Research Papers

The stability of recursively trained large language models (LLMs) is a foundational problem for AI safety. Prevailing theory predicts model collapse, a progressive degradation when models are trained on their own output. We challenge this narrative by introducing a selective feedback mechanism. Contrary to expectation, instead of merely slowing decay, our experiments provide strong evidence that this pressure reverses it, inducing a statistically significant performance improvement in a Gemma 2B model on a complex summarization task. We name this phenomenon the Anti-Ouroboros Effect. We contrast this with a foundational experiment using a simple classifier, where the theoretical degenerative loop was validated, highlighting the unique dynamics of high-dimensional models. Our findings establish that systemic resilience can be an emergent property of LLMs under simple selection pressure, suggesting a powerful and scalable principle for developing safer and more robust AI systems. Across five generations, a quality-filtered condition improved by 6.6% in ROUGE-L F1 score, whereas an unfiltered control degraded by 3.5% and a random-filter control degraded by 4.2%

Predictive models are typically trained on historical data to predict future outcomes. While it is commonly assumed that training on more historical data would improve model performance and robustness, data distribution shifts over time may undermine these benefits. This study examines how expanding historical data training windows under covariate shifts (changes in feature distributions) and concept shifts (changes in feature-outcome relationships) affects the performance and algorithmic fairness of predictive models. First, we perform a simulation study to explore scenarios with varying degrees of covariate and concept shifts in training data. Absent distribution shifts, we observe performance gains from longer training windows though they reach a plateau quickly; in the presence of concept shift, performance may actually decline. Covariate shifts alone do not significantly affect model performance, but may complicate the impact of concept shifts. In terms of fairness, models produce more biased predictions when the magnitude of concept shifts differs across sociodemographic groups; for intersectional groups, these effects are more complex and not simply additive. Second, we conduct an empirical case study of student retention prediction, a common machine learning application in education, using 12 years of student records from 23 minority-serving community colleges in the United States. We find concept shifts to be a key contributor to performance degradation when expanding the training window. Moreover, model fairness is compromised when marginalized populations have distinct data distribution shift patterns from their peers. Overall, our findings caution against conventional wisdom that "more data is better" and underscore the importance of using historical data judiciously, especially when it may be subject to data distribution shifts, to improve model performance and fairness.

It is known that big data analytics and AI pose a threat to privacy, and that some of this is due to some kind of "black box problem" in AI. I explain how this becomes a problem in the context of justification for judgments and actions. Furthermore, I suggest distinguishing three kinds of opacity: 1) the subjects do not know what the system does ("shallow opacity"), 2) the analysts do not know what the system does ("standard black box opacity"), or 3) the analysts cannot possibly know what the system might do ("deep opacity"). If the agents, data subjects as well as analytics experts, operate under opacity, then these agents cannot provide justifications for judgments that are necessary to protect privacy, e.g., they cannot give "informed consent", or guarantee "anonymity". It follows from these points that agents in big data analytics and AI often cannot make the judgments needed to protect privacy. So I conclude that big data analytics makes the privacy problems worse and the remedies less effective. As a positive note, I provide a brief outlook on technical ways to handle this situation.

The use of inappropriate language -- such as outdated, exclusionary, or non-patient-centered terms -- medical instructional materials can significantly influence clinical training, patient interactions, and health outcomes. Despite their reputability, many materials developed over past decades contain examples now considered inappropriate by current medical standards. Given the volume of curricular content, manually identifying instances of inappropriate use of language (IUL) and its subcategories for systematic review is prohibitively costly and impractical. To address this challenge, we conduct a first-in-class evaluation of small language models (SLMs) fine-tuned on labeled data and pre-trained LLMs with in-context learning on a dataset containing approximately 500 documents and over 12,000 pages. For SLMs, we consider: (1) a general IUL classifier, (2) subcategory-specific binary classifiers, (3) a multilabel classifier, and (4) a two-stage hierarchical pipeline for general IUL detection followed by multilabel classification. For LLMs, we consider variations of prompts that include subcategory definitions and/or shots. We found that both LLama-3 8B and 70B, even with carefully curated shots, are largely outperformed by SLMs. While the multilabel classifier performs best on annotated data, supplementing training with unflagged excerpts as negative examples boosts the specific classifiers' AUC by up to 25%, making them most effective models for mitigating harmful language in medical curricula.

