Research Papers

As Large Language Models (LLMs) continue to be increasingly applied across various domains, their widespread adoption necessitates rigorous monitoring to prevent unintended negative consequences and ensure robustness. Furthermore, LLMs must be designed to align with human values, like preventing harmful content and ensuring responsible usage. The current automated systems and solutions for monitoring LLMs in production are primarily centered on LLM-specific concerns like hallucination etc, with little consideration given to the requirements of specific use-cases and user preferences. This paper introduces GAF-Guard, a novel agentic framework for LLM governance that places the user, the use-case, and the model itself at the center. The framework is designed to detect and monitor risks associated with the deployment of LLM based applications. The approach models autonomous agents that identify risks, activate risk detection tools, within specific use-cases and facilitate continuous monitoring and reporting to enhance AI safety, and user expectations. The code is available at https://github.com/IBM/risk-atlas-nexus-demos/tree/main/gaf-guard.

Every publicly traded U.S. company files an annual 10-K report containing critical insights into financial health and risk. We propose Tiny eXplainable Risk Assessor (TinyXRA), a lightweight and explainable transformer-based model that automatically assesses company risk from these reports. Unlike prior work that relies solely on the standard deviation of excess returns (adjusted for the Fama-French model), which indiscriminately penalizes both upside and downside risk, TinyXRA incorporates skewness, kurtosis, and the Sortino ratio for more comprehensive risk assessment. We leverage TinyBERT as our encoder to efficiently process lengthy financial documents, coupled with a novel dynamic, attention-based word cloud mechanism that provides intuitive risk visualization while filtering irrelevant terms. This lightweight design ensures scalable deployment across diverse computing environments with real-time processing capabilities for thousands of financial documents which is essential for production systems with constrained computational resources. We employ triplet loss for risk quartile classification, improving over pairwise loss approaches in existing literature by capturing both the direction and magnitude of risk differences. Our TinyXRA achieves state-of-the-art predictive accuracy across seven test years on a dataset spanning 2013-2024, while providing transparent and interpretable risk assessments. We conduct comprehensive ablation studies to evaluate our contributions and assess model explanations both quantitatively by systematically removing highly attended words and sentences, and qualitatively by examining explanation coherence. The paper concludes with findings, practical implications, limitations, and future research directions. Our code is available at https://github.com/Chen-XueWen/TinyXRA.

Explainable AI for Radar Resource Management: Modified LIME in Deep Reinforcement Learning

Ziyang Lu, M. Cenk Gursoy, Chilukuri K. Mohan +1

Deep reinforcement learning has been extensively studied in decision-making processes and has demonstrated superior performance over conventional approaches in various fields, including radar resource management (RRM). However, a notable limitation of neural networks is their ``black box" nature and recent research work has increasingly focused on explainable AI (XAI) techniques to describe the rationale behind neural network decisions. One promising XAI method is local interpretable model-agnostic explanations (LIME). However, the sampling process in LIME ignores the correlations between features. In this paper, we propose a modified LIME approach that integrates deep learning (DL) into the sampling process, which we refer to as DL-LIME. We employ DL-LIME within deep reinforcement learning for radar resource management. Numerical results show that DL-LIME outperforms conventional LIME in terms of both fidelity and task performance, demonstrating superior performance with both metrics. DL-LIME also provides insights on which factors are more important in decision making for radar resource management.

Physics-informed neural networks (PINNs) offer a powerful approach to solving partial differential equations (PDEs), which are ubiquitous in the quantitative sciences. Applied to both forward and inverse problems across various scientific domains, PINNs have recently emerged as a valuable tool in the field of scientific machine learning. A key aspect of their training is that the data -- spatio-temporal points sampled from the PDE's input domain -- are readily available. Influence functions, a tool from the field of explainable AI (XAI), approximate the effect of individual training points on the model, enhancing interpretability. In the present work, we explore the application of influence function-based sampling approaches for the training data. Our results indicate that such targeted resampling based on data attribution methods has the potential to enhance prediction accuracy in physics-informed neural networks, demonstrating a practical application of an XAI method in PINN training.

