Research Papers

Automatic reading aloud evaluation can provide valuable support to teachers by enabling more efficient scoring of reading exercises. However, research on reading evaluation systems and applications remains limited. We present a novel multimodal approach that leverages audio and knowledge from text resources. In particular, we explored the potential of using Whisper and instruction-tuned large language models (LLMs) with prompts to improve transcriptions for child speech recognition, as well as their effectiveness in downstream reading mistake detection. Our results demonstrate the effectiveness of prompting Whisper and prompting LLM, compared to the baseline Whisper model without prompting. The best performing system achieved state-of-the-art recognition performance in Dutch child read speech, with a word error rate (WER) of 5.1%, improving the baseline WER of 9.4%. Furthermore, it significantly improved reading mistake detection, increasing the F1 score from 0.39 to 0.73.

We introduce a network-based AI framework for predicting dimensions of psychopathology in adolescents using natural language. We focused on data capturing psychometric scores of social maladjustment, internalizing behaviors, and neurodevelopmental risk, assessed in 232 adolescents from the Healthy Brain Network. This dataset included structured interviews in which adolescents discussed a common emotion-inducing topic. To model conceptual associations within these interviews, we applied textual forma mentis networks (TFMNs)-a cognitive/AI approach integrating syntactic, semantic, and emotional word-word associations in language. From TFMNs, we extracted network features (semantic/syntactic structure) and emotional profiles to serve as predictors of latent psychopathology factor scores. Using Random Forest and XGBoost regression models, we found significant associations between language-derived features and clinical scores: social maladjustment (r = 0.37, p < .01), specific internalizing behaviors (r = 0.33, p < .05), and neurodevelopmental risk (r = 0.34, p < .05). Explainable AI analysis using SHAP values revealed that higher modularity and a pronounced core-periphery network structure-reflecting clustered conceptual organization in language-predicted increased social maladjustment. Internalizing scores were positively associated with higher betweenness centrality and stronger expressions of disgust, suggesting a linguistic signature of rumination. In contrast, neurodevelopmental risk was inversely related to local efficiency in syntactic/semantic networks, indicating disrupted conceptual integration. These findings demonstrated the potential of cognitive network approaches to capture meaningful links between psychopathology and language use in adolescents.

Evaluating Large Language Models (LLMs) for safety and security remains a complex task, often requiring users to navigate a fragmented landscape of ad hoc benchmarks, datasets, metrics, and reporting formats. To address this challenge, we present aiXamine, a comprehensive black-box evaluation platform for LLM safety and security. aiXamine integrates over 40 tests (i.e., benchmarks) organized into eight key services targeting specific dimensions of safety and security: adversarial robustness, code security, fairness and bias, hallucination, model and data privacy, out-of-distribution (OOD) robustness, over-refusal, and safety alignment. The platform aggregates the evaluation results into a single detailed report per model, providing a detailed breakdown of model performance, test examples, and rich visualizations. We used aiXamine to assess over 50 publicly available and proprietary LLMs, conducting over 2K examinations. Our findings reveal notable vulnerabilities in leading models, including susceptibility to adversarial attacks in OpenAI's GPT-4o, biased outputs in xAI's Grok-3, and privacy weaknesses in Google's Gemini 2.0. Additionally, we observe that open-source models can match or exceed proprietary models in specific services such as safety alignment, fairness and bias, and OOD robustness. Finally, we identify trade-offs between distillation strategies, model size, training methods, and architectural choices.

Connecting Feedback to Choice: Understanding Educator Preferences in GenAI vs. Human-Created Lesson Plans in K-12 Education -- A Comparative Analysis

Shawon Sarkar, Min Sun, Alex Liu +4

As generative AI (GenAI) models are increasingly explored for educational applications, understanding educator preferences for AI-generated lesson plans is critical for their effective integration into K-12 instruction. This exploratory study compares lesson plans authored by human curriculum designers, a fine-tuned LLaMA-2-13b model trained on K-12 content, and a customized GPT-4 model to evaluate their pedagogical quality across multiple instructional measures: warm-up activities, main tasks, cool-down activities, and overall quality. Using a large-scale preference study with K-12 math educators, we examine how preferences vary across grade levels and instructional components. We employ both qualitative and quantitative analyses. The raw preference results indicate that human-authored lesson plans are generally favored, particularly for elementary education, where educators emphasize student engagement, scaffolding, and collaborative learning. However, AI-generated models demonstrate increasing competitiveness in cool-down tasks and structured learning activities, particularly in high school settings. Beyond quantitative results, we conduct thematic analysis using LDA and manual coding to identify key factors influencing educator preferences. Educators value human-authored plans for their nuanced differentiation, real-world contextualization, and student discourse facilitation. Meanwhile, AI-generated lesson plans are often praised for their structure and adaptability for specific instructional tasks. Findings suggest a human-AI collaborative approach to lesson planning, where GenAI can serve as an assistive tool rather than a replacement for educator expertise in lesson planning. This study contributes to the growing discourse on responsible AI integration in education, highlighting both opportunities and challenges in leveraging GenAI for curriculum development.