The development of many vision models mainly focuses on improving their performance using metrics such as accuracy, IoU, and mAP, with less attention to explainability due to the complexity of applying xAI methods to provide a meaningful explanation of trained models. Although many existing xAI methods aim to explain vision models sample-by-sample, methods explaining the general behavior of vision models, which can only be captured after running on a large dataset, are still underexplored. Furthermore, understanding the behavior of vision models on general images can be very important to prevent biased judgments and help identify the model's trends and patterns. With the application of Vision-Language Models, this paper proposes a pipeline to explain vision models at both the sample and dataset levels. The proposed pipeline can be used to discover failure cases and gain insights into vision models with minimal effort, thereby integrating vision model development with xAI analysis to advance image analysis.

Effective Cyber Threat Intelligence (CTI) relies upon accurately structured and semantically enriched information extracted from cybersecurity system logs. However, current methodologies often struggle to identify and interpret malicious events reliably and transparently, particularly in cases involving unstructured or ambiguous log entries. In this work, we propose a novel methodology that combines ontology-driven structured outputs with Large Language Models (LLMs), to build an Artificial Intelligence (AI) agent that improves the accuracy and explainability of information extraction from cybersecurity logs. Central to our approach is the integration of domain ontologies and SHACL-based constraints to guide the language model's output structure and enforce semantic validity over the resulting graph. Extracted information is organized into an ontology-enriched graph database, enabling future semantic analysis and querying. The design of our methodology is motivated by the analytical requirements associated with honeypot log data, which typically comprises predominantly malicious activity. While our case study illustrates the relevance of this scenario, the experimental evaluation is conducted using publicly available datasets. Results demonstrate that our method achieves higher accuracy in information extraction compared to traditional prompt-only approaches, with a deliberate focus on extraction quality rather than processing speed.

Large Language Models (LLMs) are widely used for downstream tasks such as tabular classification, where ensuring fairness in their outputs is critical for inclusivity, equal representation, and responsible AI deployment. This study introduces a novel approach to enhancing LLM performance and fairness through the concept of a dynamic validation set, which evolves alongside the test set, replacing the traditional static validation approach. We also propose an iterative algorithm, SMITE, to select optimal in-context examples, with each example set validated against its corresponding dynamic validation set. The in-context set with the lowest total error is used as the final demonstration set. Our experiments across four different LLMs show that our proposed techniques significantly improve both predictive accuracy and fairness compared to baseline methods. To our knowledge, this is the first study to apply dynamic validation in the context of in-context learning for LLMs.

Large Language Model (LLM)-enabled agents are rapidly emerging across a wide range of applications, but their deployment introduces vulnerabilities with security implications. While prior work has examined prompt-based attacks (e.g., prompt injection) and data-oriented threats (e.g., data exfiltration), time-of-check to time-of-use (TOCTOU) remain largely unexplored in this context. TOCTOU arises when an agent validates external state (e.g., a file or API response) that is later modified before use, enabling practical attacks such as malicious configuration swaps or payload injection. In this work, we present the first study of TOCTOU vulnerabilities in LLM-enabled agents. We introduce TOCTOU-Bench, a benchmark with 66 realistic user tasks designed to evaluate this class of vulnerabilities. As countermeasures, we adapt detection and mitigation techniques from systems security to this setting and propose prompt rewriting, state integrity monitoring, and tool-fusing. Our study highlights challenges unique to agentic workflows, where we achieve up to 25% detection accuracy using automated detection methods, a 3% decrease in vulnerable plan generation, and a 95% reduction in the attack window. When combining all three approaches, we reduce the TOCTOU vulnerabilities from an executed trajectory from 12% to 8%. Our findings open a new research direction at the intersection of AI safety and systems security.

