Archive, newest first. Filter by topic, impact and company →
Problem The paper addresses the limitations of current transformer architectures, which struggle with large context windows due to computational and memory constraints. Existing methods often require substantial resources to process…
Problem — The paper addresses the challenge of heterogeneity in psychiatric disorders and the reliance on unstructured longitudinal narratives, which complicates standardization in clinical decision-making. The authors highlight the variability…
Problem Current evaluations of large language model (LLM) agents predominantly assume static environments, which does not reflect the dynamic nature of real-world applications. This paper addresses the gap in the…
Problem The paper addresses the limitations of conventional retrieval mechanisms in grounding language models for complex reasoning tasks. Traditional methods rely on lexical or semantic similarity, which can misalign with…
Problem This work addresses the gap in the capability of existing image generators to perform interleaved generation, which is essential for applications in visual narratives, guidance, and embodied manipulation. Current…
Problem Articulated tool manipulation poses significant challenges in dexterous robotics, primarily due to the complexity of coordinating internal degrees of freedom and managing contact-rich interactions. Previous research has predominantly focused…
Problem — This work addresses the challenge of synthesizing photorealistic, cluttered scenes that require an understanding of geometry, such as perspective and relative scale. Existing methods for adapting text-to-image (T2I)…
Problem This work addresses the limitations of existing world action models (WAMs) that rely on reconstruction-oriented video tokenizers derived from pretrained video generation models. While these models maintain visual fidelity,…
Problem Spatial reasoning remains a significant challenge for vision-language models (VLMs), particularly in the context of tool-augmented agents. Existing frameworks either utilize single-pass code execution, which limits flexibility by requiring…
Problem This work addresses a gap in the understanding of truncated positional encodings (PEs) in graph neural networks (GNNs), particularly in the context of their expressive power. While existing literature…
Problem Reproducibility in the social and behavioral sciences is traditionally assessed through independent reanalysis of original data, a process that is resource-intensive and difficult to scale. This paper addresses the…
Problem Current research agents based on large language models (LLMs) have made strides in agent orchestration but largely neglect the orchestration of scientific knowledge. Existing methodologies typically condense scientific papers…
Problem The paper addresses the gap in efficient data attribution methods for large language models (LLMs), particularly focusing on the limitations of existing influence function-based approaches. Current methods, while effective…
Problem Current tool-augmented large language model (LLM) agents typically rely on step-wise atomic tool calls, leading to an execution-granularity mismatch. This mismatch arises because locally deterministic workflows are decomposed into…
Problem The paper addresses the emerging gap in the literature regarding the design of agent environments for autonomous scientific discovery, positing that as the capabilities of large language model (LLM)-based…
Problem — This work addresses the gap in understanding the cognitive and epistemic implications of AI systems, particularly through the lens of cognitive colonization. While existing frameworks like Tri-System Theory…
Problem This work addresses the gap in understanding how on-policy distillation (OPD) affects model parameters, particularly in the context of language and vision-language models. Despite OPD's growing prominence as a…
Problem This paper addresses the limitations of existing human-centric reconstruction methods that depend on explicit geometry priors such as skeletons, depth maps, or normals. The authors propose Flex4DHuman, a preprint…
Problem Current image-to-3D methods face a significant trade-off between faithfulness to the input image and completeness of the generated 3D geometry. Depth estimators are limited to visible surfaces, while generative…
Problem — The paper addresses the lack of effective, label-free methods for detecting reasoning failures in large language models (LLMs) during inference. Existing confidence baselines, such as self-consistency and semantic…
Problem This work addresses the lack of comprehensive benchmarks for text embeddings in Slovak, a low-resource West Slavic language. Existing multilingual benchmarks do not adequately cover Slovak, and prior Slovak-specific…
Problem This paper addresses the limitations of existing 3D reconstruction methods that either rely on per-view processing, resulting in overlapping and unaligned pointmaps, or on global-latent approaches that yield fixed,…
Problem This work addresses the limitations of existing recursive language models (RLMs) and coding agents, particularly in their ability to handle long-context reasoning effectively. While RLMs have demonstrated the utility…
Problem — This work addresses the gap in understanding the geometric underpinnings of recoverability in continual learning, particularly in the context of catastrophic forgetting. The authors build on the Accessibility…
Problem Current methodologies for question decomposition in large language models (LLMs) lack a rigorous mathematical foundation, which limits their effectiveness and understanding. This paper addresses this gap by proposing operads…
Problem — This work addresses the gap in existing literature regarding the integration of predictive modeling and optimization for aerial wildfire suppression. Current methods often lack a comprehensive framework that…
Problem This work addresses the gap in understanding how different speech representation methods impact the quality of 3D facial animation. The authors highlight that existing literature lacks a comprehensive evaluation…
Problem The paper addresses the growing reliance on synthetic data in scientific research, particularly in fields like social science and AI evaluations, where biases and inaccuracies in synthetic data can…
Problem This study addresses the gap in generative modeling of symbolic music, specifically focusing on Bach-style compositions. Despite the growing interest in music generation, there is limited comparative analysis of…
Problem The paper identifies a significant gap in vehicle color recognition capabilities, particularly in long-tailed surveillance scenarios where certain colors are underrepresented. The authors highlight that existing methods often fail…
