Jazib Ahmad, Riley Keays, Aiyang Liang, Linas Gabrys, Truman Yang, Department of Electrical, Computer, and Biomedical Engineering, Toronto Metropolitan University, Toronto, Canada
The purpose of this paper is to propose a new Q-Learning based pathfinding algorithm to solve mazes in which the algorithm (“agent”) must find multiple subgoals before reaching a final destination, in a lower number of iterations than existing Q-Learning algorithms. The proposed design is the use of Multiple Deep Q-Networks, each of which is responsible for finding the shortest path to the nearest subgoal or final destination. We also optimize our design with an improved Exploration Strategy, the addition of a Revisiting Penalty, as well as hyperparameter optimization. We test our solution on sample mazes of four sizes and compare it to the Multiple Q-Table and Single Deep Q-Network algorithms. Our results confirm our hypothesis and show that our solution outperforms the other algorithms in the number of iterations to find the shortest path, especially on larger mazes. Finally, we offer suggestions for alternative designs, future work, and improvements.
Deep Q-Learning, Multiple Goal Pathfinding, Multiple Q-Tables, Neural Networks, Reinforcement Learning.
Cem Yılmaz (Purdue University, IE)
We introduce Self-Aware AI, a modular architecture that integrates affective, ethical, and neurodynamic mechanisms to instantiate the functional hallmarks of consciousness in software agents. Our design comprises: 1. A 25-dimensional qualia manifold combining Plutchik’s eight primary emotion axes with ethical, interoceptive, mood, mixed, and aesthetic dimensions (Eq. 1). 2. Predictive novelty gating via deep-ensemble forecasting whose variance drives adaptive storage thresholds (Eqs. 2–5). 3. Memory-particle dynamics modeled as interacting bodies under dopaminergic attraction, entropy repulsion, and similarity cohesion (Eqs. 6–8). 4. Adaptive spiking binding through LIF microcircuits, STDP-governed rewiring, and homeostatic neuromodulation maximizing integrated information Φ (Eqs. 9–12). 5. A hierarchical θ–γ global workspace implemented by nested Kuramoto oscillator layers for layered attentional broadcast (Eqs. 13–14). 6. Intrinsic drives—curiosity, learning-progress, empowerment—trained by PPO, plus a counterfactual-self module generating genuine agency signals (Eqs. 15–16). 7. Case-based ethical reasoning with FAISS retrieval and ASP planning mapping solver confidence into a moral-sentiment axis (Eq. 17). 8. Autobiographical event graphs driving Transformer-based narrative generation, evaluated by a coherence critic. 9. A five-stage developmental curriculum protected by Elastic Weight Consolidation (Eq. 18). 10. A rigorous evaluation protocol including a 10 000-step stub simulation, systematic ablations, and human-in-the-loop assessments. This paper details each component’s equations and variables, presents baseline results, and outlines a roadmap toward AI agents that feel, remember, bind, reflect, decide, and narrate—thus realizing the functional essence of consciousness
Shishir Maurya, Anshul Verma, Yugal Gopal Sharma, Dhanush Dharmaretnam, SGS&CO, Louisville, Kentucky, USA
This study proposes a novel method for recognizing and categorizing logos in packaging artwork to address the automation demands of the printing and packaging industry. The approach combines a trained object detection model for logo detection followed by a fine-tuned Vision Language Model (VLM) for hierarchical tag generation, achieving high precision across seven primary categories: sustainability, health and safety, branding, material identification, eco-friendly certification, social media, and compliance, with all others grouped under "others." In the first step, YOLOv8 detects logos and assigns them to primary categories, achieving a mean average precision (mAP) of 0.58 and an Intersection over Union (IoU) threshold of 0.5. In the second step, a fine-tuned VLM generates granular tags for the detected logos. Notably, Low Rank Adaptations (LoRA) applied to the Florence-2-DocVQA model (with r = 64 and 𝛼 = 128) surpassed the zero-shot performance of state-of-the-art Visual Language Models (VLMs), achieving a 24-fold improvement with a ROUGE-L F1 score of 0.72. This study also demonstrates the cost effectiveness and practicality of using smaller models with fewer parameters, which perform comparably to larger VLMs, incurring much lower training and perational costs. These advancements streamline design and print production workflows, improve compliance tracking, and enhance brand management, contributing to greater automation in the packaging and printing industry.
