Large language models (LLMs) are increasingly used in assistive conversational systems but often struggle to adapt to human tone and context. While prior work emphasizes factual accuracy and safety, less attention has been given to context sensitive conversational behavior. In this work, we explore whether real world interaction signals can improve context driven reward adaptability. We use Reddit conversations, as a proxy to group conversations, to train a reward model that predicts the effectiveness of replies in context, then fine-tune a language model with Proximal Policy Optimization (PPO) to encourage responses aligned with conversational tone and user expectations. Across benchmarks, the resulting models show improved humor and engagement while maintaining comparable reasoning ability, alongside shifts in toxicity and bias consistent with the training signal. These results suggest that alignment requires not only correctness, but also sensitivity to tone, intent, and conversational context.
Quantifying modality contributions in Vision-Language Models (VLMs) remains challenging. Existing approaches rely on perturbation or gradient-based methods, which conflate inherent modality informativeness with model-specific biases and fail to capture complex cross-modal interactions. We address this gap by introducing an information-theoretic framework based on Partial Information Decomposition (PID) that decomposes internal representations into unique, redundant, and synergistic components. Our method operates directly on internal embeddings and derives an inference-only modality contribution metric from unique information scores. Applying our framework to six modern VLMs across six benchmarks, we uncover a persistent imbalance in modality contributions driven by low cross-modal synergy. Analysis reveals that fusion architecture significantly impacts the distribution of unique, redundant, and synergistic information. Our framework provides a scalable diagnostic tool for understanding and improving multimodal integration in vision-language systems.
Clinician burnout threatens patient safety, care quality, and workforce sustainability, especially in high-acuity ICUs. Existing detection approaches rely on retrospective surveys or coarse EHR metadata, limiting their ability to capture the evolution of burnout-related stress. We analyze 10,000 ICU discharge summaries from the MIMIC-IV database and propose a narrative-driven, weakly supervised framework for provider-level surveillance of burnout risk. Our approach integrates BioBERT-based sentiment modeling, lexical stress cues, latent topic structure, structured workload proxies, and temporal dynamics. In the absence of survey ground truth, we use a quantile-based ordinal labeling strategy to distinguish low, medium, and high burnout risk. A logistic regression classifier achieves an F1 score of 0.84 for conservative high-risk screening, while temporal features enable trajectory-based monitoring without degrading point-in-time performance. Specialty-specific analysis reveals elevated narrative stress indicators among Radiology, Psychiatry, and Neurology providers. ICU clinical narratives encode actionable, longitudinal signals for scalable burnout surveillance beyond static sentiment or metadata-only approaches.
This tutorial will cover an introduction to modern neural networks and how they are applied to problems in artificial intelligence. Basic terminology, history, application methods, and application case studies will be covered. Modern topics such as deep learningwill be covered. The material will be filtered and summarized for the novice. Further reading and software packages and frameworks will also be discussed.
Automated planning approaches have proven effective in performing traffic signal optimisation, and their deployability has been demonstrated by their ability to incorporate constraints and features of the real-world infrastructure on which they will operate. A major constraint is the need to know in advance, for each junction of the controlled urban region, the set of configurations (i.e., the length of all stages) that can be considered for the optimisation process. Configurations therefore play a pivotal role as they effectively allow control of the traffic flows; their quality is of crucial importance. In the literature, configurations have been generated synthetically or by leveraging historical data. In this paper, we explore the use of off-the-shelf Large Language Models (LLMs) to generate good-quality traffic signal configurations to address a range of traffic signal optimisation problems. LLMs hold the promise of generating unusual yet effective configurations with minimal human effort.
For materials engineers, it is very important to select the best material to make an object. The criteria used for such selection may be classified in several categories, including the two following: criteria related to the durability of materials against degradations caused by environments and those expressing their impacts on the environment throughout their whole life cycle. Selecting a material is therefore a multi-criteria decision problem. Due to the diversity of criteria, material analysis is carried out by design offices using systems based on human knowledge when available and, otherwise, trial and error. Currently, the large amount of materials available and the number of properties make it necessary to investigate the use of generative AI for the selection of materials. However, the information provided by the available systems is insufficient to make the appropriate decision. Traditional operational research systems based on mathematical models, such as Formal Concept Analysis (FCA), provide a support for understanding the logic used to reach the solutions. They offer decision-makers a greater possibility of arguing their choice. Our work tackles this material selection decision problem using FCA. The results obtained after applying the FCA ConExp tool are analyzed in relation to the results obtained by the AI systems ChatGPT and the Fictiv-AI material selection assistant.
There are a variety of technologies needed to create a fully synthetic training experience from the “bottom up” (ie. “a fully created synthetic experience”) and from the “top down” (ie. created for the user who needs training). This paper outlines the technologies required with the note that many of them currently exist in research form; it describes the world soon to come to computer aided instruction.
