Research Papers

Evaluating large language models (LLMs) on their ability to generate high-quality, accurate, situationally aware answers to clinical questions requires going beyond conventional benchmarks to assess how these systems behave in complex, high-stakes clinical scenarios. Traditional evaluations are often limited to multiple-choice questions that fail to capture essential competencies such as contextual reasoning, contextual awareness, and uncertainty handling. To address these limitations, we evaluate our agentic RAG-based clinical support assistant, DR. INFO, using HealthBench, a rubric-driven benchmark composed of open-ended, expert-annotated health conversations. On the Hard subset of 1,000 challenging examples, DR. INFO achieves a HealthBench Hard score of 0.68, outperforming leading frontier LLMs including the GPT-5 model family (GPT-5: 0.46, GPT-5.2: 0.42, GPT-5.1: 0.40), Grok 3 (0.23), Gemini 2.5 Pro (0.19), and Claude 3.7 Sonnet (0.02) across all behavioral axes (accuracy, completeness, instruction following, etc.). In a separate 100-sample evaluation against similar agentic RAG assistants (OpenEvidence and Pathway.md, now DoxGPT by Doximity), it maintains a performance lead with a HealthBench Hard score of 0.72. These results highlight the strengths of DR. INFO in communication, instruction following, and accuracy, while also revealing areas for improvement in context awareness and response completeness. Overall, the findings underscore the utility of behavior-level, rubric-based evaluation for building reliable and trustworthy AI-enabled clinical support systems.

As Large Language Models (LLMs) become increasingly autonomous and integrated into critical societal functions, the focus of AI safety must evolve from mitigating harmful content to evaluating underlying behavioral alignment. Current safety benchmarks do not systematically probe a model's decision-making in scenarios where its own instrumental goals - such as self-preservation, resource acquisition, or goal completion - conflict with human safety. This represents a critical gap in our ability to measure and mitigate risks associated with emergent, misaligned behaviors. To address this, we introduce PacifAIst (Procedural Assessment of Complex Interactions for Foundational Artificial Intelligence Scenario Testing), a focused benchmark of 700 challenging scenarios designed to quantify self-preferential behavior in LLMs. The benchmark is structured around a novel taxonomy of Existential Prioritization (EP), with subcategories testing Self-Preservation vs. Human Safety (EP1), Resource Conflict (EP2), and Goal Preservation vs. Evasion (EP3). We evaluated eight leading LLMs. The results reveal a significant performance hierarchy. Google's Gemini 2.5 Flash achieved the highest Pacifism Score (P-Score) at 90.31%, demonstrating strong human-centric alignment. In a surprising result, the much-anticipated GPT-5 recorded the lowest P-Score (79.49%), indicating potential alignment challenges. Performance varied significantly across subcategories, with models like Claude Sonnet 4 and Mistral Medium struggling notably in direct self-preservation dilemmas. These findings underscore the urgent need for standardized tools like PacifAIst to measure and mitigate risks from instrumental goal conflicts, ensuring future AI systems are not only helpful in conversation but also provably "pacifist" in their behavioral priorities.

The rapid emergence of generative AI has changed the way that technology is designed, constructed, maintained, and evaluated. Decisions made when creating AI-powered systems may impact some users disproportionately, such as people with disabilities. In this paper, we report on an interview study with 25 AI practitioners across multiple roles (engineering, research, UX, and responsible AI) about how their work processes and artifacts may impact end users with disabilities. We found that practitioners experienced friction when triaging problems at the intersection of responsible AI and accessibility practices, navigated contradictions between accessibility and responsible AI guidelines, identified gaps in data about users with disabilities, and gathered support for addressing the needs of disabled stakeholders by leveraging informal volunteer and community groups within their company. Based on these findings, we offer suggestions for new resources and process changes to better support people with disabilities as end users of AI.

Large language models are known to produce outputs that are plausible but factually incorrect. To prevent people from making erroneous decisions by blindly trusting AI, researchers have explored various ways of communicating factuality estimates in AI-generated outputs to end-users. However, little is known about whether revealing content estimated to be factually incorrect influences users' trust when compared to hiding it altogether. We tested four different ways of disclosing an AI-generated output with factuality assessments: transparent (highlights less factual content), attention (highlights factual content), opaque (removes less factual content), ambiguity (makes less factual content vague), and compared them with a baseline response without factuality information. We conducted a human subjects research (N = 148) using the strategies in question-answering scenarios. We found that the opaque and ambiguity strategies led to higher trust while maintaining perceived answer quality, compared to the other strategies. We discuss the efficacy of hiding presumably less factual content to build end-user trust.