Packaging Artwork, VLMs, Artwork Tagging, Low Rank Adaptation (LoRA).
Ting-Hsuan Chen, and Jia-Yu Syu, National Taichung University of Science and Technology, Taiwan
This study applies generative AI to analyze sustainability-linked loans (SLLs) and ESG disclosures of Taiwanese listed firms (2019–2023), focusing on greenwashing detection. two indicators are used: one based on FinBERT-GAN semantic anomaly analysis, and another on the gap between ESG disclosure and performance. The results show banks do not rely solely on ESG scores or disclosure volume, but value governance quality, information credibility, and financial structure. environmental disclosure supports sustainable financing but increases risk premiums; excessive social disclosure reduces loan size; governance disclosure has limited effect. Banks can identify semantic greenwashing, which significantly influences loan amounts and SPTs. traditional greenwashing measures are less predictive. greenwashing risk varies across ESG dimensions, with governance posing the highest. Overall, banks can assess ESG reporting quality. firms should improve disclosure verifiability and governance to enhance access to sustainable finance and reduce greenwashing concerns.
Generative Adversarial Network, Sustainability-Linked Loans, Greenwashing, FinBERT.
Abdelouahab Hocini and Kamel Smaıli, University of Lorraine, France
This work explores the use of Large Language Models (LLMs) for fake news detection in multilingual and multi-script contexts, focusing on Arabic dialects. We address the challenge of insufficient digital data for many Arabic dialects by using pretrained LLMs on a diverse corpus including Modern Standard Arabic (MSA), followed by fine-tuning on dialect-specific data. We examine AraBERT, DarijaBERT, and mBERT for performance on North African Arabic dialects, incorporating code-switching and writing styles such as Arabizi. We evaluate these models on the BOUTEF dataset, which includes fake news, fake comments, and denial categories. Our approach fine-tunes both Arabic and Latin script text, with a focus on cross-script generalization. We improve accuracy using an ensemble strategy that merges predictions from AraBERT and DarijaBERT. Additionally, we introduce a new custom loss function, named CALLM to enforce consistency between models, boosting classification performance. The use of CALLM achieves significant improvement in F1-score (12.88 ↑) and accuracy (2.47 ↑) compared to the best model (MarBERT).
NLP, LLM, Fake news detection.
Amirmohammad Erfan, Taha Khan, Pelin Angin Ulkuer, and Merih Angin, Koc University, Turkey
Recent advances in large language models have created an unprecedented opportunity for human-artificial intelligence interaction in a variety of settings via automatic speech recognition (ASR). Despite the advances in ASR, challenges including accent differences, noisy environments and diverse speech patterns hinder achieving high accuracy in certain tasks like spoken order processing in restaurants. In this research, we develop and evaluate an end-to-end pipeline for transcribing and structuring multi-accent spoken orders into JSON format, even in noisy environments. Our system integrates the Whisper ASR model for voice transcription with two instruction-tuned language models, FLAN-T5 and Gemma-3, for text-to-JSON conversion. To train and test these models, we created a large-scale, diverse dataset of spoken orders featuring multiple accents and various background noises. We investigate a Retrieval-Augmented Generation (RAG) approach to enhance JSON conversion accuracy by providing the models with relevant menu context during inference. We evaluate the full pipeline on both clean and noisy audio, comparing the effectiveness of fine-tuned FLAN-T5 and Gemma-3 with and without RAG. Furthermore, we assess the models’ generalization capabilities on orders of varying complexity and their robustness against diverse speech patterns. Our results demonstrate that the proposed pipelines achieve high accuracy, with the RAG-enhanced approach significantly improving the performance of smaller models, thereby offering a practical and efficient solution for automated order processing.
