863 papers
This paper demonstrates that hospital-level prevalence mismatch is a dominant, quantifiable cause of calibration failure in cross-institutional ICU mortality models, showing that a simple Bayesian correction using only local mortality rates can effectively restore calibration without requiring labeled target data.
This paper demonstrates that reasoning capabilities in foundation models exhibit layer-specific fragility patterns that vary across architectures, revealing that while uniform quantization suffices at higher bit-widths, a mixed-precision strategy guided by layer-localized fragility maps enables significant compression below four bits without sacrificing performance.
This paper introduces a lightweight, backbone-agnostic safety envelope that wraps frozen foundation models to provide provably bounded prediction deviations, calibrated confidence, and robust out-of-distribution refusal, thereby ensuring the reliability required for clinical dental diagnosis across diverse centers and architectures.
This paper introduces VozConsultAI, a provider-agnostic microservice architecture for automated SOAP note generation in Brazilian Portuguese telehealth, and presents the first comparative benchmark of open-weight large language models using a novel, clinically grounded evaluation metric that prioritizes safety and penalizes hallucinations.
This paper proposes a human-supervised meta-model framework that combines multiple XAI methods to generate more plausible and human-aligned explanations for Arabic sentiment analysis, achieving a significant F1 score improvement of 0.84 with CatBoost while trading off some faithfulness compared to previous ensemble approaches.
This paper proposes AM-RAF, a conceptual six-layer architecture that integrates agentic reasoning, retrieval-augmented generation, and uncertainty-gated deferral to create a trustworthy, privacy-preserving clinical decision support system specifically designed for the connectivity and equity challenges of resource-constrained health systems.
This paper introduces a unified three-axis observability model for quantum risk that evaluates assets based on algorithmic resistance, channel protection, and crypto-agility, supported by an open-source implementation (Sezar) and empirical validation across TLS, QKD, and agility scenarios to address the limitations of current single-axis discovery tools.
This paper proposes a mobility-driven task offloading framework for Vehicular Edge Computing that utilizes the Asynchronous Advantage Actor-Critic (A3C) deep reinforcement learning algorithm to optimize scheduling decisions, effectively minimizing latency and energy overhead while ensuring reliability in dynamic vehicular environments.
This paper proposes a novel method for selecting a small, performance-representative validation set from limited medical image data to improve model development accuracy, demonstrating through prostate cancer MRI segmentation tasks that this approach yields performance estimates significantly closer to true independent results than traditional random partitioning.
This paper proposes SDD-YOLO11n, a lightweight object detection model for traffic accident recognition that integrates a ShuffleNetV2 backbone, a DW_ADown module, and a DS_SENetV2 module to significantly reduce computational complexity while improving accuracy and real-time performance compared to the original YOLO11n and other mainstream models.