean conditions, whereas PAM recordings often contain overlapping vocalisations from multiple species together with environmental noise. This mismatch limits the transferability of weakly supervised models trained on sound libraries. This paper investigates whether synthetic multi-species training augmentation can reduce this transfer gap. We build on a weakly supervised sound event detection pipeline using BirdNET embeddings and a linear classifier, and augment the training data by synthetically mixing focal recordings into artificial multi-species soundscapes with randomized temporal overlap and mixture weights. The PAM evaluation data remain unchanged. Experiments on five anuran species show improvements in transfer performance. Compared with training on original focal recordings only, synthetic augmentation increases recall and F1-score at both bag and segment level, while causing a moderate reduction in precision. The best bag-level micro-F1 improves from 0.6792 to 0.7687, and the best segment-level micro-F1 improves from 0.5473 to 0.6120. These results indicate that simulating multi-species co-occurrence during training is a practical way to improve weakly supervised transfer from sound libraries to PAM data.
Insertion heuristics such as Nearest Insertion, Farthest Insertion, and Max-Regret are well-established constructive methods for the Traveling Salesman Problem (TSP). However, they are typically applied as standalone algorithms and are not structurally embedded within exact optimization models. In this paper, we propose a novel approach that transfers the decision principles of insertion heuristics directly into a constraint programming (CP) model. Starting from the classical Miller-Tucker-Zemlin (MTZ) formulation, we extend the model by introducing additional positional variables, a dedicated value selection strategy, and a custom constraint equipped with a specialized propagator linking insertion positions to edge-selection variables. This integration enables the systematic emulation of insertion heuristics within the search process of the CP solver. We instantiate the framework using four heuristics: Nearest Insertion, Farthest Insertion, Farthest-Nearest Insertion, and Max-Regret Insertion. An experimental evaluation on 64 benchmark instances with up to 500 nodes demonstrates that all proposed variants significantly outperform the classical MTZ model and a Savings-based adaptation. The results highlight the potential of structurally embedding heuristic knowledge into constraint models to improve robustness and scalability.
Imitation learning from human demonstration often relies on marker-based motion capture, which restricts data recording to controlled environments. Video-based motion capture presents an alternative that reduces the recording effort. To compare both, we develop a novel evaluation framework using four categories of metrics, namely pose estimation quality, trajectory distance, behaviour separability, and grasp characterization, applied to three state-of-the-art video-based 3D pose estimation models in a pick-and-place scenario with parallel recording. Using this framework, we show that video-based pose estimation using a depth map captures actions and trajectories well enough for imitation learning, but high-fidelity movement such as grasping remains beyond the current capabilities of all evaluated video-based modalities.
Neural Networks (NN) are recognized as universal approximators, yet the reasons behind their remarkable generalization capabilities remain partially unresolved. One potential explanation is the manifold hypothesis, which suggests that NNs identify low-dimensional manifolds within the high-dimensional real-world data. In each layer of a NN, the data lies on/near a manifold, a so-called representation manifold, which gets progressively flatter as the layer’s depth increases. Since NNs function on these data manifolds rather than isolated points, generalization becomes feasible. Notably, manifolds in the deeper layers of trained NNs tend to be flatter, while at the same time, deeper networks generally achieve higher accuracy. This leads to the following hypothesis: There is a correlation between the accuracy of a NN and the flatness of manifolds in the final layer of the NN. To investigate this, we empirically test the hypothesis by training a CNN with various hyperparameter sets. We assess both the flatness of the manifold and the accuracy on a test set, exploring their connection. We use correlation analysis, plots, statistical tests and examine whether flatness can be predicted from the hyperparameters.