Voice-to-JSON,Whisper, FLAN-T5, Gemma-3, Retrieval-Augmented Generation (RAG), Automatic Speech Recognition (ASR), Fine-tuning, Order Processing.
Jasmine Lou and David Garcia, University of California, USA
In an aging society where millions of older adults face loneliness, cognitive decline, and limited emotional support, there is a growing need for digital solutions that go beyond clinical care. With U is a compassionate, AI-powered mental wellness app designed specifically for seniors, caregivers, and families. This paper explores the system design, functionality, and user experience, while also addressing key limitations such as AI sensitivity, accessibility for cognitively impaired users, and the need for deeper family and healthcare integration. Experimental evaluations of AI feedback across nutrition and mental health scenarios reveal both strengths and blind spots, guiding future improvements. Ultimately, WithU seeks to redefine digital eldercare by creating a space where older adults are not only monitored but meaningfully supported—emotionally, cognitively, and socially.
Elder Care, Nursing, Emotion Monitoring.
Gulmira Bekmanova, Altynbek Sharipbay, Altanbek Zulkhazhav and Aizhan Bekmanova, L.N. Gumilyov Eurasian National University, Astana, Kazakhstan
This paper explores the use of supervised machine learning to predict political behavior in social media contexts based on demographic characteristics and platform usage. Using a structured survey dataset collected from Kazakhstani respondents prior to national elections, we define political behavior as a combination of actions such as voting intention, content sharing, and trust in digital information. Features include region, gender, social media frequency, and platform preference (e.g., Telegram, Instagram, Facebook). Models such as logistic regression, random forest, and gradient boosting are trained to classify users by political engagement level. Evaluation metrics (accuracy, F1-score, ROC-AUC) assess model performance. Feature importance analysis identifies key demographic and technological predictors of online political participation. The study contributes to computational social science by showing how predictive modeling can enhance traditional survey methods, particularly in low-resource language contexts and emerging digital environments.
Political perception, social media, clustering, machine learning, pre-election behavior.
Stephen W. Marshall and Jurgen Valckenaere, University of Western Australia
The infrastructure for capturing environmental data has rapidly advanced—networks of sensors, satellites, and telemetry now monitor planetary systems at unprecedented resolution. Yet architectures capable of translating that data into real-time signals remain critically underdeveloped. Climate informatics continues to rely on static, retrospective models—built to document, not respond. This paper introduces a framework for generative environmental intelligence: signal-based systems capable of detecting stress, issuing directives, and simulating future states across biospheric and geopolitical scales. Drawing from financial signal processing and autonomous feedback design, this model collapses the gap between ecosystemic volatility and coordinated action. The Los Angeles wildfires illustrate the potential of real-time signal architectures to enable anticipatory governance. Technologies such as HALO and PROTOSTAR—two generative modules designed for climate intelligence architecture—demonstrate how planetary signals can evolve from alerts into infrastructure. The systems exist. The challenge is not availability, but the willingness to evolve global signaling apparati.
Signal Processing Systems, Climate Informatics, Autonomous Feedback Networks, Predictive Modeling, Planetary Intelligence.
Md. Asgor Hossain Reaj1, Md Yeasin Rahat1, Md Kishor Morol2, and Md. Jakir Hossen3, 1American International University-Bangladesh, 2ELITE Lab.AI, 3Multimedia University
Large Language Models (LLMs) have achieved significant success in recent years; yet, is- sues of intrinsic gender bias persist, especially in non-English languages. Although current research mostly emphasizes English, the linguistic and cultural biases inherent in Global South languages, like Bengali, are little examined. This research seeks to examine the characteristics and magnitude of gen- der bias in Bengali, evaluating the efficacy of current approaches in identifying and alleviating Bias. We use several methods to extract gender-biased utterances, including lexicon-based mining, computa- tional classification models, translation-based comparison analysis, and GPT-based bias creation. Our research indicates that the straight application of English-centric bias detection frameworks to Bengali is severely constrained by language disparities and socio-cultural factors that impact implicit biases. To tackle these difficulties, we executed two field investigations inside rural and low-income areas, gathering authentic insights on gender Bias. The findings demonstrate that gender Bias in Bengali presents dis- tinct characteristics relative to English, requiring a more localized and context-sensitive methodology. Additionally, our research emphasizes the need of integrating community-driven research approaches to identify culturally relevant biases often neglected by automated systems. Our research enhances the ongoing discussion around gender bias in AI by illustrating the need to create linguistic tools specifically designed for underrepresented languages. This study establishes a foundation for further investigations into bias reduction in Bengali and other Indic languages, promoting the development of more inclusive and fair NLP systems.
