Autonomous robots play an important role in modern indoor navigation, but existing systems often struggle with seamless human interaction and semantic understanding of environments. This paper presents an Artificial Intelligence (AI)-driven object recognition system enhanced by Large Language Models (LLMs), such as GPT-4 Vision and Gemini, to bridge this gap. Our approach combines vision-based mapping techniques with natural language processing and interactions to enable intuitive collaboration in solving navigation tasks. By leveraging multimodal input and vector space analysis, our system achieves enhanced object recognition, semantic embedding, and context-aware responses, setting a new standard for autonomous indoor navigation. This approach provides a novel framework for improving spatial understanding and dynamic interaction, making it suitable for complex indoor environments.
02261nas a2200169 4500000000100000000000100001008004100002260001200043100001800055700001900073700001700092245007800109856005000187490000700237520183300244022001402077 2025 d c07/2025 1 aNur Kelesoglu 1 aMarzena Halama 1 aAnna Strzoda 00 aEnhancing LoRa-based Outdoor Localization Accuracy Using Machine Learning uhttps://ieeexplore.ieee.org/document/11080010 0 v13 3 aThe Internet of Things is gaining significant relevance, driving increasing interest in location-
based services using wireless signals, particularly Low Power Wide Area Network (LPWAN) technology.
LoRa (Long Range), together with LoRaWAN, is a prominent LPWAN standard that provides long-range
connectivity and low energy consumption, making it viable for IoT-based positioning systems in smart
cities. For localization systems leveraging LoRa signals, Machine Learning (ML) approaches are being
increasingly explored, as ML-based solutions offer a powerful way to enhance the accuracy of positioning.
In this study, we propose various ML approaches for LoRa-based positioning in outdoor environments. We
evaluate six different ML models: k-NN, CNN, SVR, ANN, XG-Boost, and LightGBM-using an open-
source urban LoRaWAN dataset. We further propose a Hybrid Model that combines convolutional feature
extraction with gradient-boosted regression. This architecture integrates the strengths of deep learning and
tree-based models, aiming to capture both temporal signal patterns and structured input correlations for
improved localization accuracy. The models are trained offline and tested for performance in terms of
localization accuracy, mean square error, and computational efficiency. Additionally, we investigate the
impact of different Feature Vector (FV) subsets on localization performance by analyzing the significance of
LoRaWAN signal attributes. Our results highlight the effectiveness of ML models in enhancing localization
accuracy for LoRa-based outdoor positioning systems, demonstrating performance improvements ranging
from 10% to 73% compared to previous ML studies in outdoor localization.
Artificial Intelligence (AI) systems, including robotic assistants and voice interfaces, are increasingly present in public spaces such as airports, train stations, hospitals, and government offices, offering new opportunities to enhance user safety and comfort. Detecting aggression and hate speech in audio data from such locations is critical to protecting users and creating safe spaces. Previous approaches in this area have mainly relied on classification using classical Machine Learning (ML) methods and Deep Learning (DL) techniques, primarily focused on pure text analysis. However, most systems do not take into account real-world challenges such as background noise, incomplete utterances, or transcription errors resulting from automatic speech recognition. In this work, we propose a comprehensive approach that exploits the potential of Large Language Models (LLMs) to classify aggressive content in degraded audio transcriptions. We create a realistic test environment by converting tagged social media posts with offensive content into recordings with simulated noise and compare the performance of the models under different data quality conditions. Our research provides practical guidance on how to build more resilient and accessible voice moderation systems using LLM.
