Archive, newest first. Filter by topic, impact and company →
Problem This work addresses a critical gap in the literature regarding the impact of chain-of-thought (CoT) supervised fine-tuning (SFT) on long-range recall capabilities in hybrid linear-attention models. The authors demonstrate…
Problem The paper addresses a critical gap in the literature regarding the alignment preservation of large language models (LLMs) when they are converted into reasoning models through post-training techniques. While…
Problem This work addresses the apparent lack of overfitting in benchmark-driven machine learning (ML), particularly in the context of large language model (LLM)-driven research agents. The authors propose that successful…
Problem The paper addresses the limitations of existing conformal predictive systems (CPS) that operate under the assumption of exchangeability, which is often violated in real-world scenarios. Specifically, it tackles the…
Problem This work addresses the lack of comprehensive benchmarks for evaluating AI agents on long-horizon, high-value tasks within professional domains. Existing benchmarks primarily focus on general-purpose software and short-horizon tasks,…
Problem Existing Spiking Neural Network (SNN) motor control systems typically address bipedal locomotion and arm control independently, creating a gap in integrated control solutions for humanoid robots. This paper addresses…
Problem This paper addresses the limitations of existing approaches for integrating speech inputs into large language models (LLMs), which typically rely on cascaded automatic speech recognition (ASR) and LLM pipelines,…
Problem Existing deep learning models for Positron Emission Tomography (PET) image denoising exhibit significant performance degradation when faced with distribution shifts, such as variations in dose levels or scanner types.…
Problem This work addresses the limitations of existing reinforcement learning (RL) methods, specifically the ratio clipping used in Flow-GRPO and CPS for flow matching models in image and video generation.…
Problem The paper addresses the limitations in sequential recommendation systems, particularly the inadequate quality of item representations that hinder predictive accuracy. Existing methods primarily utilize static encodings of fixed attributes,…
Problem The paper addresses the challenge of temporal inconsistency in learning-based motion planners for autonomous driving, which can lead to unstable trajectories and compromised safety and comfort. Existing methods that…
Problem This work addresses the challenge of measuring subjective constructs, specifically human values, in social media texts. Existing literature lacks robust annotation procedures that are theoretically grounded and empirically validated,…
Problem — This work addresses the gap in understanding whether large language models (LLMs) possess a systematic decision-making structure or merely imitate rationales when making choices. The authors highlight that…
Problem Tuning controllers for strongly coupled multi-input multi-output (MIMO) systems presents significant challenges due to the complexities of loop interactions and non-convex cost landscapes. Traditional decentralized auto-tuning methods often overlook…
Problem The paper addresses the limitations in cumulative dose estimation during adaptive radiotherapy (ART), particularly in the context of cervical cancer treatment using Cone Beam Computed Tomography (CBCT). Existing methods…
Problem This preprint addresses the gap in understanding how large language models (LLMs) adapt math word problems across diverse languages and cultural contexts. While LLMs are increasingly employed for personalized…
Problem The paper identifies a significant gap in the literature regarding the accessibility of AI risk management for non-technical users of the OpenClaw framework. While OpenClaw has demonstrated its capability…
Problem The paper addresses the lack of annotated datasets and effective segmentation methods for intermediate phases in zinc-based alloy microstructures, which are critical for understanding their mechanical and functional properties.…
Problem This work addresses the challenge of unlearning in large language models (LLMs), specifically the need to suppress undesirable knowledge while maintaining benign capabilities. Existing methods primarily focus on suppressing…
Problem This paper addresses the significant bottleneck in 3D asset production related to the synthesis of category-agnostic 3D animations. While recent advancements in generative AI have improved static 3D model…
Problem The paper addresses the limited evaluation of visual in-context learning (VICL) models, which have primarily been tested in narrow setups that do not require real adaptation to new tasks…
Problem The paper addresses the gap in evaluating the capabilities of Large Language Models (LLMs) in navigating complex, professional-grade productivity software, specifically in the context of document automation. Despite the…
Problem The paper addresses the critical gap in explainable artificial intelligence (XAI) for network operations, particularly the challenge of translating complex model outputs into actionable insights for non-specialists. Existing XAI…
Problem The paper addresses the gap in accessible AI tools for biodiversity monitoring, specifically for UK fauna, which are often restricted to commercial platforms or trained on non-relevant species. The…
Problem This work addresses the gap in trajectory data augmentation methods, particularly the lack of systematic approaches for selecting which trajectories to augment. Prior research has shown the potential of…
Problem The paper addresses the challenge of tracking provenance in AI compilers, which often perform aggressive rewrites of computation graphs through normalization, lowering, and optimization. Existing methods for provenance tracking…
Problem The paper addresses the limitations of existing multi-token prediction (MTP) methods in large language models (LLMs), which suffer from a fundamental architectural flaw: the competition between the MTP head…
Problem The paper addresses a significant gap in the literature regarding the limitations of predictive models in structural causal inference, particularly in representing uncertainty over counterfactual couplings. The authors identify…
Problem This work addresses the gap in evaluating large language model (LLM)-based coding agents in unfamiliar programming languages, a scenario often overlooked in existing benchmarks. Most contemporary evaluations focus on…
Problem — This work addresses a significant gap in understanding the robustness of alignment mechanisms in large language models (LLMs) against biased inputs. Despite extensive post-training alignment efforts, the authors…