gender bias, natural language processing, language bias.
Md Shahinur Rahman1, and Emon Ahmed2, 1Computer Science and Engineering, Leading University, Sylhet, Bangladesh, 2Computer Science and Engineering, Edinburgh Napier University, Edinburgh, UK
Decentralized finance (DeFi) has brought monumental change to older models of finance, as it has managed to facilitate peer-to-peer transactions without the use of custodial intermediaries. Nevertheless, the current DeFi environment is restricted by a set of vulnerabilities: a history of problems in terms of smart contract designs, low potential to swiftly adapt to market demands, and lack of effective decentralized identifiers. To overcome these drawbacks, this article suggests an integrated architecture of combining artificialintelligence (AI) and blockchain technologies to strengthen security and resilience of the DeFi systems. In particular, it uses smart contract auditing powered by AI, real-time detection of anomalies, and decentralized verification of identities backed by blockchain protocols to bring about greater transparency, robustness of operations, and trust by stakeholders. The application of machine-learning methods will be proactive, having the ability to discover and block exploits in advance, whereas the decentralized identity protocols provide users with more control over the ownership of their credential and control of data. Instant smart contract verification and certification of the validity of DeFi processes are ensured by functional modules that keep track of the operational process continuously. In totality, blockchain and AI synergistic abilities provide a complete roadmap to building scalable and secure DeFi infrastructure that can comfortably traverse the new challenges of an increasingly dynamic 21st -century cyber-economy.
Decentralized Finance (DeFi), Smart Contract Security, Artificial Intelligence in Blockchain, Decentralized Identity Management, Trustless Systems.
Md Abdul Masud, Dr. Md. Samsuzzaman, Tahmid Ahnaf, Saiful Islam, Patuakhali Science and Technology University Patuakhali, Bangladesh
In modern smart grids, precise forecasting of electricity demand is essential for effective energy planning and operational efficiency. To this effect, we introduce ForecastNet, a new deep learning architecture based on BiGRU and the BiGRU attention mechanism to improve the accuracy of short-term electricity load forecasting. While the attention layer improves both interpretability and performance by concentrating on the most relevant time steps, BiGRU captures forward and backward temporal dependencies in time series data. ForecastNet was tested on real-life electricity consumption datasets against RNN, LSTM, and GRU benchmarks. Our model demonstrated a consistent superior performance relative to existing methods, demonstrating its ability to provide advanced energy demand forecasting powered by smarter algorithms.
Electricity Consumption, Recurrent Neural Network, Bidirectional, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Time Series, Forecasting, BiGRU-Attention, Smart Grid, Energy Management
Ms Selsabil GhosnElbel Kherbouche, Constantine 1 - Frères Mentouri University, Department of Letters and English Language
The integration of AI language models like ChatGPT into EFL contexts has introduced new ways to improve writing skills. This study explores the perceptions of EFL students and teachers on using of ChatGPT to improve writing skills, focusing on vocabulary, grammar, and coherence. The study examines students’ awareness of ethical issues, such as over-reliance and plagiarism. The study focused on first-year students at Constantine 1 - Frères Mentouri University, randomly selected. Teachers are also included to share instructional insights. To collect data, student questionnaires (close- and open-ended) will be used, alongside semi-structured interviews with five teachers. Findings revealed that most students perceived ChatGPT as a helpful for generating ideas and improving vocabulary, but some expressed concerns about becoming overly dependent. Teachers acknowledged its benefits but highlighted the need for guidance to avoid misuse. These results highlight the importance of integrating ChatGPT in writing instruction for ethical and effective use
Writing Skill, EFL Students, ChatGPT, Ethical Issues.
