Archive, newest first. Filter by topic, impact and company →
Problem Blind image restoration aims to recover clean images from observations that have been corrupted by unknown and potentially mixed degradations. Existing deterministic flow-based methods typically model this restoration process…
Problem This work addresses the lack of systematic comparisons between feedback provided by large language models (LLMs) and human experts in the context of writing revision. Specifically, it focuses on…
Problem This work addresses a significant gap in the understanding of circuit discovery methods in neural networks, particularly the assumption that structural differences in discovered circuits imply distinct functional mechanisms.…
Problem Hate speech detection is subjective, with significant variability in interpretation across demographic groups. Existing literature lacks comprehensive evaluations of how well persona-conditioned Large Language Models (LLMs) can simulate these…
Problem The paper addresses the limitations of existing generative recommendation models, particularly in their reasoning capabilities, which are underutilized due to the inability to construct effective Chain-of-Thought (CoT) sequences from…
Problem The paper addresses the inefficiencies of classical Farthest Point Sampling (FPS) in processing large point clouds, which are critical for applications in robotic perception, including LiDAR-based autonomous driving and…
Problem Current transformer-based multimodal models primarily utilize pairwise dot-product interactions or concatenate modalities for attention score computation. This results in either quadratic complexity with respect to the number of modalities…
Problem The paper addresses a significant gap in the literature regarding the extraction of semantically meaningful visual artifacts from institutional documents, which contain critical operational and analytical information. Current methodologies…
Problem Automatic Speech Recognition (ASR) systems have achieved significant performance improvements for standard speech; however, they struggle with pathological speech resulting from neurological conditions. This paper addresses the gap in…
Problem This work addresses a gap in the understanding of factors influencing the effective context window of large language models (LLMs), specifically focusing on lexical density, which has been largely…
Problem The paper addresses the inefficiencies in the Locally Aligned Ant Technique (LAAT) when applied to high-dimensional point clouds, particularly in the context of astronomical data analysis. The authors highlight…
Problem Current question answering (QA) systems, particularly those leveraging large language models (LLMs), exhibit limitations in accurately extracting and generating precise answers from contextual information. These systems often struggle with…
Problem Recent research has focused on understanding the reasoning capabilities of Large Language Models (LLMs), yet a systematic, model-intrinsic signal that captures layer-wise reasoning dynamics remains underexplored. This paper addresses…
Problem This work addresses the substantial hardware resource utilization and energy consumption associated with training spiking neural networks (SNNs) using existing spike-timing-dependent plasticity (STDP) algorithms. The authors highlight that while…
Problem The paper addresses the limitations of existing motion planning algorithms, particularly the Sample-efficient Cross-Entropy Method (iCEM), in complex robotic manipulation tasks such as stacking, sliding, and shelf placement. The…
Problem — The paper addresses the challenge of generating high-quality synthetic question-answer (Q&A) pairs for pretraining language models, specifically focusing on the Nemotron architecture. Existing methods often rely on limited…
Problem The paper addresses the gap in efficient uncertainty quantification methods for neural network predictions, particularly in the context of wide two-layer networks. Traditional approaches, such as deep ensembles, require…
Problem The paper addresses the gap in existing DNA sequence modeling approaches that inadequately capture the dynamic interactions between DNA strands. Current models often overlook the complexities of double-strand dynamics,…
Problem This paper addresses the challenge of Training Data Attribution (TDA), which aims to trace model predictions back to specific training examples. The authors highlight the limitations of existing methods…
Problem The paper addresses the challenge of generating dynamic 3D shapes that accurately reflect both textual descriptions and specified motion trajectories. Existing methods struggle with the ambiguity of natural language,…
Problem Audio-language models (ALMs) frequently exhibit a tendency to favor text-based answers over audio-supported ones, even when the audio evidence is unequivocal. This paper addresses the gap in understanding whether…
Problem The paper addresses the limitations of existing multi-agent reasoning systems that utilize a "generate-then-transfer" paradigm, which results in linear scaling of end-to-end latency with pipeline depth. This work is…
Problem The paper addresses the limitations of traditional reinforcement learning from verifiable rewards (RLVR), which typically relies on binary feedback indicating correctness. This approach is narrow and does not exploit…
Problem This work addresses the scalability limitations of traditional radial basis function neural networks (RBFNs) trained with gradient descent and particle swarm optimization (PSO) methods, particularly in handling large datasets.…
Problem The paper addresses the lack of automated tools for classifying vehicles into fine-grained categories relevant to cyclist injury risk from naturalistic roadway video. Existing object detection benchmarks provide only…
Problem This work addresses a gap in the understanding of failed reasoning in post-trained language models, particularly in how to leverage failed inference attempts for improved performance. The authors highlight…
Problem Existing multi-view image editing techniques primarily focus on rigid transformations or appearance-only modifications, which limits their applicability to nonrigid scene changes. While some methods target specific tasks like object…
Problem The paper addresses the lack of standardized benchmarks for hyperparameter optimization (HPO) in unsupervised representation learning, particularly in the context of biological data. Existing literature often relies on reconstruction…
Problem Generating realistic financial time series is a significant challenge due to the limited availability of historical data, which often leads to overfitting, particularly in adversarial training scenarios. Existing methods…
Problem This work addresses the gap in the literature regarding the application of transformer activation signals for optimizing in-context sample selection in deep active learning. Previous studies have explored active…
