This article presents a Bayesian framework for modeling Linear Time-Variant systems using neural networks and Gaussian processes. By treating impulse responses as stochastic processes, it allows for better uncertainty quantification and improved data efficiency in system identification, even from single noisy observations.

This paper presents a reinforcement learning approach for optimizing resource management in microservice systems. By dynamically adjusting resource allocation, the method improves response times, throughput, and energy efficiency, outperforming traditional static models in various load scenarios and resource conditions.

This article presents VAR-MATH, a new framework designed to evaluate the mathematical reasoning capabilities of large language models (LLMs). By using symbolic benchmarks, VAR-MATH addresses issues like benchmark contamination and evaluation fragility, revealing that many current models may rely on superficial heuristics rather than true reasoning skills.

This article presents VIDAR, a two-stage framework for bimanual robotic manipulation that utilizes video diffusion models and masked inverse dynamics for action prediction. With minimal human demonstration data, VIDAR achieves strong generalization to new tasks and environments, showcasing the promise of video-based approaches in robotics.

This paper presents a new dataset of 23,326 images from commercial sources designed for Federated Learning, alongside two innovative algorithms, Fed-Cyclic and Fed-Star. Both algorithms enhance image classification performance while ensuring data privacy, marking a significant step forward in collaborative machine learning practices.

This article presents a new ultra-low-power Coarse-Grained Reconfigurable Array (CGRA) designed to accelerate transformer models on edge devices. By optimizing General Matrix Multiplication operations and enhancing data flow, this architecture addresses the energy constraints of deploying complex deep learning applications.

Fremer introduces a new, efficient deep forecasting model designed for workload forecasting in cloud services. By addressing challenges in the frequency domain, it outperforms existing Transformer models in accuracy and efficiency, while also reducing latency and resource consumption in real-world applications.

This article addresses energy-efficient solutions for low-altitude mobile edge computing (MEC) systems using aerial support. It explores a novel framework integrating generative AI with deep reinforcement learning to optimize task offloading and reduce uplink interference, ultimately enhancing performance in emerging 6G networks.

LaViPlan introduces a framework that enhances Vision-Language Models for autonomous driving by addressing the misalignment between high-level language commands and low-level actions. Using Reinforcement Learning with Verifiable Rewards, it aims to improve decision-making in out-of-distribution scenarios, promising better situational awareness for autonomous vehicles.

This article explores the integration of prediction uncertainty in machine learning to enhance model transparency. By differentiating between aleatoric and epistemic uncertainty, the authors propose new methods for providing reliable explanations, potentially improving the robustness and reliability of AI systems in various applications.

Argus is a novel framework that enhances 3D scene understanding by integrating multi-view images with Large Language Models. By addressing the information loss common in 3D point clouds, Argus improves detail and accuracy in scene reconstruction, demonstrating superior performance in various related tasks.

This article surveys architectural backdoors in deep learning, identifying critical vulnerabilities that traditional defenses overlook. It reviews detection and defense strategies while highlighting existing challenges. The authors aim to guide future research efforts to enhance security in deep learning systems against these subtle threats.

This article presents TReconLM, a novel approach for trace reconstruction that uses language models to recover original sequences from corrupted copies. By fine-tuning models on specific error patterns, TReconLM demonstrates significant improvements over existing algorithms, particularly in the context of DNA data storage.

This article explores a new approach to language models by introducing a hierarchical decoder architecture. By adapting pretrained models to include multiple decoding layers, the authors show improved performance in various tasks, including text classification and generation, demonstrating the potential of hierarchical reasoning in AI.

This article explores enhancements to the Differentiable Automatic Post-editing Optimization (DAPO) algorithm through two innovative modifications. By integrating a stable guiding policy and optimizing the use of zero-reward samples, the authors aim to improve training stability and sample efficiency in machine learning applications.