Ray Data LLM achieves double the AI throughput of vLLM's synchronous engine by using asynchronous execution at both batch and token levels, eliminating pipeline bottlenecks in mixed reasoning workloads, with benchmarks showing performance gains that continue to grow as decode lengths increase.
New research reveals that base large language models possess strong semantic confidence accuracy, but popular techniques like fine-tuning and chain-of-thought reasoning actively destroy this calibration, raising urgent questions about the reliability of widely deployed AI systems.
Columbia University researchers are deploying cutting-edge AI tools to decode the molecular complexity of human biology, uncovering hidden genetic drivers of pediatric leukemia, Alzheimer's, and cancer in a landmark shift toward predictive, personalized medicine.
A billion-dollar race to build world models is accelerating as AI giants and startups pour massive funding into technology that can simulate physical reality, with AMI Labs raising $1.03 billion and World Labs securing $1 billion to overcome the critical limits large language models face in robotics and autonomous driving.
OpenAI is racing to deploy an autonomous AI research intern by September 2025, with ambitions to build a fully automated AI researcher by 2028, even as major safety and misuse concerns remain unresolved.
Over 300 gather at MIT's 'Gender, Empire, and AI' symposium where experts challenge Big Tech's AI expansion, arguing that smaller, community-centered tools are more beneficial and sustainable than massive large language models.
MIT researchers unveil a breakthrough uncertainty quantification method that detects overconfident AI responses by measuring disagreement across multiple similar LLMs, significantly reducing hallucinations and outperforming existing approaches across 10 tasks including math reasoning and question-answering.
Reinforcement learning environments for AI agents are rapidly evolving, with smarter verifier designs, automated generation dropping costs to just $4 per environment, and emerging standards like Open Reward decoupling infrastructure from training — fundamentally reshaping what large language models can learn to do.
A powerful new open-source framework called vla-evaluation-harness launches with the ability to evaluate 500+ robot AI models across 17 benchmarks at up to 47x faster speeds, completing 2,000 simulation episodes in just 18 minutes on a single GPU.
Google DeepMind unveils a groundbreaking AGI progress framework identifying 10 key cognitive benchmarks — including reasoning, memory, and social cognition — while launching a $200K Kaggle hackathon challenging researchers to build evaluations for under-measured AI abilities.