Road fatalities pose significant public safety and health challenges worldwide, with pedestrians being particularly vulnerable in vehicle-pedestrian crashes due to disparities in physical and performance characteristics. This study employs explainable artificial intelligence (XAI) to identify key factors contributing to pedestrian fatalities across the five U.S. states with the highest crash rates (2018-2022). It compares them to the five states with the lowest fatality rates. Using data from the Fatality Analysis Reporting System (FARS), the study applies machine learning techniques-including Decision Trees, Gradient Boosting Trees, Random Forests, and XGBoost-to predict contributing factors to pedestrian fatalities. To address data imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is utilized, while SHapley Additive Explanations (SHAP) values enhance model interpretability. The results indicate that age, alcohol and drug use, location, and environmental conditions are significant predictors of pedestrian fatalities. The XGBoost model outperformed others, achieving a balanced accuracy of 98 %, accuracy of 90 %, precision of 92 %, recall of 90 %, and an F1 score of 91 %. Findings reveal that pedestrian fatalities are more common in mid-block locations and areas with poor visibility, with older adults and substance-impaired individuals at higher risk. These insights can inform policymakers and urban planners in implementing targeted safety measures, such as improved lighting, enhanced pedestrian infrastructure, and stricter traffic law enforcement, to reduce fatalities and improve public safety.

Federated Learning (FL) enables collaborative model training while preserving data privacy; however, balancing privacy preservation (PP) and fairness poses significant challenges. In this paper, we present the first unified large-scale empirical study of privacy-fairness-utility trade-offs in FL, advancing toward responsible AI deployment. Specifically, we systematically compare Differential Privacy (DP), Homomorphic Encryption (HE), and Secure Multi-Party Computation (SMC) with fairness-aware optimizers including q-FedAvg, q-MAML, Ditto, evaluating their performance under IID and non-IID scenarios using benchmark (MNIST, Fashion-MNIST) and real-world datasets (Alzheimer's MRI, credit-card fraud detection). Our analysis reveals HE and SMC significantly outperform DP in achieving equitable outcomes under data skew, although at higher computational costs. Remarkably, we uncover unexpected interactions: DP mechanisms can negatively impact fairness, and fairness-aware optimizers can inadvertently reduce privacy effectiveness. We conclude with practical guidelines for designing robust FL systems that deliver equitable, privacy-preserving, and accurate outcomes.

Bias Detection via Maximum Subgroup Discrepancy

Jiří Němeček, Mark Kozdoba, Illia Kryvoviaz +2

Bias evaluation is fundamental to trustworthy AI, both in terms of checking data quality and in terms of checking the outputs of AI systems. In testing data quality, for example, one may study the distance of a given dataset, viewed as a distribution, to a given ground-truth reference dataset. However, classical metrics, such as the Total Variation and the Wasserstein distances, are known to have high sample complexities and, therefore, may fail to provide a meaningful distinction in many practical scenarios. In this paper, we propose a new notion of distance, the Maximum Subgroup Discrepancy (MSD). In this metric, two distributions are close if, roughly, discrepancies are low for all feature subgroups. While the number of subgroups may be exponential, we show that the sample complexity is linear in the number of features, thus making it feasible for practical applications. Moreover, we provide a practical algorithm for evaluating the distance based on Mixed-integer optimization (MIO). We also note that the proposed distance is easily interpretable, thus providing clearer paths to fixing the biases once they have been identified. Finally, we describe a natural general bias detection framework, termed MSDD distances, and show that MSD aligns well with this framework. We empirically evaluate MSD by comparing it with other metrics and by demonstrating the above properties of MSD on real-world datasets.

In the rapidly advancing field of artificial intelligence, machine perception is becoming paramount to achieving increased performance. Image classification systems are becoming increasingly integral to various applications, ranging from medical diagnostics to image generation; however, these systems often exhibit harmful biases that can lead to unfair and discriminatory outcomes. Machine Learning systems that depend on a single data modality, i.e. only images or only text, can exaggerate hidden biases present in the training data, if the data is not carefully balanced and filtered. Even so, these models can still harm underrepresented populations when used in improper contexts, such as when government agencies reinforce racial bias using predictive policing. This thesis explores the intersection of technology and ethics in the development of fair image classification models. Specifically, I focus on improving fairness and methods of using multiple modalities to combat harmful demographic bias. Integrating multimodal approaches, which combine visual data with additional modalities such as text and metadata, allows this work to enhance the fairness and accuracy of image classification systems. The study critically examines existing biases in image datasets and classification algorithms, proposes innovative methods for mitigating these biases, and evaluates the ethical implications of deploying such systems in real-world scenarios. Through comprehensive experimentation and analysis, the thesis demonstrates how multimodal techniques can contribute to more equitable and ethical AI solutions, ultimately advocating for responsible AI practices that prioritize fairness.