Predictive Process Monitoring (PPM) often uses deep learning models to predict the future behavior of ongoing processes, such as predicting process outcomes. While these models achieve high accuracy, their lack of interpretability undermines user trust and adoption. Explainable AI (XAI) aims to address this challenge by providing the reasoning behind the predictions. However, current evaluations of XAI in PPM focus primarily on functional metrics (such as fidelity), overlooking user-centered aspects such as their effect on task performance and decision-making. This study investigates the effects of explanation styles (feature importance, rule-based, and counterfactual) and perceived AI accuracy (low or high) on decision-making in PPM. We conducted a decision-making experiment, where users were presented with the AI predictions, perceived accuracy levels, and explanations of different styles. Users' decisions were measured both before and after receiving explanations, allowing the assessment of objective metrics (Task Performance and Agreement) and subjective metrics (Decision Confidence). Our findings show that perceived accuracy and explanation style have a significant effect.

Fairness is a critical component of Trustworthy AI. In this paper, we focus on Machine Learning (ML) and the performance of model predictions when dealing with skin color. Unlike other sensitive attributes, the nature of skin color differs significantly. In computer vision, skin color is represented as tensor data rather than categorical values or single numerical points. However, much of the research on fairness across sensitive groups has focused on categorical features such as gender and race. This paper introduces a new technique for evaluating fairness in ML for image classification tasks, specifically without the use of annotation. To address the limitations of prior work, we handle tensor data, like skin color, without classifying it rigidly. Instead, we convert it into probability distributions and apply statistical distance measures. This novel approach allows us to capture fine-grained nuances in fairness both within and across what would traditionally be considered distinct groups. Additionally, we propose an innovative training method to mitigate the latent biases present in conventional skin tone categorization. This method leverages color distance estimates calculated through Bayesian regression with polynomial functions, ensuring a more nuanced and equitable treatment of skin color in ML models.

VARSHAP: Addressing Global Dependency Problems in Explainable AI with Variance-Based Local Feature Attribution

Mateusz Gajewski, Mikołaj Morzy, Adam Karczmarz +1

Existing feature attribution methods like SHAP often suffer from global dependence, failing to capture true local model behavior. This paper introduces VARSHAP, a novel model-agnostic local feature attribution method which uses the reduction of prediction variance as the key importance metric of features. Building upon Shapley value framework, VARSHAP satisfies the key Shapley axioms, but, unlike SHAP, is resilient to global data distribution shifts. Experiments on synthetic and real-world datasets demonstrate that VARSHAP outperforms popular methods such as KernelSHAP or LIME, both quantitatively and qualitatively.

Artificial intelligence has transformed numerous industries, from healthcare to finance, enhancing decision-making through automated systems. However, the reliability of these systems is mainly dependent on the quality of the underlying datasets, raising ongoing concerns about transparency, accountability, and potential biases. This paper introduces a scorecard designed to evaluate the development of AI datasets, focusing on five key areas from the system card framework data development life cycle: data dictionary, collection process, composition, motivation, and pre-processing. The method follows a structured approach, using an intake form and scoring criteria to assess the quality and completeness of the data set. Applied to four diverse datasets, the methodology reveals strengths and improvement areas. The results are compiled using a scoring system that provides tailored recommendations to enhance the transparency and integrity of the data set. The scorecard addresses technical and ethical aspects, offering a holistic evaluation of data practices. This approach aims to improve the quality of the data set. It offers practical guidance to curators and researchers in developing responsible AI systems, ensuring fairness and accountability in decision support systems.

