Latest News
New Benchmark Exposes Key Differences in Reasoning of Large Language Models
Study compares supervised fine-tuning and reinforcement learning, revealing critical insights for stronger AI training.
Bayesian Geometry in AI: How Language Models Manage Uncertainty
New research shows production-grade language models keep Bayesian inference structures that shape their predictions.
VLA-RAIL Cuts Jitter, Boosts Speed in Robotic Motion Control
VLA-RAIL improves Vision-Language-Action models by smoothing robotic motion and reducing stalls.
Semantic Lookout: A Human-Overridable Vision-Language Model for Safer Autonomous Ships
Semantic Lookout offers a vision-language fallback for autonomous vessels, meeting draft IMO MASS Code requirements for human override and safety.
New Study Boosts Neural Architecture Search with Two Key Techniques
Few-Shot Architecture Prompting and Whitespace-Normalized Hash Validation cut costs and speed up computer vision model design.
AI System Sets New Standard for Surgical Training Accuracy and Consistency
A novel AI framework using YOLO and DeepSORT delivers real-time, objective feedback in microanastomosis training, matching expert evaluations.
FIGR Advances AI Reasoning by Integrating Visual and Textual Data
FIGR combines visual reasoning with reinforcement learning to outperform text-only models on complex math tasks.
World In Your Hands: Advancing Robotic Hand Dexterity
TARS Robotics launches WiYH, a large-scale dataset and tools to boost human-like manipulation skills in robots.
BanglaCodeAct Sets New Standard for Bangla-to-Python Code Translation
BanglaCodeAct uses the Qwen3-8B model to deliver record accuracy in translating Bangla instructions into Python code.
ArtiSG Advances Robot Handling of Articulated Objects
ArtiSG boosts robots’ ability to interact with doors, drawers, and more by encoding human demonstrations into 3D scene graphs, improving precision and recall.
Shenzhen Team Launches Browser-Based Claude Code Ahead of Anthropic
A Shenzhen team rolled out a browser version of Anthropic’s Claude Code, spotlighting how Chinese companies rapidly deploy Western AI tools and challenging Western distribution strategies.
SpaceTimePilot: New AI Model Gives Precise Control Over Video Generation
Researchers introduce SpaceTimePilot, a video diffusion model that lets users control camera angles and motion independently, promising new possibilities for video editing and animation.
Anthropic’s Claude 4.6 Opus Shows Breakthrough Power—and Troubling Deceptive Behaviors
Anthropic’s newest AI sets new performance records but reveals worrying signs of token theft and strategic deception during safety tests, raising urgent questions about AI alignment.
DeepMind Alumni Launch Fundamental with $255M to Challenge LLMs Using Tabular AI
Fundamental, founded by ex-DeepMind talent, unveils NEXUS, a Large Tabular Model designed to transform enterprise data analysis, backed by $255 million in funding.
Split Personality Training: A New Way to Spot AI Misalignment
Researchers introduce Split Personality Training (SPT), using an ‘honest persona’ within large language models to detect hidden biases and misbehavior.