Centralised Traffic Routing offers a global perspective of the entire region, enabling effective route finding for vehicles that navigate through the region. Recently, automated planning techniques were leveraged for centralised traffic routing. However, centralised routing techniques might generate different routes for vehicles even if they share the same origin and destination and arrive at a similar time. Also, centralised routing techniques are usually computationally demanding. In this paper, we analyse the outputs of a planning-based centralised traffic routing method in terms of the number of different routes per journey, differences between traveled distance, and travel time among these (different) routes. Then, we propose a method that leverages the found routes and their relative distribution to vehicles for an online centralised routing.
Generative AI tools tend to be used as if they were built to gather or confirm truthful information, as if they were knowledge-based systems. As such, there is a discrepancy between how generative AI (e.g. ChatGPT) is conceived and used by the general public, and what it really is and can accomplish. Given a lack of proper legal framework and the widespread usage of these tools, organizations have raised red flags and urged academic institutions to reflect on governance principles for the use of generative AI. In this paper, we present the principles adopted by an Institutional AI committee to guide usage of generative AI, as well as the theoretical and practical considerations motivating their introduction.
This study investigates whether a neural network can approximate human perception of proximity in the context of air traffic surveillance involving suspicious aircraft near regular and irregular airstrips. Using a DenseNet model trained on a synthetic dataset of graphical representations, the study evaluates the network's ability to classify visual proximity relationships without explicit distance computation. While the model achieved moderate performance (F1-Score of 78.4%), results were limited by overfitting and the low variability of the data. Larger datasets did not improve performance, suggesting the importance of visual diversity over quantity. These findings validate the feasibility of modeling human-like spatial reasoning through neural networks in controlled environments. The research establishes an experimental baseline for future studies involving more complex data and architectures, such as EfficientNet or Transformers, to further improve model generalization and practical applicability.
Coastal environments are dynamic and ecologically significant, yet monitoring across multiple sites and analysis remain challenging due to the lack of domain-specific datasets tailored to their unique features. General-purpose models, including those used for scene graph generation, often fail to capture the semantic details necessary for meaningful comparisons in this context. This paper outlines the process of creating a domain-specific dataset for coastal environments, focusing on the challenges posed by crowdsourced imagery, such as variability in image sizes, lighting conditions, and camera quality. By leveraging scene graph generation to capture semantic meaning, this research seeks to create a domain-specific dataset suitable for the comparison of coastal environments. This work demonstrates how domain-specific datasets can drive innovation in computer vision and semantic understanding, contributing to the broader field of artificial intelligence by bridging the gap between generalized tools and specialized applications. Ultimately, this effort lays the groundwork for future planned research to develop a pipeline capable of generating comparison metrics based on the semantic content of scenes. Using raw standardized images of coastal environments from the Coastie Initiative, this pipeline aims to go beyond superficial appearance comparisons, offering more meaningful analyses that could enhance our understanding and support conservation efforts.
Hierarchical planning is a knowledge-based approach to automated planning, where high-level (compound) tasks are decomposed into simpler tasks until primitive tasks – actions – forming the plan are obtained. Hierarchical Task Networks (HTN) is a modeling formalism for hierarchical planning, where task decompositions are described using decomposition methods with specific constraints. These decomposition methods are very close to rewriting (production) rules of context-free grammars, specifically when the domain is totally ordered. Recursive rules, when a task decomposes to it self, are generally allowed. Generating plans from left to right brings advantage of the first actions in the plan to be ready for execution even if the plan is not yet fully generated. However, left recursion goes against this idea as it tends to generate the plan from right (end) till left (beginning). The paper shows how to remove left recursion from HTN decomposition methods using the ideas from the Greibach Normal Form.
The adoption of large language models (LLMs) has facilitated significant advancements in natural language processing. In a short space of time, LLMs have permeated a wide array of disciplines including healthcare, finance, education, etc. However, in their native form, LLMs retain information in their parameters, which sometimes causes the underlying models to produce inaccurate results or hallucinations. To that end, Retrieval-Augmented Generation (RAG) has been proposed to address some of the challenges of LLMs by referencing an external knowledge base while formulating a response to queries. Still, traditional RAG fails to handle the complex structure of relationships among different entities in structured data such as knowledge graphs. GraphRAG, a successor of basic RAG, leverages structural information contained in graphs to enable more precise and comprehensive retrieval thereby facilitating more accurate, context-aware responses. GraphRAG has been applied in many domains, but its use in a cybersecurity context has not been widely explored. In this research, we propose a framework that applies GraphRAG to network security monitoring. By generating knowledge graphs from network logs, we provide LLMs with more structured data, backed by an ontology, that enables the models to perform high-level reasoning to answer questions regarding the security posture of an organization more accurately.
This study aims to use machine learning to find miniature excitatory postsynaptic currents (EPSCs) in neurons of a Drosophila to find behavior markers of a seizure. Using MATLAB, we are training a machine learning model on electrophysiological data to recognize patterns of post-synaptic events that show potential seizure activity. We have faced challenges applying this method and we are planning to present these in our poster. The results of this research may help develop a further understanding of seizure mechanisms in Drosophila that could translate into a more in-depth understanding for neurological disorders in humans.
