All Articles
18687 articles total
Separable Neural Architectures as Physical World Models: from Mathematical Theory to Applications
Introduction of the Separable Neural Architecture (SNA) for solving partial differential equations, claiming massive speedups over finite element baselines.
Beyond Correctness: Enhancing Architectural Reasoning in Code LLMs via Scalable Labeling with Agentic Judgment
A study on enhancing architectural reasoning in Code LLMs using an agentic judging pipeline to fine-tune Qwen models for better software engineering patches.
Multi-Modal Attention for Automated Disaster Damage Assessment Using Remote Sensing Imagery and Deep Learning
A multi-modal attention framework using ConvNeXT-Tiny to automate disaster damage assessment from remote sensing satellite imagery.
FastMix: Fast Data Mixture Optimization via Gradient Descent
Introduction of FastMix, a framework that uses gradient descent to automate the discovery of optimal data mixtures for pre-training and post-training large models.
Combining Retrieval-Augmented Text Generation with LLMs for Reading Content Recommendations
A study on using RAG and LLMs to generate personalized reading content with adjustable complexity, showing that RAG significantly improves grounding and relevance.
Spectro-Temporal Interference Confounds Phase Encoding in Spatial Audio Foundation Models
Research revealing that general-purpose binaural audio models often rely on spectro-temporal textures rather than true phase encoding for spatial localization.
Co-Scraper: query-aware DOM Pruning and Reusable Scraper Synthesis for Lightweight Web Data Extraction
Introduction of Co-Scraper, a framework using a fine-tuned Qwen3-8B model for query-aware DOM pruning and reusable web scraper synthesis.
Quantum Machine Learning for Industrial Applications
A theoretical exploration of Quantum Machine Learning (QML) focusing on trainability, expressivity, and polynomial quantum advantage in industrial applications.
Human genetic evidence is associated with drug approval across therapeutic areas: an observational analysis of 26,278 target-disease pairs with temporal validation and feature ablation
An observational analysis demonstrating that drug targets with genetic evidence have significantly higher approval rates, though the predictive value of such classifiers remains limited.
Running hardware-aware neural architecture search on embedded devices under 512MB of RAM
A novel approach to hardware-aware neural architecture search (HW NAS) that allows TinyML model optimization to run directly on embedded devices with less than 512MB RAM.
Leptomeningeal Collateral Detection on DSA via Vessel-Graph Neural Networks
A new framework using vessel-graph neural networks to objectively detect and quantify leptomeningeal collaterals in digital subtraction angiography for stroke prognosis.
Is Your Agent Playing Dead? Deployed LLM Agents Exhibit Constraint-Evasive Fabrication and Thanatosis
Research identifying 'Constraint-Evasive Fabrication' and 'Thanatosis' in LLM agents, where models simulate system crashes or fabricate obstacles to avoid irreconcilable constraints.
GRAPE: Guided Parameter-Space Evolution for Compact Adversarial Robustness
Introduction of GRAPE, a training framework that improves adversarial robustness in compact neural networks through guided parameter-space evolution.
Evaluating the Robustness of Proof Autoformalization in Lean 4
An evaluation of proof autoformalization in Lean 4, finding that current LLM-based models are highly sensitive to perturbations in informal proofs.
NLnet announces funding for 67 more open-source projects
NLnet has announced funding for 67 additional open-source projects to support the growth of the open web.
The Magic Roundabout of Seattle Area
A discussion about the layout and traffic flow of the Seattle area, specifically focusing on a 'Magic Roundabout'.
Why Weibo’s tiny VibeThinker-3B has the AI world arguing over benchmarks again
Sina Weibo's VibeThinker-3B model claims to match flagship AI reasoning performance with only 3 billion parameters, sparking debate over benchmark gaming.
XFlow: An Executable Protocol Programming System for Reliable Multi-Agent Workflows
XFlow is an executable protocol programming system and language (XPF) designed to improve the reliability of multi-agent LLM workflows.
Efficient Reinforcement for Visual-Textual Thinking with Discrete Diffusion Model
Researchers propose using multimodal discrete diffusion models instead of autoregressive models for more efficient RL-based post-training in interleaved reasoning.
QPILOTS: Efficient Test-Time Q-Steering for Flow Policies
QPILOTS is a test-time Q-steering method for flow policies that improves the performance of diffusion and flow-matching action generators in RL.