Technology
LLMs for Cardiovascular Risk Prediction from Structured Clinical Data
This study compares LLMs and traditional ML for CAD prediction, finding Random Forest most accurate. However, LLMs offer privacy benefits by processing natural language narratives derived from structured data.
TCAR-Gen: Temporal Graph Retrieval with Evidence Fusion for Knowledge-Grounded Generation
TCAR-Gen enhances temporal graph retrieval for criminal case QA via evidence fusion, outperforming baselines on the Victorian Crime Diaries benchmark. However, output quality degrades significantly as model scale decreases.
A Multi-Domain Red Teaming Framework for Safety, Robustness, and Fairness Evaluation of Medical Large Language Models
This study introduces a multi-domain red teaming framework to evaluate 11 medical LLMs, revealing that aggregate metrics obscure critical safety and fairness failures. It advocates for hybrid evaluation strategies combining automated scoring with clinician oversight to ensure clinical reliability.
Update Opacity: Epistemic Accessibility and Governance Under AI System Change
This paper addresses "update opacity" in AI, proposing a governance framework integrating the EU AI Act and MLOps to track and disclose materially significant system changes. It ensures epistemic accessibility through threshold-based transparency and trustworthiness profiles.
Make Mechanistic Interpretability Auditable: A Call to Develop Guidelines via Continuous Collaborative Reviewing
Mechanistic interpretability lacks auditing standards, hindering safety-critical use. The authors propose a collaborative platform for continuous review and guideline development to ensure reliability.
Tracing GenAI Literacy: Uncovering Student-AI Interaction Patterns in Academic Writing through Epistemic Network Analysis
Using Epistemic Network Analysis on 162 students, this study reveals distinct GenAI interaction patterns: high-literacy users employ iterative refinement, while low-literacy users rely on direct prompts.
Improving Hospital Process Management through Process Mining: A Case Study on COVID-19 Clinical Pathways
This study applies process mining to COVID-19 data, revealing care variability and monitoring’s role. Findings support evidence-based hospital governance for triage, capacity planning, and patient transitions.
Beyond Tool Adoption: A Practical Five-Stage Developmental Continuum for AI Literacy in Higher Education
This study introduces a five-stage AI literacy continuum for higher education, validated through a case study at NC State University. It shifts focus from mere tool adoption to developmental critical engagement, offering a practical diagnostic roadmap for educators.
Beyond Categories of Caste: Examining Caste Bias and Morality in Text-to-Image AI Models
This study critiques static caste categorizations in Text-to-Image AI, advocating for a relational, anti-caste methodology to address Brahminical normativity and systemic bias.
Comprehensive AI governance requires addressing non-model gains
This paper argues that AI governance must expand beyond model-centric frameworks to address "non-model gains" like inference and systems improvements. It proposes multi-layered strategies and societal resilience to manage risks arising from these broader capability enhancements.
Algorithmic Authority and the Clinical Standard of Care
This paper argues that clinical AI acts as implicit regulation, requiring a new standard of care where the AI-physician dyad is jointly accountable. It advocates for integrating algorithmic accuracy with human expertise under strict governance.
When Jokes Cross the Line: Analyzing Regular Humor and Dark Humor in YouTube Shorts
This study introduces TwistedHumor, a dataset of 1,211 YouTube Shorts, analyzing how dark humor polarizes audiences compared to conventional jokes. It highlights the need for context-sensitive moderation to address the nuanced boundary between humor and harm.
From Human Videos to Robot Manipulation: A Survey on Scalable Vision-Language-Action Learning with Human-Centric Data
This survey reviews converting human videos into Vision-Language-Action models, categorizing extraction methods and addressing challenges like embodiment gaps and evaluation protocols.
Measuring and Mitigating Bias in Code Generated by Large Language Models
This study evaluates bias in GPT-4o and Gemini-generated code using CBS and ACR metrics. Despite testing four mitigation techniques, bias remains widespread, highlighting the need for more robust solutions.
Business Utility of Large Language Models as Exploratory Data Analysis Agents
This study evaluates LLMs as EDA agents, revealing most lack the reliability for autonomous business use. Only GPT-5.4 demonstrated sufficient quality and repeatability to achieve viable business utility.
Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks
This study identifies rare, synchronized spiking ensembles as key computational units in deep SNNs. These functional groups encode class information reliably, unlike random noise, revealing how learning shapes network structure.
Physics-Informed Neural Networks for Radial Consolidation of Combined Electroosmotic, Vacuum and Surcharge Preloading Considering Smear Effects
This study develops a physics-informed neural network framework to model radial consolidation with smear effects under combined preloading. The modified hard-constraint PINN demonstrates superior accuracy and robustness compared to standard architectures.
Beyond Text and Tables: Vision-Language Model Integration in ComProScanner for Extracting Materials Data from Scientific Figures with High Accuracy
ComProScanner integrates vision-language models to extract materials data from scientific figures, achieving 0.97 accuracy with Gemini-3-Flash. This multimodal platform automates composition-property database construction from literature.
CLSP-REQA: A Real-Time Quality-Aware Closed-Loop Seizure Prediction Framework with Mamba-BiLSTM and Confidence-Gated Intervention
CLSP-REQA integrates real-time EEG quality assessment with Mamba-BiLSTM to predict seizures robustly. It outperforms baselines on CHB-MIT and SIENA datasets, enabling reliable closed-loop neurostimulation.
Improved Belief-Attention in Vision Task
Belief2-Attention enhances visual recognition by retaining projected token correlations and introducing a $ZZ^T$ matrix for complex relationships. It outperforms standard attention in classification and segmentation.