Creativity is a fundamental skill of human cognition. We use textual forma mentis networks (TFMN) to extract network (semantic/syntactic associations) and emotional features from approximately one thousand human- and GPT3.5-generated stories. Using Explainable Artificial Intelligence (XAI), we test whether features relative to Mednick's associative theory of creativity can explain creativity ratings assigned by humans and GPT-3.5. Using XGBoost, we examine three scenarios: (i) human ratings of human stories, (ii) GPT-3.5 ratings of human stories, and (iii) GPT-3.5 ratings of GPT-generated stories. Our findings reveal that GPT-3.5 ratings differ significantly from human ratings not only in terms of correlations but also because of feature patterns identified with XAI methods. GPT-3.5 favours 'its own' stories and rates human stories differently from humans. Feature importance analysis with SHAP scores shows that: (i) network features are more predictive for human creativity ratings but also for GPT-3.5's ratings of human stories; (ii) emotional features played a greater role than semantic/syntactic network structure in GPT-3.5 rating its own stories. These quantitative results underscore key limitations in GPT-3.5's ability to align with human assessments of creativity. We emphasise the need for caution when using GPT-3.5 to assess and generate creative content, as it does not yet capture the nuanced complexity that characterises human creativity.

South Africa and the Democratic Republic of Congo (DRC) present a complex linguistic landscape with languages such as Zulu, Sepedi, Afrikaans, French, English, and Tshiluba (Ciluba), which creates unique challenges for AI-driven translation and sentiment analysis systems due to a lack of accurately labeled data. This study seeks to address these challenges by developing a multilingual lexicon designed for French and Tshiluba, now expanded to include translations in English, Afrikaans, Sepedi, and Zulu. The lexicon enhances cultural relevance in sentiment classification by integrating language-specific sentiment scores. A comprehensive testing corpus is created to support translation and sentiment analysis tasks, with machine learning models such as Random Forest, Support Vector Machine (SVM), Decision Trees, and Gaussian Naive Bayes (GNB) trained to predict sentiment across low resource languages (LRLs). Among them, the Random Forest model performed particularly well, capturing sentiment polarity and handling language-specific nuances effectively. Furthermore, Bidirectional Encoder Representations from Transformers (BERT), a Large Language Model (LLM), is applied to predict context-based sentiment with high accuracy, achieving 99% accuracy and 98% precision, outperforming other models. The BERT predictions were clarified using Explainable AI (XAI), improving transparency and fostering confidence in sentiment classification. Overall, findings demonstrate that the proposed lexicon and machine learning models significantly enhance translation and sentiment analysis for LRLs in South Africa and the DRC, laying a foundation for future AI models that support underrepresented languages, with applications across education, governance, and business in multilingual contexts.

Gender bias in machine translation (MT) systems poses significant challenges that often result in the reinforcement of harmful stereotypes. Especially in the labour domain where frequently occupations are inaccurately associated with specific genders, such biases perpetuate traditional gender stereotypes with a significant impact on society. Addressing these issues is crucial for ensuring equitable and accurate MT systems. This paper introduces a novel approach to studying occupation-related gender bias through the creation of the GOSt-MT (Gender and Occupation Statistics for Machine Translation) Knowledge Graph. GOSt-MT integrates comprehensive gender statistics from real-world labour data and textual corpora used in MT training. This Knowledge Graph allows for a detailed analysis of gender bias across English, French, and Greek, facilitating the identification of persistent stereotypes and areas requiring intervention. By providing a structured framework for understanding how occupations are gendered in both labour markets and MT systems, GOSt-MT contributes to efforts aimed at making MT systems more equitable and reducing gender biases in automated translations.

Linguistically inclusive LLMs -- which maintain good performance regardless of the language with which they are prompted -- are necessary for the diffusion of AI benefits around the world. Multilingual jailbreaks that rely on language translation to evade safety measures undermine the safe and inclusive deployment of AI systems. We provide policy recommendations to enhance the multilingual capabilities of AI while mitigating the risks of multilingual jailbreaks. We examine how a language's level of resourcing relates to how vulnerable LLMs are to multilingual jailbreaks in that language. We do this by testing five advanced AI models across 24 official languages of the EU. Building on prior research, we propose policy actions that align with the EU legal landscape and institutional framework to address multilingual jailbreaks, while promoting linguistic inclusivity. These include mandatory assessments of multilingual capabilities and vulnerabilities, public opinion research, and state support for multilingual AI development. The measures aim to improve AI safety and functionality through EU policy initiatives, guiding the implementation of the EU AI Act and informing regulatory efforts of the European AI Office.