Archive, newest first. Filter by topic, impact and company →
Problem The paper addresses the lack of effective methodologies for zero-day anomaly detection and classification in optical networks, particularly in scenarios involving unseen anomaly types. Existing approaches often require extensive…
Problem The paper addresses the challenge of fine-tuning diffusion policies (DPs) in reinforcement learning (RL) settings, particularly when transitioning from offline imitation learning (e.g., behavioral cloning) to online RL. While…
Problem The paper addresses the limitations of existing methods for encoding the Euler Characteristic Transform (ECT) in neural networks, particularly the conventional discretization of the Euler Characteristic Curve (ECC). This…
Problem The paper addresses the inefficiencies of autoregressive (AR) language modeling during inference, particularly in high-load batch serving scenarios critical for industrial applications. Existing acceleration methods, such as speculative decoding…
Problem Existing remote sensing multimodal large language models (RS-MLLMs) are limited in their support for diverse sensor modalities and tasks, resulting in a fragmented understanding of geoscientific data. This paper…
Problem This work addresses the vulnerability of agent execution traces, which contain sensitive procedural details that can be exploited to recover proprietary skills and strategies. The authors highlight a gap…
Problem This work addresses the gap in autonomous mathematical research agents capable of generating and proving conjectures, a largely unexplored area in AI. The authors present Moonshine, a system designed…
Problem This work addresses a significant gap in the literature regarding the capabilities of Multimodal Large Language Models (MLLMs) in physical tool use, an essential aspect for their application in…
Problem The paper addresses the challenge of data scarcity in training deep neural networks (DNNs) for electrocardiogram (ECG) classification, particularly in the medical domain where acquiring large datasets is hindered…
Problem Existing pretrained time series foundation models (TSFMs) often exhibit high computational costs and limited adaptability to diverse variable types, particularly in scenarios where covariates influence target variability. This paper…
Problem Automatic Depression Detection (ADD) from clinical interviews is a critical task in computational mental health, yet it faces significant challenges. The primary issues include the difficulty of modeling complex,…
Problem Zero-shot learning (ZSL) for inertial measurement unit (IMU)-based human activity recognition (HAR) is hindered by the modality gap between sensor embeddings and semantic class representations. This paper systematically evaluates…
Problem The paper addresses the challenge of selecting appropriate security tools in heterogeneous open-source networks, where the complexity of interconnecting components necessitates a deep understanding of various security mechanisms. The…
Problem The paper addresses the limitations of existing unsupervised term discovery methods, particularly the reliance on centre-based clustering techniques like K-means, which yield uniform distributions rather than the expected Zipfian…
Problem The paper addresses the challenge of secure aggregation in decentralized federated learning (DFL), particularly focusing on the communication overhead associated with gradient transmission. Traditional secure aggregation methods scale linearly…
Problem The paper addresses the limitations of local deployment of large Mixture-of-Experts (MoE) models, which fail to meet the service quality of cloud environments, particularly under low-concurrency workloads. The authors…
Problem Existing neural architecture search (NAS) methods are predominantly designed for specific hardware families, which restricts their applicability and generalization across different platforms. This paper addresses the gap in the…
Problem Current machine learning approaches in drug discovery struggle with out-of-distribution (OOD) generalization, particularly in identifying novel bioactive molecules that do not conform to the training data distribution. This paper…
Problem The paper addresses the limitations of existing video world models that utilize explicit point cloud memory in RGB space, which is computationally intensive and results in information loss due…
Problem This work addresses the limitations of existing vision-language-action (VLA) models in handling long-horizon, temporally dependent tasks in robotic manipulation. Current models primarily rely on immediate observations, which hampers their…
Problem The paper identifies significant gaps in the evaluation of vision-language model (VLM) agents within interactive gaming contexts. Existing benchmarks typically provide a single score for each (agent, game) pair,…
Problem The paper addresses the inefficiencies in training reinforcement learning (RL) policies from scratch, which often require extensive computational resources, careful reward design, and fine-tuning. Many control problems already have…
Problem The authors address the gap in understanding the relationship between data frequency and task learnability in language models, particularly in the context of formal languages. Existing literature often relies…
Problem The paper addresses the limitations of existing reinforcement learning (RL) methods for post-training large language models (LLMs), particularly the issues arising from off-policy training due to training-inference mismatch and…
Problem This work addresses a gap in the literature regarding the universal approximation theorem (UAT) for functional input neural networks (FNN) applied to differentiable maps on infinite-dimensional manifolds. Previous formulations…
Problem Standard diffusion models rely on a single time-homogeneous Gaussian terminal distribution, which lacks the capacity to explicitly model data concentrated on low-dimensional manifolds. This limitation necessitates that the reverse…
Problem Conventional embodied world models predominantly utilize low-dimensional structured action vectors, such as joint angles and end-effector poses, which limit expressive capacity and generalization across diverse robotic embodiments. These frameworks…
Problem The paper addresses the limitations of existing world-action models in robot manipulation, which couple world prediction and action execution at the same temporal resolution. This coupling leads to inefficiencies,…
Problem The paper addresses the inconsistency in AI evaluation reporting, which hampers the ability to compare results across various sources such as leaderboards, model cards, and benchmark papers. This inconsistency…
Problem The paper addresses the limitations of existing neural operators (NOs) that operate primarily on point or edge functions, lacking the ability to effectively model interactions in topological domains. This…
