Archive, newest first. Filter by topic, impact and company →
Problem This work addresses the gap in understanding the linear structures present in the weights and activations of neural networks, particularly in the context of task adaptation. The authors challenge…
Problem The paper addresses the reliability risks associated with using large language models (LLMs) for generating solver code in finite element simulations, particularly in multi-physics contexts. While LLMs can streamline…
Problem Long-document question answering (QA) presents challenges for large language models (LLMs) due to the need for reasoning over extensive evidence that is often dispersed throughout the text. Existing retrieval-augmented…
Problem The paper addresses the limitations in existing robotics datasets, which often lack linguistic and action sequence diversity, hindering the ability of learned policies to follow complex instructions. This issue…
Problem The paper addresses the limitations of Large Language Model (LLM) agents in learning efficiency due to suboptimal interaction feedback and static training environments. These constraints hinder the generalization capabilities…
Problem This work addresses the limitations of existing federated learning methods, particularly in the context of linear systematic components. The authors identify a gap in the literature regarding the need…
Problem — This work addresses the gap in understanding how intrinsic symmetries of training data influence the emergence of conserved quantities during gradient-flow training of neural networks. The authors highlight…
Problem Deepfake speech detection has advanced significantly, yet existing models typically provide a single score without elucidating the underlying decision-making process. This lack of interpretability hampers trust and understanding of…
Problem — This work addresses the lack of comprehensive evaluation of deepfake speech datasets, which are crucial for training and assessing deepfake speech detectors. The authors highlight that existing literature…
Problem Reconstructing local stress fields in heterogeneous microstructures subjected to non-linear, history-dependent loading is a significant computational challenge in multi-scale simulations. Existing methods often struggle with the scale mismatch between…
Problem This work addresses the challenge of anti-spoofing in automatic speaker verification (ASV) systems, specifically focusing on the limitations of existing architectures that rely on speaker-reference recordings. The authors note…
Problem This work addresses the gap in understanding how statistical calibration impacts human-AI teaming frameworks. The authors highlight that existing literature does not adequately explore the implications of calibration on…
Problem The paper addresses the limitations of existing image generation methods in achieving effective pose control for customized subjects, particularly in scenarios where users provide reference images and text prompts.…
Problem This paper addresses the limitations of existing Gaussian Mixture Model (GMM) implementations, particularly their inefficiency in handling large-scale datasets due to high memory consumption. The authors highlight that traditional…
Problem Existing text-driven Foreground Conditioned Outpainting (FCO) methods suffer from significant output artifacts, which arise from misalignment between text-derived concept embeddings and specific visual instances. These artifacts compromise the quality…
Problem The paper addresses the limitations of existing post-training quantization (PTQ) methods, which typically rely on simplistic, data-free heuristics for selecting quantization scales. This gap is particularly relevant in the…
Problem The early detection of neurodegeneration is a significant clinical challenge, particularly in identifying individuals at risk of cognitive decline before clinical symptoms manifest. This paper addresses the gap in…
Problem This preprint addresses the significant gap in the understanding and measurement of learner agency and autonomy within educational research. The authors highlight the "jingle-jangle" fallacy, where similar terms may…
Problem This paper addresses the limitations of existing occlusion-based feature attribution methods, which suffer from biases introduced by externally selected baselines, out-of-distribution samples, and unstable explanations. The authors highlight a…
Problem This work addresses the limitations of large language models (LLMs) in executing multi-step tasks using tools, particularly their deficiencies in tool-related knowledge and knowledge activation. The authors highlight that…
Problem This work addresses the challenge of long-horizon autoregressive forecasting of oscillatory physical signals, specifically synthetic seismograms. The authors highlight the issue of error accumulation in causal models, where small…
Problem Current Vision-Language-Action (VLA) models are evaluated primarily under the assumption that all task-relevant objects are visible, which does not reflect real-world scenarios where occlusion is prevalent. This paper addresses…
Problem The paper addresses the gap in literature regarding the sustainability of large language model (LLM) chatbots, particularly focusing on user-interface (UI) interventions to foster energy awareness. While existing strategies…
Problem Current large language models (LLMs) lack a robust mechanism to manage conflicting instructions from sources with varying trust levels, leading to vulnerabilities such as prompt injection attacks. Existing training…
Problem This work addresses the lack of a unified control policy capable of managing various multirotor configurations, such as quadrotors and hexarotors, with a single set of weights. Existing methods…
Problem The paper identifies a significant gap in the evaluation of large language models (LLMs) regarding their propensity for goal-conditioned information distortion. Traditional assessments focus on overt deception, such as…
Problem This work addresses the limitations of existing conversational memory retrieval systems, specifically focusing on the need for improved ranking mechanisms that preserve recall while enhancing retrieval effectiveness. The authors…
Problem This work addresses the gap in stability mechanisms for Q-learning with linear function approximation, particularly focusing on the effectiveness of periodic hard target updates. While existing literature has explored…
Problem This work addresses the gap in evaluating Vision-Language Models (VLMs) specifically for engineering reasoning, a domain that requires interpreting technical diagrams, applying governing physical principles, and executing multi-step reasoning.…
Problem The paper addresses the limitations of existing training-free samplers for masked diffusion language models, which struggle to effectively manage the trade-off between parallelism and prediction accuracy. While methods like…
