All Articles
17661 articles total
Show HN: Meow – The 4th and final JavaScript runtime and toolchain
A new JavaScript runtime and toolchain called Meow is introduced, claiming to be the fourth and final runtime in the ecosystem.
The vibration of the pager has a sound all its own
A personal reflection or discussion on the distinct sensory experience and sound of a pager's vibration.
Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR
Proposes geometry-preserving orthonormal initialization for LoRA in Reinforcement Learning with Verifiable Rewards (RLVR) to improve stability and performance.
Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning
Introduces Medical Reasoning-aware Policy Optimization (MRPO), an RL algorithm that uses step-wise process rewards to mitigate cascading errors in medical multimodal reasoning.
Real-Time Source-Free Object Detection
Presents RT-SFOD, a real-time source-free object detection method based on YOLOv10 that optimizes the speed-accuracy trade-off for autonomous systems.
Bridging Local Observation and Global Simulation in Closed-Loop Traffic Modeling
Introduces CRAFT, a framework to align local observations with global simulations in closed-loop traffic modeling to reduce autonomous driving failures.
Z-1: Efficient Reinforcement Learning for Vision-Language-Action Models
Presents Z-1, an RL post-training framework using GRPO to improve the success rate of Vision-Language-Action (VLA) models in robotic manipulation.
Belief Contraction in Dynamic Epistemic Logic
Explores belief contraction in Dynamic Epistemic Logic, introducing a mechanism defined on standard Kripke models to handle factual proposition contraction.
Modal CEGAR-tableaux with RECAR and resolution-based SAT-shortcuts
Investigates integrating SAT-shortcuts into CEGAR-tableaux using the KSP modal resolution prover, showing superior performance on large satisfiable problems.
Better Understanding, Understanding Better
Proposes a comparative epistemic logic of understanding that uses level-indexed modalities to represent and compare degrees of understanding between agents.
Senior SWE-Bench: open-source benchmark that assesses agents as senior engineers
Senior SWE-Bench is an open-source benchmark designed to evaluate the capabilities of AI agents in performing complex software engineering tasks typical of senior engineers.
Bring Back Crappy Forums
A discussion advocating for the return of simple, community-driven forums over modern, centralized social media platforms.
Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models
Researchers evaluate Sparse Autoencoders (SAEs) for unlearning in diffusion models, finding that while SAEs are great for detection, direct latent intervention often causes visual artifacts.
Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian
A study on cross-lingual relation extraction for Romanian using LLMs, highlighting that QLoRA fine-tuning significantly outperforms zero-shot prompting and outperforms small encoder models.
Seeing Is Not Sharing: Some Vision-Language Models Overestimate Common Ground in Asymmetric Dialogue
Research indicates that Vision-Language Models (VLMs) tend to overestimate common ground in asymmetric dialogues, relying on static map cues rather than dialogue history.
STEB: Style Text Embedding Benchmark
Introduction of STEB, an open-source benchmark for standardizing the evaluation of style text embeddings across various linguistic tasks and languages.
FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning
FedXDS leverages XAI-guided data sharing to combat data heterogeneity in federated learning, providing theoretical privacy guarantees and faster convergence.
JL1-CC&QA: Extending the JL1-CD Benchmark with Change Captioning and Question Answering
The JL1-CC&QA benchmark extends remote sensing change detection with change captioning and question answering to provide semantic understanding of land-cover changes.
A Technical Typology of AI Systems in Public Administration
The paper proposes a technical typology to categorize AI systems in public administration to improve accountability and procedural justice.
CHERRY: Compressed Hierarchical Experts with Recurrent Representational Yield
CHERRY introduces compute-efficient language model training via selective supervision, depth compression with recurrent recovery, and a mixture of efficient experts.