Archive, newest first. Filter by topic, impact and company →
Problem — This work addresses the gap in understanding how image classifiers utilize Fourier phase and magnitude in their internal representations. Previous studies, such as Oppenheim and Lim (1981), demonstrated…
Problem This work addresses the assumption that differential privacy (DP) enhances the robustness of federated learning (FL) against backdoor attacks. Prior literature suggests that DP mechanisms can effectively filter out…
Problem The paper addresses the challenge of post-hoc context erasing in KV caches for long-context language models (LLMs), where local edits can have global repercussions. Specifically, once a span is…
Problem This work addresses the gap in unified frameworks for embodied world modeling that leverage natural language for action representation. Existing models often lack the ability to predict future visual…
Problem This work addresses the inefficiency in reinforcement learning (RL) for deep research agents, particularly when using rubric-based rewards. Existing methods typically rely on large language models (LLMs) to generate…
Problem This work addresses the gap in the literature regarding the effectiveness of complex learned models for long-horizon time-series forecasting. It challenges the prevailing assumption that high-capacity models, such as…
Problem This paper addresses the limitations of existing non-rigid registration methods, which often rely on computationally expensive per-instance optimization, are restricted to narrow object categories, or only handle pairwise inputs.…
Problem The paper addresses the limitations of existing sparse reward reinforcement learning (RL) techniques for large language models (LLMs), particularly in the context of mid-training. Current methods rely on manually…
Problem This paper addresses the gap in machine learning methodologies for analyzing geometric data, particularly in contexts where traditional techniques fail to capture the nonlinear structures inherent in such datasets.…
Problem Existing remote sensing vision-language models predominantly focus on RGB imagery, neglecting the complementary information provided by infrared (IR) data. This oversight limits the understanding of Earth observation, as infrared…
Problem The paper addresses the inefficiencies in context management for large language model (LLM) agents during long-horizon sessions, where context accumulation leads to increased inference costs. Existing methods, such as…
Problem The paper addresses the challenge of generating joint prediction sets for multivariate time series that effectively control for a single event while adapting to cross-coordinate dependencies. Existing methods often…
Problem The paper addresses the well-documented issues of exploding and vanishing gradients in deep neural networks, which hinder effective training, particularly in deep architectures. Despite extensive literature on these phenomena,…
Problem The paper addresses the challenge of integrating human interventions into post-training Vision-Language-Action (VLA) models for humanoid manipulation. Current methods struggle with the complexities of humanoid kinematics and dexterous hand…
Problem The paper addresses the gap in the identification of heterogeneous treatment effects (HTE) in controlled experiments, particularly in the context of policy optimization. Existing methodologies often compromise between expressivity…
Problem This paper addresses the lack of effective metrics for aligning music generation with human preferences, particularly in the context of text-to-music models. Existing methods often fail to provide reliable…
Problem The paper addresses the gap in understanding the reliability and validity of public AI evaluation leaderboards, which are often perceived as definitive rankings. The authors highlight that these evaluations…
Problem This work addresses a significant gap in the computational complexity literature regarding min-max optimization problems, specifically for quadratic polynomials. The authors demonstrate that computing approximate stationary points in this…
Problem The paper addresses the limitations of frozen small code models (≤1.5B parameters) that are designed for offline and privacy-constrained applications but often generate plausible yet incorrect code. The authors…
Problem This paper addresses the inefficiency of applying the Segment Anything Model 3 (SAM 3) directly to open-vocabulary semantic segmentation (OVSS). The authors highlight that traditional methods require full-resolution decoding…
Problem Reinforcement Learning (RL) policies often exhibit performance degradation in novel environments due to their lack of explicit deliberation. This paper addresses the gap in the literature regarding the integration…
Problem This paper addresses the limitations of existing interactive world models in generating long-horizon video content with controllable camera navigation and event prompts. Prior models often lack the ability to…
Problem The paper addresses the limitations of multiphasic contrast-enhanced CT (CECT) in abdominal imaging, which poses risks such as contrast-induced nephropathy and increases the workload for radiologists. Despite the widespread…
Problem The paper addresses the limitations of persistent Laplacians (PL) in machine learning tasks, particularly the challenges posed by high dimensionality and the "varying length" problem across different filtration scales.…
Problem This work addresses the gap in understanding how AI can enhance economic research workflows, specifically in the context of public goods, as outlined in the EC 2025 paper "Stable…
Problem This work addresses the gap in understanding how AI agents achieve their performance, particularly in software engineering tasks. Existing benchmarks primarily report success rates without elucidating the underlying behavioral…
Problem The paper addresses the challenge of accurate Harmonized Tariff Schedule (HTS) code classification, which is critical for customs clearance and regulatory compliance in maritime logistics. Existing methods struggle with…
Problem — The paper addresses the gap in the integration of state-space models (SSMs) within modern probabilistic programming languages (PPLs), which has hindered the application of advanced Bayesian methods in…
Problem The paper addresses the challenge of high latency and operational costs in streaming data systems that require frequent state updates due to incoming events. In production environments, each event…
Problem The paper addresses the limitations of existing pairwise learning methods, particularly their computational and memory inefficiencies when applied to large datasets. Despite the effectiveness of kernel methods in capturing…