01322nas a2200133 4500000000100000000000100001008004100002260009600043100001900139245003800158856026000196520071000456020002201166 2025 d c01/2025 b1st International Conference PUT STEM Day 2025: Book of Abstracts aPoznań, Polska 1 aMarzena Halama 00 aOptimization for Pre-trained CNNs uhttps://putpoznanpl-my.sharepoint.com/personal/stemday_put_poznan_pl/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fstemday%5Fput%5Fpoznan%5Fpl%2FDocuments%2FKN%20PUT%20STEM%2FStemDay%2FSTEM%20Day%202026%2FPUT%20STEM%20Day%202025%2F1st%20International%20Confe 3 aConvolutional neural networks (CNNs) constitute a fundamental cornerstone of computer vision. With increasing complexity, the need for effective optimisation strategies remains crucial. Techniques such as transfer learning (TL), utilising pre-trained networks, enable the deployment of advanced models on mobile devices with limited computing capacity, including autonomous vehicles and educational applications. The study explores optimisation strategies for Keras models, focusing on the impact of different algorithms on performance and accuracy. The results demonstrate that appropriate optimiser selection enhances learning efficiency, mitigates overlearning, and supports accurate image recognition. a978-83-955437-7-7 01037nas a2200133 4500000000100000000000100001008004100002100001900043700002100062700002000083700002700103245005100130520072200181 2025 d 1 aMarzena Halama 1 aJoanna Domańska 1 aAnna Gancarczyk 1 aMirzokhid Abdirakhimov 00 aTworzenie bazy danych zeolitów z użyciem LLM 3 aPraca dotyczy opracowania zautomatyzowanego systemu tworzenia i zarządzania bazą danych zeolitów z wykorzystaniem dużych modeli językowych (LLM). Zaprojektowane rozwiązanie umożliwia ekstrakcję, ujednolicanie i strukturyzację informacji dotyczących syntezy, właściwości oraz zastosowań zeolitów na podstawie dokumentów tekstowych, takich jak publikacje naukowe i raporty badawcze. W procesie budowy bazy wykorzystano dane International Zeolite Association – Structure Commission oraz model LLM do identyfikacji kluczowych parametrów syntezy. Opracowana baza danych stanowi fundament pod przyszły system rekomendacyjny wspierający dobór odpowiednich zeolitów do konkretnych zastosowań.
01663nas a2200121 4500000000100000008004100001260007000042100001900112700001800131700002100149245008600170520128500256 2026 d b2nd International Conference PUT STEM Day 2026: Book of Abstracts 1 aMarzena Halama 1 aKonrad Połys 1 aJoanna Domańska 00 aAI-Assisted Zeolite Design through Retrieval-Augmented Scientific Language Models 3 aDesigning materials such as zeolites is a major challenge due to the wide range of possible structures and synthesis conditions. Furthermore, knowledge on this subject is widely scattered across numerous publications and experimental notes. Therefore, large language models (LLMs) represent a promising tool for building knowledge bases, supporting the analysis of scientific literature, although their responses may be incomplete or fallible. In addition, in the case of zeolites, the relationship between material properties and their topology is of crucial importance, which limits the effectiveness of approaches based solely on text. In this study, we investigate a Retrieval-Augmented Generation (RAG) approach extended with a multimodal component that combines textual and structural information.
00343nas a2200121 4500000000100000008004100001260002500042100001900067700001800086700002000104700002100124245007600145 2026 d bAGH University Press 1 aMarzena Halama 1 aKonrad Połys 1 aSławomir Nowak 1 aJoanna Domańska 00 aExternal Semantic Memory as a Cognitive Layer for Large Language Models 01478nas a2200109 4500000000100000008004100001100001900042700002200061245010500083520116600188022001401354 2026 d 1 aMarzena Halama 1 aNatalia Polańska 00 aBielik i GPT w analizie polskiego prawa karnego: wpływ RAG i grafów wiedzy na jakość odpowiedzi 3 a
Rozwój dużych modeli językowych stwarza nowe możliwości wspomagania wyszukiwania i analizy informacji prawnych. Problemem pozostaje ryzyko, że modele generują wiarygodnie brzmiące odpowiedzi, które nie mają podstaw w przepisach lub orzecznictwie, zwłaszcza gdy trzeba połączyć informacje z kilku źródeł prawnych. Celem badania było porównanie jakości odpowiedzi modeli językowych na pytania z zakresu polskiego prawa karnego z wykorzystaniem różnych sposobów dostępu do wiedzy zewnętrznej. Analizie poddano 4 warianty: model bez kontekstu; Retrieval-Augmented Generation (RAG) oparty na wyszukiwaniu wektorowym; rozwiązanie wykorzystujące graf wiedzy oraz architekturę hybrydową łączącą wyszukiwanie wektorowe i grafowe. Badanie obejmowało ocenę podobieństwa semantycznego i kompletności odpowiedzi, jakości wyszukiwania źródeł oraz zdolności do rozpoznawania braku informacji w korpusie. Szczególną uwagę poświęcono pytaniom wymagającym łączenia informacji pochodzących z wielu źródeł.
a1734-5391