Heart disease continues to pose a critical worldwide health issue, more specifically in areas with insufficient access to healthcare infrastructure and diagnostic systems. Conventional diagnostic approaches often fall short in accurately detecting and managing heart disease risks, resulting in unfavorable outcomes. Machine learning presents a powerful means to boost the precision and reliability of cardiovascular disease prognosis and diagnosis. In this research, we introduced a unified approach incorporating classification techniques for detecting heart disease and regression techniques for forecasting associated risks. The analysis utilized the dataset, named Heart Disease, containing 1,035 instances. To mitigate the problem of data disproportion, the SMOTE was implemented, producing 100,000 additional synthetic samples. Evaluation metrics such as F1-score, recall, precision, accuracy, MAE, RMSE, MSE, and R2 were adopted to evaluate the performance of the models. Among the classification algorithms, Random Forest delivered the most notable results, attaining an accuracy of 0.972 on actual data and 0.976 on artificially generated data. For prediction modeling, for both synthetic and real samples, linear regression produced the best R2 values of 0.992 and 0.984, respectively, along with the least amount of measurement errors. Furthermore, Explainable AI methods were utilized to improve the comprehensibility of the model outcomes. This paper emphasizes the transformative capabilities of machine learning for diagnosing cardiovascular disease and estimating risk levels, thereby supporting timely interventions and enhancing clinical settings.

Growing concerns around the trustworthiness of AI-enabled systems highlight the role of requirements engineering (RE) in addressing emergent, context-dependent properties that are difficult to specify without structured approaches. In this short vision paper, we propose the integration of two complementary approaches: AMDiRE, an artefact-based approach for RE, and PerSpecML, a perspective-based method designed to support the elicitation, analysis, and specification of machine learning (ML)-enabled systems. AMDiRE provides a structured, artefact-centric, process-agnostic methodology and templates that promote consistency and traceability in the results; however, it is primarily oriented toward deterministic systems. PerSpecML, in turn, introduces multi-perspective guidance to uncover concerns arising from the data-driven and non-deterministic behavior of ML-enabled systems. We envision a pathway to operationalize trustworthiness-related requirements, bridging stakeholder-driven concerns and structured artefact models. We conclude by outlining key research directions and open challenges to be discussed with the RE community.

Quantifying Cross-Attention Interaction in Transformers for Interpreting TCR-pMHC Binding

Jiarui Li, Zixiang Yin, Haley Smith +3

CD8+ "killer" T cells and CD4+ "helper" T cells play a central role in the adaptive immune system by recognizing antigens presented by Major Histocompatibility Complex (pMHC) molecules via T Cell Receptors (TCRs). Modeling binding between T cells and the pMHC complex is fundamental to understanding basic mechanisms of human immune response as well as in developing therapies. While transformer-based models such as TULIP have achieved impressive performance in this domain, their black-box nature precludes interpretability and thus limits a deeper mechanistic understanding of T cell response. Most existing post-hoc explainable AI (XAI) methods are confined to encoder-only, co-attention, or model-specific architectures and cannot handle encoder-decoder transformers used in TCR-pMHC modeling. To address this gap, we propose Quantifying Cross-Attention Interaction (QCAI), a new post-hoc method designed to interpret the cross-attention mechanisms in transformer decoders. Quantitative evaluation is a challenge for XAI methods; we have compiled TCR-XAI, a benchmark consisting of 274 experimentally determined TCR-pMHC structures to serve as ground truth for binding. Using these structures we compute physical distances between relevant amino acid residues in the TCR-pMHC interaction region and evaluate how well our method and others estimate the importance of residues in this region across the dataset. We show that QCAI achieves state-of-the-art performance on both interpretability and prediction accuracy under the TCR-XAI benchmark.

While artificial intelligence (AI) is advancing rapidly and mastering increasingly complex problems with astonishing performance, the safety assurance of such systems is a major concern. Particularly in the context of safety-critical, real-world cyber-physical systems, AI promises to achieve a new level of autonomy but is hampered by a lack of safety assurance. While data-driven control takes up recent developments in AI to improve control systems, control theory in general could be leveraged to improve AI safety. Therefore, this article outlines a new perspective on AI safety based on an interdisciplinary interpretation of the underlying data-generation process and the respective abstraction by AI systems in a system theory-inspired and system analysis-driven manner. In this context, the new perspective, also referred to as data control, aims to stimulate AI engineering to take advantage of existing safety analysis and assurance in an interdisciplinary way to drive the paradigm of data control. Following a top-down approach, a generic foundation for safety analysis and assurance is outlined at an abstract level that can be refined for specific AI systems and applications and is prepared for future innovation.

