Postersession im Foyer
17 Forschungsposter - Aktuelle Entwicklungen in ML & KI
Entdecken Sie innovative Forschungsarbeiten und praktische Anwendungen
The CRISTAL Method: Fast, reliable analytical problem-solving with pre-synthesized grounded world models
Felix Neubürger
We introduce CRISTAL (Coherent Reliable Intentional Synthesis of Truthful Analysis Logic), a neurosymbolic framework for automating complex analysis workflows.Investment analysis is challenging: high uncertainty, noisy/subjective data, limited attention, and the need for justified, reproducible decisions. Human analysts face cognitive biases, while LLM-based agents struggle with poor numerical reasoning, lack of uncertainty awareness, and weak reproducibility.
CRISTAL bridges these gaps by combining:- Statistical model synthesis with full Bayesian inference
- Active, budget-aware data acquisition
- LLM-assisted code generation for continuous refinement
From a natural-language knowledge base, CRISTAL builds a dynamic, interpretable probabilistic program that evolves during analysis.
On a synthetic equities benchmark, CRISTAL achieves Bayes-optimal accuracy with only 5 examples and a 5-second budget, surpassing state-of-the-art LLMs that plateau at ~40% despite far greater data and compute.
Bias Unveiled: Enhancing Fairness in German Word Embeddings with Large Language Models
Yasser Saeid
Gender bias in word embeddings can reinforce harmful stereo-types in machine learning systems. This research addresses gender bias in German word embeddings, proposing mitigation using large language models (LLMs) to promote fairness and inclusivity. Bias in language models impacts applications like hiring, translation, and customer service. Addressing biases at the embedding level is critical, as embeddings underpin many NLP tasks.
An Open-Source Agentic Reasoning Framework for German Nuclear Decommissioning Documents
Rohil Rao
We present an agentic framework that performs iterative reasoning over German-language documents in the nuclear decommissioning domain. The system orchestrates specialized language models for answering complex questions that require multi-step reasoning. To assess performance, we created a dataset: the German-Nuclear-Decom-QA, a domain-specific evaluation, comprising highly objective questions with varying difficulty. In our evaluation, the workflow consistently gathers document-grounded evidence, and generates reasoned, accurate answers. These results indicate that agentic workflows can support reliable question answering in safety-critical settings without overreliance on a single model.
KIKO: A GenAI Platform for Adaptive, Collaborative Learning
Sanjay Gupta
We present KIKO, a research-driven, open-source platform that explores how Generative Artificial Intelligence (GenAI) can support adaptive, collaborative learning. Rather than solely focusing on technical innovation, KIKO serves as a testbed for evaluating how collective intelligence can be leveraged when humans are guided by GenAI. The platform will include multiple assistant personas to support both novices and experts during their onboarding and learning journey. We investigate how GenAI identifies learner errors, adapts feedback to address cognitive gaps, and dynamically adjusts learning paths to maximize skill development. Through these experiments, KIKO contributes to understanding the design of user-tailored, trustworthy AI companions in education.
Data Engineering for LLMs
Amir Kazemzadeh Moghanjoughi
LLM-based Knowledge Graph Construction
Ole Katzenberger
Large Language Models (LLMs) are increasingly used for Knowledge Graph (KG) construction, but their outputs often lack semantic consistency and structural coherence. Ontologies can address this by constraining extraction and enabling validation, reasoning, and integration. This study evaluates Neo4j, LangChain, and a custom approach for ontology-guided KG construction from German texts on nuclear decommissioning. A modular pipeline was developed to process inputs of varying complexity, using custom ontologies and ground truths for evaluation with precision, recall, and F1. The pipeline provides a reusable, systematic way to benchmark and improve LLM-based extraction strategies.
GraphRAG for regulatory data
Christian Vahrenkamp
The decommissioning of nuclear facilities involves intricate legal, procedural, and safety-related documentation, making effective knowledge retrieval essential for regulatory compliance. This thesis explores the application of graph-based RetrievalAugmented Generation (GraphRAG) as a means to enhance information access in this complex and sensitive domain. Traditional RAG methods, which rely on vector-based retrieval, often fail to capture inter-document relationships and implicit contextual dependencies that are critical in regulatory texts. To address these limitations, five retrieval pipelines were developed and evaluated: three based on standard RAG architectures and two employing novel GraphRAG designs with distinct graph structures and retrieval strategies.
Experiments were conducted on both a general-purpose Wikipedia dataset and a
domain-specific corpus of legal and procedural texts related to nuclear decommissioning. All evaluations were performed in a fully local, privacy-preserving environment using open-source models. The results indicate that graph-enhanced retrieval methods significantly outperform traditional approaches in terms of contextual fidelity, answer relevance and interpretability - particularly in scenarios requiring broad semantic understanding. However, challenges persist in handling narrowly scoped, detail-oriented queries.
The findings underscore the potential of GraphRAG to support knowledge retention and decision-making in high-stakes, knowledge-restricted environments. This work contributes a systematic comparison of retrieval strategies and introduces design principles for constructing context-aware, compliant AI systems tailored to regulated domains.
Reliable AI for Safety-Critical Fact-Checking with LLMs, RAG, and Uncertainty Estimation
Daniel Gierse
The increasing popularity of Large Language Models (LLMs) in Natural Language Processing has led to significant advancements, but ensuring the reliability and trustworthiness of their outputs remains a critical challenge. This work is set in the domain of nuclear decommissioning and addresses the challenge by proposing an agent architecture that integrates LLMs for knowledge retrieval and generation as well as probabilistic, logit- and similarity-based methods for uncertainty estimation. The agent utilizes Retrieval-Augmented Generation (RAG) to access and process relevant information from a knowledge base, enabling it to look up relevant data to identify inconsistencies.
