Voice AI Systems Ignore Emotion Despite Detecting It
New research reveals leading real-time voice assistants fail to act on vocal cues they can perceive, creating risks in high-stakes interactions.
Category
AI Research · Page 3
Papers, breakthroughs, benchmarks, and the long-arc trends shaping artificial intelligence. arXiv highlights and lab announcements, distilled.
Novel Method Improves 3D Video Generation by Tracking Motion Cues
Researchers develop geometric supervision technique that enhances consistency in AI-generated videos from single camera angles.
Self-Distillation Training Narrows AI Model Diversity, New Research Finds
A popular technique for improving language model accuracy may inadvertently reduce output variety and hurt performance on unfamiliar tasks.
AI System Enables Interactive Camera Control in Virtual Try-on Videos
New framework lets users explore garments from any angle, advancing e-commerce and fashion technology.
Researchers Develop Method to Reverse-Engineer AI Agent Strategies
New benchmark reveals how well LLMs can reconstruct hidden decision-making policies through behavioral observation and controlled experiments.
Two-Stage Training Approach Gives Robots Better Motion Intuition
Researchers propose pretrained action modules to accelerate robot learning across different body types and real-world tasks.
New Benchmark Exposes Cancer AI Models' Blind Spots in Minority Patients
Large-scale study reveals state-of-the-art tumor detection systems fail for underrepresented demographic groups, raising concerns about equitable medical AI deployment.
Google Extends Computer Control to Gemini 3.5 Flash Model
The fast, efficient AI model can now interact with computer interfaces, expanding automation possibilities for enterprise and consumer applications.
New Framework Boosts Language Model Performance by 120% in Low-Data Settings
Researchers introduce automated method for matching training objectives to specific AI tasks, dramatically improving encoder-decoder models.
Robots Learn New Skills Without Human Training Data
Researchers develop a framework enabling robots to autonomously acquire manipulation abilities by decomposing tasks into learnable primitives.
Vector Databases for RAG: Pinecone, pgvector, Chroma, and How to Choose
A technical guide to selecting the right embeddings index for retrieval-augmented generation systems.
New Method Adapts Classical Stereo Vision Algorithms to 360-Degree Cameras
Researchers show how to flatten spherical imagery for depth estimation, enabling real-time 3D perception in omnidirectional systems.
Researchers Challenge How We Measure AI Success in Speech Tech
New study exposes flaws in evaluating AI-enhanced communication devices for people with speech disabilities.
Researchers Challenge ImageNet-Only Testing for AI Image Models
New benchmark reveals that methods improving class-conditional generation don't necessarily excel at text-to-image tasks, forcing a rethink of how diffusion transformers are evaluated.
New Training Method Unlocks Math and Reasoning Skills in Vision AI
Researchers demonstrate how reinforcement learning can teach multimodal AI systems to solve complex numerical problems through code generation.
New Training Method Helps AI Models Handle Drastically Longer Text
Randomized YaRN technique enables language models to reason effectively across context windows 16 times larger than training data.
New AI System Reconstructs Moving Objects From Single Video Angles
Researchers introduce Lift4D, a technique that combines 3D modeling with AI priors to capture dynamic deformations and hidden surfaces from monocular footage.
Humanoid Robots Learn to Manipulate Objects While Walking
New training method enables continuous dexterous manipulation during locomotion, moving beyond stop-and-grasp limitations.