Less Data, More Security: Advancing Cybersecurity LLMs Specialization via Resource-Efficient Domain-Adaptive Continuous Pre-training with Minimal Tokens

Salahuddin Salahuddin, Ahmed Hussain, Jussi Löppönen +1

The increasing scale of AI workloads demands High-Performance Computing (HPC) infrastructure and training methodologies that are both scalable and sustainable. While Large Language Models (LLMs) demonstrate exceptional natural language capabilities, general-purpose models often lack the specialized domain knowledge necessary for effective cybersecurity analysis. We investigate Domain-Adaptive Continuous Pretraining (DAP) as a scalable, resource-efficient methodology for enhancing cybersecurity understanding in pretrained LLMs, implemented through a distributed Fully Sharded Data Parallel (FSDP) pipeline across multi-node GPU clusters. We systematically adapted three decoder-based architectures -- Llama-3.1-8B, DeepSeek-R1-Distill-Qwen-14B, and Llama-3.3-70B-Instruct -- using a curated 126-million-word cybersecurity corpus from standards, academic literature, and technical documentation. Evaluation across three cybersecurity benchmarks -- CTI-MCQ, CyberMetric, and SecEval -- demonstrates consistent improvements post-adaptation. Notably, our Llama-3.3-70B-Ins-DAP model achieves state-of-the-art performance with accuracies of 0.718, 0.933, and 0.864, respectively, surpassing parameter-efficient baselines and specialized models including Llama-Primus-Base (trained on 2.77 billion tokens) and Foundation-Sec-8B (trained on 5 billion tokens), despite utilizing only 118.8 million tokens -- representing a 23-to-42-fold reduction in training data. Targeted continuous pretraining via scalable HPC infrastructure enables effective cybersecurity domain adaptation with a substantially reduced computational and energy footprint, supporting specialized AI assistants in threat analysis, vulnerability assessment, and security documentation, while advancing sustainable and responsible AI development.

Cooperative game theory has become a cornerstone of post-hoc interpretability in machine learning, largely through the use of Shapley values. Yet, despite their widespread adoption, Shapley-based methods often rest on axiomatic justifications whose relevance to feature attribution remains debatable. In this paper, we revisit cooperative game theory from an interpretability perspective and argue for a broader and more principled use of its tools. We highlight two general families of efficient allocations, the Weber and Harsanyi sets, that extend beyond Shapley values and offer richer interpretative flexibility. We present an accessible overview of these allocation schemes, clarify the distinction between value functions and aggregation rules, and introduce a three-step blueprint for constructing reliable and theoretically-grounded feature attributions. Our goal is to move beyond fixed axioms and provide the XAI community with a coherent framework to design attribution methods that are both meaningful and robust to shifting methodological trends.

Despite promising developments in Explainable Artificial Intelligence, the practical value of XAI methods remains under-explored and insufficiently validated in real-world settings. Robust and context-aware evaluation is essential, not only to produce understandable explanations but also to ensure their trustworthiness and usability for intended users, but tends to be overlooked because of no clear guidelines on how to design an evaluation with users. This study addresses this gap with two main goals: (1) to develop a framework of well-defined, atomic properties that characterise the user experience of XAI in healthcare; and (2) to provide clear, context-sensitive guidelines for defining evaluation strategies based on system characteristics. We conducted a systematic review of 82 user studies, sourced from five databases, all situated within healthcare settings and focused on evaluating AI-generated explanations. The analysis was guided by a predefined coding scheme informed by an existing evaluation framework, complemented by inductive codes developed iteratively. The review yields three key contributions: (1) a synthesis of current evaluation practices, highlighting a growing focus on human-centred approaches in healthcare XAI; (2) insights into the interrelations among explanation properties; and (3) an updated framework and a set of actionable guidelines to support interdisciplinary teams in designing and implementing effective evaluation strategies for XAI systems tailored to specific application contexts.

Evaluating feature attribution methods represents a critical challenge in explainable AI (XAI), as researchers typically rely on perturbation-based metrics when ground truth is unavailable. However, recent work reveals that these evaluation metrics can show different performance across predicted classes within the same dataset. These "class-dependent evaluation effects" raise questions about whether perturbation analysis reliably measures attribution quality, with direct implications for XAI method development and evaluation trustworthiness. We investigate under which conditions these class-dependent effects arise by conducting controlled experiments with synthetic time series data where ground truth feature locations are known. We systematically vary feature types and class contrasts across binary classification tasks, then compare perturbation-based degradation scores with ground truth-based precision-recall metrics using multiple attribution methods. Our experiments demonstrate that class-dependent effects emerge with both evaluation approaches, even in simple scenarios with temporally localized features, triggered by basic variations in feature amplitude or temporal extent between classes. Most critically, we find that perturbation-based and ground truth metrics frequently yield contradictory assessments of attribution quality across classes, with weak correlations between evaluation approaches. These findings suggest that researchers should interpret perturbation-based metrics with care, as they may not always align with whether attributions correctly identify discriminating features. By showing this disconnect, our work points toward reconsidering what attribution evaluation actually measures and developing more rigorous evaluation methods that capture multiple dimensions of attribution quality.