Artificial neural networks are one of the key technologies in machine learning. In particular, networks using the ReLU-activation function are widely used in the community of formal verification and theoretical machine learning. Such networks can be represented by their linear decomposition, which consists of a set of convex polytopes together with associated linear functions that capture the network’s behavior. This representation provides detailed insights into the learned structure of the model and enables the quantitative analysis of both linear and non-linear verification problems. However, obtaining the linear decomposition poses a computational challenge. To the best of our knowledge, only two algorithms have been proposed for this task, yet a systematic and objective comparison between them is still missing. In this work, we implemented both approaches and conducted a comprehensive empirical evaluation of their runtime across a range of benchmark settings. Our results show that while one algorithm outperforms the other in many scenarios, both exhibit distinct strengths and weaknesses. Consequently, no single algorithm emerges as the clear winner, highlighting that performance is fundamentally shaped by the network’s structural properties.
Recent advances have introduced speech-to-speech (S2S) conversational assistants capable of producing natural-sounding interactions, including non-verbal cues like tonality and mood. In the automotive domain, this enables intuitive and humanlike in-car dialogue experiences. However, integrating these end-to-end assistants limits architectural options for programmable domain-specific safeguards. This paper discusses two implementation approaches for S2S guardrails: transcript-based and tool-based. Through an empirical evaluation, we demonstrate that both strategies are insufficient for industrial deployment in most cases due to prohibitive latency (delaying each answer by 0 to 1.4 s even for computationally cheap checks) and technical impediments (like potentially non-deterministic tool call behavior). Finally, we outline open challenges for S2S guardrails in the automotive context.
Industrial acoustic monitoring in real manufacturing environments is challenging due to strong background noise, heterogeneous operating conditions, and limited labeled fault data. In such settings, the normal operating conditions often vary acoustically. Background sounds may resemble malfunction events. All of these factors may lead to unstable decision boundaries and increased false alarm rates. To address this challenge, we formulate industrial audio anomaly detection as a multi-state problem rather than a binary normal vs. anomalous task. We propose to adapt the PatchCore retrieval framework for industrial audio, using log-Mel spectrogram representations and class-specific memory banks to model different operating states. The approach is evaluated under multiple temporal segmentation strategies using a pretrained CNN14 encoder on AudioSet, and compared with a supervised CNN baseline. Our experiments show that fault state detection performance depends strongly on the alignment between temporal segmentation and the encoder receptive field, rather than on the temporal context length alone. When this alignment is preserved, the retrieval-based approach produces well-separated embeddings of operating states and achieves strong ranking performance while maintaining competitive classification accuracy.
AI planning is a central problem in Knowledge Representation (KR), as it requires expressive action theories, structured domain models, and formal reasoning mechanisms. With the increase of domain complexity, manually building planning models becomes a tedious and error-prone task, highlighting the need for methods that can learn symbolic knowledge from unstructured data while preserving formal semantics. Recent advances in language processing, especially large language models (LLMs), promise great potential for model extraction from text. They, however, are prone to hallucination and lack the ability to validate the extracted knowledge required for formal planning and reasoning. To address the above challenges, we introduce a novel methodology for automatic extraction of formal models expressed in the Planning Domain Definition Language (PDDL) from textual instructions. Our methodology combines linguistic, statistical, symbolic and LLM-based techniques within a knowledge-driven framework. In doing so, it enables the learning of symbolic abstractions such as action schemas, type hierarchies, preconditions, and effects, while also providing an explicit mechanism for explainability by making the derivation of the final formal representation transparent and traceable. As an intermediate representation, we construct a language-independent knowledge graph, called a Situation Model (SM), which captures relational dependencies and domain constraints in a structured and reusable form. We evaluate the proposed hybrid approach by investigating syntactic and semantic correctness via AI planning, as well as validation against manually created plans from everyday-life domains, demonstrating the benefits of integrating KR techniques with LLM-based extraction.
The integration of Artificial Intelligence (AI) techniques in e-learning ecosystems is expanding, although the algorithms used and decision steps often represent an enigma to learners and educators. Equity issues, plausible privacy risks, scarcity of transparency and trust are major concerns for the adoption of AI in digital education, about individualized teaching and learning recommendations, feedback, and e-assessment. The aim of this work-in-progress is to explore the importance of integrating explainable AI (xAI) methods in Digital Ecosystems in Research and Education (DERE) for Education 4.0 (E4.0) and present an approach for structuring the main stakeholders and xAI use scenarios constructed following a design-science approach.