Lifelong learning algorithms enable models to incrementally acquire new knowledge without forgetting previously learned information. Contrarily, the field of machine unlearning focuses on explicitly forgetting certain previous knowledge from pretrained models when requested, in order to comply with data privacy regulations on the right-to-be-forgotten. Enabling efficient lifelong learning with the capability to selectively unlearn sensitive information from models presents a critical and largely unaddressed challenge with contradicting objectives. We address this problem from the perspective of simultaneously preventing catastrophic forgetting and allowing forward knowledge transfer during task-incremental learning, while ensuring exact task unlearning and minimizing memory requirements, based on a single neural network model to be adapted. Our proposed solution, privacy-aware lifelong learning (PALL), involves optimization of task-specific sparse subnetworks with parameter sharing within a single architecture. We additionally utilize an episodic memory rehearsal mechanism to facilitate exact unlearning without performance degradations. We empirically demonstrate the scalability of PALL across various architectures in image classification, and provide a state-of-the-art solution that uniquely integrates lifelong learning and privacy-aware unlearning mechanisms for responsible AI applications.

Responsible AI (RAI) efforts increasingly emphasize the importance of addressing potential harms early in the AI development lifecycle through social-technical lenses. However, in cross-functional industry teams, this work is often stalled by a persistent coordination challenge: how technical roles hand off technical intent, how teams establish shared structures for collaboration, and how non-technical roles are supported in systematically evaluating harms. Through literature review and a semi-structured interview study with 8 practitioners, we unpack how this challenge manifests---technical design choices are rarely handed off in ways that support meaningful engagement by non-technical roles; collaborative workflows lack shared, visual structures to support mutual understanding; and non-technical practitioners are left without scaffolds for systematic harm evaluation. Existing tools like JIRA or Google Docs, while useful for product tracking, are ill-suited for supporting joint harm identification across roles, often requiring significant extra effort to align understanding. To address this, we developed AI LEGO, a web-based prototype that operationalizes the boundary object theory to support cross-functional AI practitioners in effectively facilitating knowledge handoff and identifying harmful design choices in the early design stages. Technical roles use interactive blocks to draft development plans, while non-technical roles engage with those blocks through stage-specific checklists and LLM-driven persona simulations to surface potential harms. In a study with 18 cross-functional practitioners, AI LEGO increased the volume and likelihood of harms identified compared to baseline worksheets. Participants found that its modular structure and persona prompts made harm identification more accessible, fostering clearer and more collaborative RAI practices in early design.

Large Language Models (LLMs) are increasingly used not only to generate text but also to evaluate it, raising urgent questions about whether their judgments are consistent, unbiased, and robust to framing effects. In this study, we systematically examine inter- and intra-model agreement across four state-of-the-art LLMs (OpenAI o3-mini, Deepseek Reasoner, xAI Grok 2, and Mistral) tasked with evaluating 4,800 narrative statements on 24 different topics of social, political, and public health relevance, for a total of 192,000 assessments. We manipulate the disclosed source of each statement to assess how attribution to either another LLM or a human author of specified nationality affects evaluation outcomes. We find that, in the blind condition, different LLMs display a remarkably high degree of inter- and intra-model agreement across topics. However, this alignment breaks down when source framing is introduced. Here we show that attributing statements to Chinese individuals systematically lowers agreement scores across all models, and in particular for Deepseek Reasoner. Our findings reveal that framing effects can deeply affect text evaluation, with significant implications for the integrity, neutrality, and fairness of LLM-mediated information systems.

Assessments of trustworthiness have become a cornerstone of responsible AI development. Especially in high-stakes fields like healthcare, aligning technical, evidence-based, and ethical practices with forthcoming legal requirements is increasingly urgent. We argue that developers and deployers of AI systems for the medical domain should be proactive and take steps to progressively ensure that such systems, both those currently in use and those being developed or planned, respect the requirements of the AI Act, which has come into force in August 2024. This is necessary if full and effective compliance is to be ensured when the most relevant provisions of the Act become effective (August 2026). The engagement with the AI Act cannot be viewed as a formalistic exercise. Compliance with the AI Act needs to be carried out through the proactive commitment to the ethical principles of trustworthy AI. These principles provide the background for the Act, which mentions them several times and connects them to the protection of public interest. They can be used to interpret and apply the Act's provisions and to identify good practices, increasing the validity and sustainability of AI systems over time.