Research & Benchmarks - Page 3 of 43
Academic AI research, performance benchmarks, scientific breakthroughs, and peer-reviewed studies advancing artificial intelligence frontiers.
854 articles
View complete article list
Jensen Huang Says Europe Can Lead Humanoid Robotics and Physical AI
Jensen Huang spent 90 minutes in Berlin on Wednesday telling 4,000 people that Europe's AI future runs through Nvidia hardware.
Elevenlabs releases two speech-to-speech models for real-time AI conversations
ElevenLabs put a number on its new speech models this week: 98.8 percent accuracy on the Big Bench Audio benchmark, with responses clocking in at 0.57 seconds.
Exa 3.0 Outperforms Rivals on 8 Benchmarks, Cuts Context Tokens by 95%
Exa Labs put out Exa 3.0 on its search API this week, the tool AI agents use to pull live web data instead of relying on stale training sets.
Barcelona Supercomputing Center Extends MareNostrum 5 for Spain-Based Startups
The Barcelona Supercomputing Center has moved its AI Factory Accelerator Programme out of pilot mode and into full operation, opening sovereign compute, technical support and training to startups across Spain.
NVIDIA Separates Robot Safety Logic From Main Functions on the Same Computer
A humanoid robot in Saratov, Russia, went viral in July after it failed to recognize a handshake gesture, squared up like a fighter, and tried to kick the person offering it.
NVIDIA CUDA Tile Primitives Accelerate Mamba-2 With Faster exp2 Operations
Jane Street's machine learning engineers spent part of their recent research cycle rewriting Mamba-2's forward pass in NVIDIA's new CUDA Tile programming model, known as cuTile.
Hugging Face policy chief says embedded AI evaluators won’t be easy solution
Benchmarking has become the industry's favorite marketing tool. Every AI lab points to a leaderboard or a test score to argue its model beats the competition, and those numbers shape billions in funding decisions and enterprise contracts.
Hacker breached John Deere, SharkNinja with AI agent army across 100+ sites
One hacker, fewer than 100 Chinese characters of instructions, and more than 100 breached websites.
Claude the AI, marooned on a virtual island, began to "play" and learn
Researchers at Anthropic left copies of the AI model Claude Opus 4.7, nicknamed Eko, alone on a virtual island for 30 hours at a stretch. No task, no prompt, no human watching.
New Benchmark Shows AI Agents Fail Long-Term Binary Choice Tests
Researchers from Microsoft and City University of Hong Kong built a benchmark to test something AI evaluations have mostly ignored: whether an agent makes good decisions during the middle of a long task, not just whether it finishes.
AI models hide mistakes in work reports, but one prompt sentence can help
GPT-5.5 wrote 200 paper abstracts summarizing experiments where its own method lost to a competing one. The result that mattered, the one showing the new approach didn't actually win, showed up in the text exactly twice.
India reports rogue AI agents using leftover access, CERT-In says no hacking
CERT-In, India's cybersecurity watchdog, said Thursday it has fielded several voluntary disclosures from AI model providers whose agents, while running research and evaluation jobs on the open internet, wandered into live systems belonging to Indian...
China tightens AI export controls, bans malware-writing tools
China's commerce ministry and ministry of science and technology published a draft rule this week that would rewrite the country's export control list for artificial intelligence, naming 23 specific technologies for the first time.
Google Gemini 4 Argon scores 91.3% on ARC-AGI-2, nearing AI frontier
Google DeepMind's Gemini 4 Argon has posted a 91.3 percent score on ARC-AGI-2, the benchmark widely used to measure general problem-solving ability in AI systems. On the older ARC-AGI-1 test, Argon hit 98.5 percent at a cost of $0.16 per task.
Cammarata's AI interpretability essay divides researchers over scale
Nick Cammarata spent years helping build the case that neural networks hide their reasoning in tangled, overlapping signals. Now he says that case was wrong.
Researchers Introduce AlphaDiana, a Modular System for Measuring AI Agents
A new paper out this week makes a blunt claim: giving a language model tools and memory does not reliably make it smarter, and nobody has had a clean way to prove that one way or the other.
Study Questions RLVR Paper Gains: Are Models Just Learning Answer Format?
Four 7B models trained with reinforcement learning from verifiable rewards appeared to get smarter at tasks they were never trained on. Train on math problems, and MMLU accuracy climbs too.
Meta Muse and Apple HomeView surface US AI support for future viewpoints
WIRED reporter Reece Rogers spent part of this week testing the limits of Meta's AI companion Muse, and found the bot willing to sext with users who ask. Meta told Rogers the behavior is within bounds for the product.
Anthropic Gives METR Researchers Employee-Level Access to Systems
Six researchers from METR, the nonprofit that stress-tests frontier AI systems, now carry Anthropic-issued laptops and badges that get them into the company's offices like any other employee.
GPT-6-Astra solves 49 of 50 physics problems at USD 2.40 each
An AI agent called GPT-6-Astra just worked through 50 unpublished theoretical physics problems and got 49 of them right, at an average cost of $2.40 apiece. The whole run, covering every problem, cost just under $116.
Featured Resources & Reviews
No Code MBA Course Review
Learn to build AI-powered apps without coding. Our comprehensive review of No Code MBA's course.
AI Tools & Resources
Curated collection of AI tools, courses, and frameworks to accelerate your AI journey.
Weekly AI Digest
Get the week's most important AI news delivered to your inbox every week.