This article introduces GALDS, a novel model that streamlines the simulation of material transport in neuronal networks using a graph autoencoder. By optimizing data representation and reducing computational demands, GALDS significantly improves accuracy and processing speed, achieving impressive results on unseen neural geometries.

This article presents a domain-adaptive small language model designed to enhance the prediction of structured tax codes, crucial for tax compliance. The proposed encoder-decoder architecture effectively addresses hierarchical dependencies in tax code generation, outperforming traditional classifiers and offering potential applications across various tax systems.

This article introduces SiGMoID, a simulation-based generative model that enhances the inference of nonlinear dynamic systems using noisy and incomplete data. By combining physics-informed neural networks and Wasserstein adversarial networks, this approach improves parameter estimation and system understanding across various applications.

This article explores a novel approach to modeling outbound inventory flows and shipping costs in complex fulfillment systems. By framing the problem as a probabilistic forecasting task, it enhances reinforcement learning applications in regional inventory management, with preliminary results showing promising accuracy in real-world settings.

This study reviews research on artificial intelligence in journalism from 2010 to 2025, highlighting trends, ethical concerns, and regional disparities. By analyzing 72 articles, it reveals growing interest in AI’s impact on automation, misinformation, and governance, while also noting challenges like bias and transparency.

This study presents a modernized CNN-based model, KAI-a, for global weather forecasting that achieves competitive accuracy with reduced computational demands. Incorporating innovative design enhancements, it effectively captures extreme weather events while maintaining lightweight performance, making it a promising alternative to traditional models in medium-range forecasting.

This research tackles Timescale Dependent Label Inconsistency in EEG emotion recognition by introducing two innovative regularization strategies. The methods enhance model performance and explainability, validated through experiments on popular datasets. Results show significant improvements over existing models, emphasizing a balance between interpretability and predictive accuracy.

This paper explores a new approach to stochastic optimization, combining non-smooth outer functions with smooth inner ones. Through the use of smoothing functions, the authors present a method that enhances convergence rates in various settings, demonstrating competitive results for specific problem classes in machine learning.

LiLM-RDB-SFC presents an innovative method for optimizing Service Function Chain (SFC) provisioning using a combination of Lightweight Language Models and Relational Databases in a Deep Reinforcement Learning framework. The approach enhances decision-making in Software-Defined Networking by improving accuracy and reducing processing time significantly.

This article explores the concept of class-difficulty separability in machine learning, highlighting how data difficulty varies across classes. The authors present a new method for data pruning that enhances model performance, particularly in complex domains like network intrusion detection and medical imaging, achieving state-of-the-art results.

This study evaluates the effectiveness of a Large Language Model (LLM)-based multi-agent system for therapy recommendations in chronic patients with multimorbidity. It compares the performance of multi-agent systems to single-agent approaches, revealing strengths and limitations in providing safe, comprehensive treatment advice.

HanjaBridge introduces a new technique for improving Korean language understanding in large language models by addressing semantic ambiguity. By integrating Hanja characters into pre-training, this method facilitates better contextual learning and enhances model performance, achieving significant improvements on benchmarks while ensuring practical efficiency.

Protein language models offer significant benefits in generating sequences, but they carry risks of harmful outcomes. This article presents a Knowledge-guided Preference Optimization framework to enhance the safety of protein generation while preserving functionality, addressing crucial biosafety and ethical challenges in biotechnology.