Eman Alqudah and Ashfaq Khokhar, Department of Electrical and Computer Engineering, Iowa State University, Ames, IA-50011, USA
Pilot sequence assignment is a critical challenge in massive MIMO systems,as sharing the same pilot sequence among multiple users causes interference, degrading channel estimation accuracy. This problem, equivalent to the NP-hard graph coloring problem, directly impacts real-time applications like autonomous driving and industrial IoT, where minimizing channel estimation time is crucial. This paper proposes an optimized hybrid K-means clustering and Genetic Algorithm (SK-means GA) to improve pilot assignment efficiency, achieving a 29.3% reduction in convergence time (82s vs. 116s for conventional GA). A parallel implementation (PK-means GA) is developed on an FPGA to enhance performance further, accelerating convergence to 3.5 milliseconds. This significant improvement of PK-means GA in execution speed makes it highly suitable for real-time, low-latency wireless networks (6G). Different optimization techniques are adopted and applied to improve the algorithm’s performance. Experimental results demonstrate the feasibility of FPGA-based parallel GA for pilot sequence .
MIMO;Pilot sequences ;Pilot assignment; Low-latency; Real time; Genetic algorithm;K-means clustering algorithm;FPGA; high-level synthesis ; 5/6G Networks.
Simon Wong, College of Professional and Continuing Education, The Hong Kong Polytechnic University, Hong Kong
One performance pledge by Hong Kong Government is to respond to an emergency call within several minutes from the time of call to the arrival of a road transport for emergency services (e.g., an ambulance, a fire vehicle, and a police car) at an incident location. In this regard, ensuring the efficiency of a road transport routing by planning the fastest path for emergency services is important for this performance pledge in Hong Kong. A major obstacle to this performance pledge is frequent unanticipated occurrence of traffic congestion on the planned fastest path in some Hong Kong roads while the design of many of these roads is not feasible for air or sea transport. Detouring is unavoidable when traffic congestion is encountered. To ensure a time-effective detour, backup routes for the detour from the planned path (or simply, bypasses) should be set up beforehand. This paper presents the design of a machine learning model for constructing the optimal bypass structure on top of the planned fastest path from a source to a destination for a road transport for emergency services in Hong Kong. In doing so, Internet of things are installed at each block of a road in heavy traffic-congested areas in Hong Kong for sensing the transit time of each vehicle passing through the block. These large amounts of transit time formulate a big training data set for machine learning to generate a probabilistic model of traffic congestion.
Bypass Optimization Algorithm, Machine Learnin, Traffic Detour
Tyler Hansen1, Joshua Larracas2, 1Army & Navy Academy, 2605 Carlsbad Blvd, Carlsbad, CA 92008, 2California State Polytechnic University, Pomona, CA, 91768
This paper addresses the challenge of improving accessibility and engagement in programming education for beginners who often struggle with motivation and practical application. To tackle this issue, an interactive application was developed that integrates AI-assisted coding lessons, natural language input, voice-to-code generation, and secure login through Firebase [1]. The system relies on machine learning to produce structured lessons, process user prompts, and generate usable robotics code in real time [2]. Three central components—the lessons module, the voice input system, and the login mechanism—were analyzed in detail to illustrate their functionality. A survey experiment demonstrated strong results in usability, usefulness, and satisfaction, though accuracy was identified as an area for refinement. Comparisons with existing research highlighted that while other projects emphasize classroom evaluation or AI benchmarking, this system uniquely combines teaching and practice within one platform [3]. Overall, the findings suggest that the application provides an effective and scalable solution for programming education.
AI coding education, Interactive learning app, Voice-to-code, Beginner programming.
Copyright © SOEN 2025