Archive, newest first. Filter by topic, impact and company →
Problem This work addresses the inadequacy of existing memory mechanisms in action-conditioned world models, particularly their failure to maintain scene consistency when the camera leaves and returns. The authors highlight…
Problem This work addresses the challenge of efficiently recommending actions in a linear contextual stochastic multi-armed bandit setting, where user preferences are personalized and context distributions may drift over time.…
Problem The paper addresses the limitations of existing methods for quantifying uncertainty in multi-agent code generation systems, particularly the reliance on costly LLM-driven equivalence checks. Current approaches struggle with the…
Problem The paper addresses the limitations of existing spherical harmonics (SH) methods in view-dependent appearance modeling for novel-view synthesis and reconstruction. Traditional SH approaches often require high-order expansions to capture…
Problem This work addresses the gap in understanding the conditions under which end-to-end optimization of optical front-ends, such as metasurfaces, can outperform traditional lens-based imaging systems for object classification tasks.…
Problem Current table extraction (TE) methods for large-scale document processing are often resource-intensive, requiring billions of parameters, extensive autoregressive steps, or expensive API calls. This paper addresses the gap in…
Problem The paper addresses the cold start problem in the Cloud-Edge Continuum (CEC), where newly discovered nodes lack sufficient historical data for training localized predictive models. Existing generalized models fail…
Problem This study addresses the limitations in existing sound generation tools that hinder composers and sound designers from effectively exploring and refining their sonic creations. The authors highlight the gap…
Problem The paper addresses a critical gap in the literature regarding the evaluation of safety in learned control policies, particularly in the context of quantum predictive control. Traditional safety filters…
Problem The paper addresses the challenge of agent-tool interface grounding in scientific simulation, specifically the need for coding agents to adapt to specialized input languages of simulators. Current coding agents…
Problem The paper addresses the limitations of existing semantic change detection (SCD) methods, which struggle with cross-temporal alignment, multi-scale representation, and robustness to pseudo-changes induced by factors such as illumination,…
Problem This work addresses the limitations of existing topographic models that are unimodal and spatially constrain each layer separately, resulting in fragmented cortical maps. The authors highlight the need for…
Problem This work addresses the challenge of neural machine translation (NMT) for low-resource Indigenous languages, specifically Q'eqchi' Mayan, which suffers from extreme data scarcity. Traditional methods often rely on extractive…
Problem Existing benchmarks for mobile agents primarily focus on isolated task execution without considering user personalization, which is critical for developing effective phone agents. This paper addresses the gap in…
Problem The paper addresses the challenge of maintaining plasticity in deep neural networks during continual learning, particularly under non-stationary conditions. The authors highlight that existing methods often lead to a…
Problem The paper addresses the limitations of parametric imitation learning, specifically behavior cloning, which often struggles with generalization to out-of-distribution states due to compounding errors during deployment. The authors highlight…
Problem This work addresses the gap in existing machine learning frameworks that leverage physical systems for learning, particularly the lack of methods that utilize perturbative responses to derive learning signals.…
Problem The paper addresses the inadequacy of existing frameworks for multi-human, multi-agent collaborations in operational settings, particularly in the context of foundation models transitioning from response generation to more complex…
Problem — This work addresses the lack of guarantees against collisions with task-irrelevant objects in Vision-Language-Action (VLA) models, which have shown strong performance in robotic manipulation tasks. Existing safety filters…
Problem Existing benchmarks for deep research agents (DRAs) primarily assess single-shot outputs, neglecting the potential for iterative improvement through feedback mechanisms. This study addresses the gap in understanding how DRAs…
Problem The paper addresses the gap in formal robustness verification for spatio-temporal neural networks, particularly in safety-critical applications such as autonomous driving and medical imaging. Existing methods often rely on…
Problem This work addresses the gap in understanding the dynamics of gradient descent in feed-forward ReLU networks, particularly focusing on the collective behavior of the network rather than the individual…
Problem Existing LLM-based systems for text-driven indoor scene generation and editing often utilize scene graphs or global constraint lists, which inadequately specify local geometry and complicate localized instruction-based edits. This…
Problem This work addresses the gap in understanding the effects of reinforcement learning from human feedback (RLHF) on the alignment of large language models, specifically regarding the retention of partisan…
Problem The training of parameterized quantum circuits (PQCs) on quantum hardware is hindered by the high measurement costs associated with gradient estimation, particularly when using the parameter-shift rule, which scales…
Problem This work addresses the gap in understanding the sample complexity and VC dimension of depth-$L$ Transformers, particularly in the context of chain-of-thought learning. Prior literature lacked precise characterizations of…
Problem The paper addresses the challenge of enabling large language models (LLMs) to effectively manage complex, long-horizon tasks that require extensive contextual understanding. Current LLMs are limited by finite context…
Problem Disentanglement, the process of isolating distinct factors of variation in data, remains a significant challenge in machine learning, particularly in the context of neural networks. Existing approaches, primarily utilizing…
Problem This paper addresses the limitations of Retrieval-Augmented Generation (RAG) in the legal domain, highlighting its failures such as fabricated citations and outdated legal content. The authors argue that these…
Problem — This work addresses the underexplored relationship between the generative capabilities and self-supervised representation learning of diffusion models. While diffusion models have shown promise in both areas, a systematic…