Breast cancer remains one of the leading causes of mortality among women globally, and early detection is critical for improving survival rates. Breast MRI, the most sensitive imaging modality for detection, often involves manual review of numerous slices, which is time-intensive and prone to human error. Machine learning (ML) algorithms offer a transformative solution by automating this process, improving efficiency, and enhancing diagnostic accuracy. In this study, we propose a machine learning approach to enhance breast cancer prediction and diagnosis. We utilize a pre-trained multiscale vision transformer, Wave-ViT, to classify MRI slices as healthy or unhealthy. The model was trained and tested on MRI scans from 922 patients in the Duke Breast Cancer MRI dataset and independently validated on 143 patients from the MAMA-MIA dataset. To ensure high-quality data, both datasets were carefully curated to exclude noisy or mislabeled slices. The model's performance was evaluated using accuracy, F1-score, precision, recall, and confusion matrices under various experimental conditions. These included randomized training and testing splits using the Fisher-Yates shuffle, exploration of different Wave-ViT variants, and testing across multiple training set configurations. Our approach consistently demonstrated over 94\% accuracy on the external validation dataset, showcasing the potential of machine learning algorithms like Wave-ViT to reduce diagnostic workloads and improve breast cancer detection outcomes.
Prior research showed the instability of word embeddings. That is, the neighborhoods of word vectors differ depending on corpora and training methods. In this article we compute, using the correlation dimension algorithm, as well as a clustering dimension algorithm, the fractal dimensions of word embeddings, such as Glove vectors and FastText and CoNLL. We note the differences in fractal dimensions reported by us and prior work using other techniques, thereby showing the dependence of the geometry of word embeddings on algorithms used for computation. In addition, this article answers two questions about the dimension of the local manifold of word embeddings around polysemous words. Namely, the dimension is relatively small, i.e. 4 or less, and it does not differ from neighborhoods of non-polysemous words. We also observe in a few examples, that fractal dimensions are higher if we restrict ourselves to most frequent words, and hypothesize that this could be a more general pattern. This article also reviews recent publications in the area, including applications of fractals to an analysis of deep neural networks.
This poster presents an AI-powered intelligent tutoring system to enhance programming education through personalized and adaptive learning experiences. Using Generative AI, Machine Learning, and Natural Language Processing (NLP), our AI Tutor dynamically adjusts instructional content to match student needs, providing real-time feedback and a more interactive learning experience. The system integrates a chat-based interface that fosters seamless interaction, maintains learning context, and delivers personalized responses. Built on LangChain and Retrieval-Augmented Generation (RAG), the AI Tutor improves response accuracy and coherence, making it a powerful tool for programming education.
This study introduces SeniorSafeAI, an open-source chatbot trained on a curated dataset of cybersecurity Q&As to assist senior citizens in identifying and responding to cybercrimes. We trained and evaluated eight large language models using both quantitative metrics (F1, BertScore, n-gram overlap) and qualitative assessments (clarity, accuracy, relevance, and usefulness). While quantitative results indicate modest performance, qualitative evaluations of top models, including ChatGPT-4o and Qwen2.5 variants, reveal a notable discrepancy between numerical precision/recall and perceived response quality. Future work will focus on further finetuning with an expanded evaluation dataset and conducting user testing to improve usability and interface accessibility.
In this paper, we demonstrate an intelligent, personalized user interface for enterprise users who are inundated with humongous amount of business relevant news data. These users need interfaces that enable fast and interactive exploration of large and evolving corpora of news articles. Existing user interfaces either focus on relative frequency or co-occurrences of keywords (e.g., word clouds, clustering, Topic modeling) or organize the news using timeline representation and its recent variants such as Story Forest. All these representations ignore that there are different latent aspects or dimensions about a news article which are of particular interest from an enterprise user's perspective. In this paper, we propose NewsCube, a multi-dimensional hypercube representation that overcomes the limitations of existing interfaces to explore news corpora. Comparison of NewsCube based user-interface with other UI representations by experienced business users and senior management has confirmed the ease of navigation, high utility and on-demand focused interactivity of the proposed NewsCube representation. The accompanying video gives an overview of the NewsCube.
Efficient exploration remains a critical challenge in Reinforcement Learning (RL), significantly affecting sample efficiency. This paper demonstrates that biasing exploration towards state-action pairs with positive temporal difference error speeds up convergence and, in some challenging environments, has the potential to result in an improved policy. We show that this Positive Error Bias (PEB) method achieves statistically significant performance improvements across various tasks and estimators. Empirical results demonstrate PEB’s effectiveness in bandits, grid worlds, and classic control tasks with exact and approximate estimators. PEB is particularly effective when unbiased exploration struggles with policy discovery.