A deep learning model is often considered a black-box model, as its internal workings tend to be opaque to the user. Because of the lack of transparency, it is challenging to understand the reasoning behind the model's predictions. Here, we present an approach to making a deep learning-based solar storm prediction model interpretable, where solar storms include solar flares and coronal mass ejections (CMEs). This deep learning model, built based on a long short-term memory (LSTM) network with an attention mechanism, aims to predict whether an active region (AR) on the Sun's surface that produces a flare within 24 hours will also produce a CME associated with the flare. The crux of our approach is to model data samples in an AR as time series and use the LSTM network to capture the temporal dynamics of the data samples. To make the model's predictions accountable and reliable, we leverage post hoc model-agnostic techniques, which help elucidate the factors contributing to the predicted output for an input sequence and provide insights into the model's behavior across multiple sequences within an AR. To our knowledge, this is the first time that interpretability has been added to an LSTM-based solar storm prediction model.

Conformalized Exceptional Model Mining: Telling Where Your Model Performs (Not) Well

Xin Du, Sikun Yang, Wouter Duivesteijn +1

Understanding the nuanced performance of machine learning models is essential for responsible deployment, especially in high-stakes domains like healthcare and finance. This paper introduces a novel framework, Conformalized Exceptional Model Mining, which combines the rigor of Conformal Prediction with the explanatory power of Exceptional Model Mining (EMM). The proposed framework identifies cohesive subgroups within data where model performance deviates exceptionally, highlighting regions of both high confidence and high uncertainty. We develop a new model class, mSMoPE (multiplex Soft Model Performance Evaluation), which quantifies uncertainty through conformal prediction's rigorous coverage guarantees. By defining a new quality measure, Relative Average Uncertainty Loss (RAUL), our framework isolates subgroups with exceptional performance patterns in multi-class classification and regression tasks. Experimental results across diverse datasets demonstrate the framework's effectiveness in uncovering interpretable subgroups that provide critical insights into model behavior. This work lays the groundwork for enhancing model interpretability and reliability, advancing the state-of-the-art in explainable AI and uncertainty quantification.

Recommender systems help users discover new content, but can also reinforce existing biases, leading to unfair exposure and reduced diversity. This paper introduces and investigates thematic bias in book recommendations, defined as a disproportionate favouring or neglect of certain book themes. We adopt a multi-stage bias evaluation framework using the Book-Crossing dataset to evaluate thematic bias in recommendations and its impact on different user groups. Our findings show that thematic bias originates from content imbalances and is amplified by user engagement patterns. By segmenting users based on their thematic preferences, we find that users with niche and long-tail interests receive less personalised recommendations, whereas users with diverse interests receive more consistent recommendations. These findings suggest that recommender systems should be carefully designed to accommodate a broader range of user interests. By contributing to the broader goal of responsible AI, this work also lays the groundwork for extending thematic bias analysis to other domains.

The heart's electrical activity, recorded through Electrocardiography (ECG), is essential for diagnosing various cardiovascular conditions. However, many existing ECG segmentation models rely on complex, multi-layered architectures such as BiLSTM, which are computationally intensive and inefficient. This study introduces a streamlined architecture that combines spectral analysis with probabilistic predictions for ECG signal segmentation. By replacing complex layers with simpler ones, the model effectively captures both temporal and spectral features of the P, QRS, and T waves. Additionally, an Explainable AI (XAI) approach is applied to enhance model interpretability by explaining how temporal and frequency-based features contribute to ECG segmentation. By incorporating principles from physics-based AI, this method provides a clear understanding of the decision-making process, ensuring reliability and transparency in ECG analysis. This approach achieves high segmentation accuracy: 97.00% for the QRS wave, 93.33% for the T wave, and 96.07% for the P wave. These results indicate that the simplified architecture not only improves computational efficiency but also provides precise segmentation, making it a practical and effective solution for heart signal monitoring.