3031 papers
This paper proposes CG-GNN, a robust graph neural network framework that mitigates label noise by integrating community-guided label refinement with edge pruning and a progressive contrastive learning scheme to outperform state-of-the-art methods in noisy and data-scarce scenarios.
This study demonstrates that converting SHAP-based feature importance scores into direct managerial actions without accounting for intervention costs, causal validity, and downside risks creates significant accountability hazards, revealing that explainability alone is insufficient for responsible AI-driven decision-making.
This survey systematically reviews poisoning attacks and defenses in the era of Large Language Models by proposing a comprehensive taxonomy that categorizes threats into weight and context poisoning, analyzing representative techniques and defense strategies, and outlining future research directions for securing compound AI systems.
FedDAAW is a lightweight, privacy-preserving framework for heterogeneous federated learning that employs dynamic client selection and accuracy-adaptive aggregation to significantly improve convergence efficiency and performance across diverse tasks (classification, sentiment analysis, and regression) under non-IID settings without requiring gradient sharing or auxiliary data.
This bibliometric analysis of 1,225 publications from 2000 to 2024 reveals the evolution of robotic sports research into a mature, globally collaborative field driven by key institutions and a thematic shift from foundational robot soccer to advanced AI applications like deep learning and human–robot interaction.
This paper introduces the Composite Tasks Challenge (CTC), a new benchmark suite designed to rigorously evaluate division of labor and cooperation in multi-agent reinforcement learning, revealing that current state-of-the-art methods fail to solve these tasks while demonstrating that a solvable yet challenging testbed is essential for advancing the field.
SousCoder is a local-first, proactive developer assistant that monitors cross-application behavioral signals across VS Code, Chrome, and the terminal to detect developer confusion and intervene with context-aware suggestions, achieving low latency and high relevance in field evaluations while extending prior work with persistent memory and multi-source perception.
This study evaluates a large language model pipeline for extracting 14 actuarial variables from workers' compensation claims, revealing moderate overall agreement with human reviewers (quadratic weighted kappa of 0.53) and significant variability across dimensions, which underscores the need for calibration and phased deployment before formal validation.
This study demonstrates that a transformer-based deep learning model outperforms classical machine learning algorithms in predicting electrical grid stability, achieving high accuracy in both binary classification and regression tasks while supporting global sustainability goals for resilient smart grids.
This study identifies that lower lesion-background contrast on darker skin creates a structural class overlap causing AI dermatological models to systematically over-predict malignancy and generate excess referrals for darker-skinned patients, a fairness bottleneck that persists even after correcting for confounding class distributions and is best mitigated by per-tone class balancing rather than tone conditioning.