Multi-agent path finding (MAPF) is the problem of finding collision-free paths for a team of agents. Conflict-based search (CBS) is a state-of-the-art algorithm for solving MAPF optimally that repeatedly picks a collision (known as conflict) between two agents and resolves it by constraining the movement of the agents. One of the earliest runtime enhancement of CBS categorized conflicts into three groups (cardinal, semi-cardinal, and non-cardinal) and resolved them in that priority order. When multiple collisions share the same type, CBS typically breaks ties by selecting one at random. In this work, we show that CBS with disjoint splitting (which is another CBS’ runtime enhancement) is further improved by introducing a more nuanced conflict prioritization strategy when multiple cardinal (or semi-cardinal) conflicts are present. More specifically, we prioritize conflicts that involve agents that participate in multiple cardinal (or semi-cardinal) conflicts. Our experiments show that the new conflict prioritization strategy reduces both the number of node expansions and the runtime of CBS with disjoint splitting for many instances from the standard MAPF benchmark suite.
Large language models are increasingly used for knowledge graph question answering (KGQA), but can fail to correctly ground answers in the underlying graph. Current approaches to LLM-based KGQA either rely on full semantic parsing into executable queries such as SPARQL, which is brittle in practice due to complex schemas or incompleteness of real-world KGs, or on LLM-reasoning and answer generation over KGs, which can be more robust but lacks formal guarantees. In this work, we study a complementary setting in which candidate answers are generated by an LLM-based system and subsequently verified using lightweight symbolic constraints derived from the question. We introduce Constrained Entity Selection under Partial Knowledge (CES-PK), a problem formulation that focuses on eliminating invalid answers and providing symbolic support for valid ones without requiring construction of executable logical forms. To account for incomplete KGs, we employ a three-valued constraint semantics (satisfied, violated, unknown) that avoids incorrect rejections under open-world assumptions. To demonstrate the effects of our method, we instantiate CES-PK over the Hetionet biomedical knowledge graph and evaluate both controlled and LLM-generated candidate sets. Experiments show that constraint-based verification improves precision while preserving recall and provides additional symbolic evidence for remaining candidates.
This article proposes a methodology for evaluating AI systems as components of social mechanisms rather than isolated artefacts. The research question we answer in the paper is: How can an AI system’s contribution to a human–agent mechanism be assessed with respect to collective outcomes, welfare, norms, and risks? Combining mechanism design with analytical sociology, we propose the Star–Durkheim test as a diagnostic instrument for asking which collective goal is implemented, how welfare and fairness are affected, and which causal social dynamics may emerge. The test does not ask whether an AI appears human as done in the well-known Turing test, but investigates whether it contributes to, distorts, or endangers a socially situated mechanism. NotebookLM, Wikipedia, and autonomous vehicles illustrate the method; LLM prompting is treated as prototyping a Star-Durkheim test, not as full test automation.
Large language models fail systematically on modal and conditional reasoning tasks, yet whether such failures are foreseeable remains unexplored. Selective prediction, the ability to abstain when a prediction is likely wrong, has been studied for factual NLP tasks but not for structured logical reasoning. We investigate whether LLM failures on seven modal and conditional inference patterns are predictable from the structural properties of reasoning problems, without relying on model outputs. Using 3,776 inference instances across nine models and three prompting strategies, we train an external classifier to predict per-instance failure from input-side features. Leave-one-pattern-out cross-validation reveals substantial variation in predictability: AUCROC ranges from 0.69 to 0.93, with patterns involving “must” operators being most predictable. We further show that targeted prompting (LogiCue), while reducing overall failures substantially, makes remaining failures structurally harder to predict. These findings suggest that input-side and output-side reliability mechanisms serve complementary roles in LLM reasoning pipelines.