Projekt Urban Climate Twin
Ein Projekt in Kooperation mit der Stadt Soest
Serious Games meets KI: Programmierenlernen adaptiv gedacht
Dehbia Kouadria
Transfer Learning in Multi-Agent Systems: Evolutionary Algorithms and Dynamic Grid Structures for Industrial Applications
Marlon Löppenberg
Distributed production systems for process measurement and control are increasingly required to reconcile economic objectives, such as energy efficiency and productivity, with critical technical demands, including flexibility, real-time capability, and reliability. This paper proposes a novel methodology that employs dynamic grid structures for state space representation, embedded within a centralized control framework for multi-agent systems in the context of multi-objective industrial automation. The approach integrates action-oriented decision-making with evolutionary mechanisms by selection, recombination, and mutation, to enable adaptive optimization. Operating within an agent-based interaction cycle, these structures facilitate continuous heterogeneous learning. In addition, the methodology supports heterogeneous knowledge transfer among agents through graph-theoretic data structures, ensuring the efficient dissemination of specialized expertise. The developed strategies are systematically validated in an industrial laboratory environment, addressing the challenges associated with multi-objective optimization in complex, highly interconnected manufacturing processes. The results demonstrate significant advancements in evolutionary knowledge transfer, notable improvements in production efficiency through reduced energy consumption, and the practical applicability of dynamic network structures in real-world industrial settings.
Resampling Multi-Resolution Signals Using the Bag of Functions Framework
David Salazar
This paper introduces the Multi-Resolution Bag of Functions (MR-BoF) framework, a novel approach for analyzing time series data with varying sampling rates. Unlike traditional methods requiring uniform sampling, MR-BoF uses a flexible, sampling-rate-independent technique to decompose time series signals of different resolutions. Our experiments demonstrate that this framework accurately reconstructs original data and improves resampling capabilities by leveraging decomposed components. The results highlight MR-BoF's effectiveness in handling irregular sampling rates and data from diverse acquisition systems, making it a valuable tool for applications in finance, healthcare, industrial monitoring, and sensor networks.
Ref: Salazar Torres, D.O.; Altinses, D.; Schwung, A. Resampling Multi-Resolution Signals Using the Bag of Functions Framework: Addressing Variable Sampling Rates in Time Series Data. Sensors 2025, 25, 4759. https://doi.org/10.3390/s25154759
Continual Learning by Gradient Monitoring
Rafay Bin Shah
Adaptive Fusion of Intermodal and Intramodal Correlations for Fault-tolerance
Diyar Altinses
Minimizing disruptions in large-scale industrial operations is essential, as unexpected downtime leads to considerable expenses from wasted energy, labor, and materials. Traditional predictive maintenance strategies are effective for managing gradual degradation but are insufficient when failures occur abruptly and without warning. To address this limitation, we present the Multimodal Concept Fusion Architecture, a deep learning framework specifically designed to integrate diverse, noisy, and multimodal data representations. The framework emphasizes the fusion of latent features across structurally different modalities, using a learned concept that dynamically balances shared and contrasting information. Beyond improving fusion, this approach naturally supports anomaly detection and enhances model interpretability. By incorporating concept-guided fusion, the method allows decision processes to be traced back to the influence of individual modalities, thereby improving transparency and reliability. We validate the effectiveness and robustness of this architecture through experiments on three carefully designed multimodal datasets that simulate industrial scenarios, including sensor malfunctions introduced via targeted data augmentation.
Flexible Manufacturing Systems intralogistics: Dynamic optimization of AGVs and tool sharing using Colored-Timed Petri Nets and actor–critic RL with actions masking
Sofiene Lassoued, Laxmikant Shrikant Baheti, Nathalie Weiß-Borkowski
We present a novel approach for optimizing Flexible Manufacturing Systems by combining Colored-Timed Petri Nets with Reinforcement Learning. Our method enables real-time machine scheduling while coordinating automated guided vehicles and shared tools, addressing complex shop floor operations efficiently. Unlike traditional heuristics or metaheuristics, it retains learned knowledge for faster decision-making, adapts to dynamic changes, and scales to large production scenarios, significantly reducing computation time while maintaining high-quality schedules.
Machine Learning Driven Combinatorial Optimization of Logistical Operations in Drone-based Delivery Logistics
Laxmikant Shrikant Baheti
This research addresses real-time multimodal drone logistics using adaptive learning, integrating reinforcement learning (RL), graph neural networks (GNNs), and hierarchical architectures for complex, uncertain tasks. It progressively delivers: (1) an RL framework for adaptive routing and scheduling; (2) GNN-based spatial awareness for context-sensitive decisions; (3) a hierarchical system for scalable coordination and real-time adaptation. This yields structure-aware, dynamically responsive agents, advancing intelligent drone delivery systems. Models are validated on a custom simulation platform.
Statistical methods for smarter, transparent AI
Michael Walters
There is currently no consensus (and scarce reason to claim) that language models contain symbolic structures internally. Yet these black boxes are regularly chained together in sequence and packaged as agents so they may act in the world, while their word is increasingly unquestioned. We present how mathematically grounded reasoning can be brought into the fold, through our architectures GRAIL and GAIA Project, to enhance agentic abilities and grant transparency for stakeholders.
Zurück zum Programm
