Research & Benchmarks - Latest AI News & Updates
Academic AI research, performance benchmarks, scientific breakthroughs, and peer-reviewed studies advancing artificial intelligence frontiers.
502 articles
View complete article list
Auto-FL-Research Uses Agents to Automate Federated Learning Algorithm Search
Federated learning research hinges on countless subtle yet impactful decisions, from optimizer tweaks and aggregation protocols to regularization strategies and architectural nuances.
60% of Experts Say Humanity's Last Exam Is Necessary and Useful
Imagine a test so difficult that even the most advanced AI models fail more than half the time, a benchmark designed not just to measure intelligence, but to push it to its absolute limit.
Study Evaluates AI Retrieval Techniques for Finding Models Across Formats
Finding the right simulation model among dozens—or hundreds—of candidates has long been a bottleneck for engineers and researchers.
Researchers unveil RSEA, a three‑layer self‑evolving language agent
What if an agent could not just learn from experience, but rewrite the very structure of its own reasoning?
Automate Web Research and Brief Writing with a Python Project from 2026 Guide
The weekly grind of market research is brutal. You open a dozen tabs, skim endless articles, and try to stitch together a coherent brief from fragmented signals. Hours vanish. The result? Often a shallow summary, not a strategic insight.
Meta AI launches Brain2Qwerty v2, MEG pipeline hits 61% word accuracy
Meta AI’s latest brain-to-text model doesn’t read minds, it reads MEG signals, and it reads them disturbingly well. Brain2Qwerty v2 hits 61% word accuracy, a jump that leaves prior non-invasive methods, stuck at 8%, in the dust.
Birkhoff’s 1930s ‘measure’ and AICAN’s ‘novelty’ probe AI aesthetics
The MIT Keller Gallery will host “Beyond Data‑Driven Aesthetics” through June 30, a show that pulls together philosophy, mathematics, computer science and design computation into tangible installations and interactive visualizations.
MiniMax Token Plan offers extensive coding model access for USD 20/month
Twenty dollars a month. That’s the price of a streaming subscription, a couple of coffees, or, if you’re a developer, unrestricted access to MiniMax’s coding models across an entire ecosystem of tools.
Sina's VibeThinker-3B probes limits, shows reasoning compresses, knowledge weak
Size is no longer the sole arbiter of intelligence. Sina’s VibeThinker-3B is a deliberate provocation, a bet that raw scale has been masking a simpler truth: reasoning compresses, but knowledge does not.
MRAgent beats RAG, A-MEM, MemoryOS, LangMem, Mem0 with 118K tokens/query
MRAgent is the latest entry in a crowded field of agentic memory frameworks. While A‑MEM relies on a graph‑based approach and MemoryOS layers memory hierarchically, LangMem and Mem0 also promise persistent context across long interactions.
OpenAI unveils Jalapeño custom inference chip, challenging Nvidia's AI dominance
Nvidia has lorded over AI chips for years. That grip is loosening. OpenAI just dropped the Jalapeño, a custom inference chip built with Broadcom, and it’s more than a spicy metaphor.
Prerequisites for NVIDIA AI‑Q Blueprint on OCI: Cluster and Volume Limits
The path to deploying a production-ready NVIDIA AI‑Q Blueprint on Oracle Cloud Infrastructure begins not with code, but with capacity.
Episode 11 Explores Overfitting as RAG Evaluation Scores Keep Rising
RAG evaluation scores are climbing. That should be good news, proof your system is getting smarter, more reliable, more production-ready. But pause. Look closer at those rising numbers.
LLM pipeline compares DAO ERC‑8004 and Google A2A governance, 4,323 records
Two visions of agentic AI governance are colliding. One is permissionless, on-chain, and built for decentralized autonomy, ERC-8004. The other is corporate-led, protocol-driven, and designed by Google, A2A.
Calibration uses NVIDIA Triton Llama-3-8B A10 and vLLM Qwen2.5-7B RTX 4090 data
Inference doesn’t scale the way you think. Not when the clock starts ticking at the ISO 8583 budget. Five thousand single-transaction calls to a GBDT scorer on one CPU core at batch size 1. That’s the hot path.
Figma launches AI motion graphics, shader tools, code layers, and new creative materials
Figma just detonated a bomb under the design industry. Not with a simple update, but with a full creative arsenal. Motion graphics, shader effects, and live code layers are no longer separate disciplines; they’re canvas tools now.
NVIDIA RTX PRO 4500 Blackwell GPUs Power New Amazon EC2 G7 Instances
The GPU arms race has a new winner. Amazon Web Services just lit the fuse. The new EC2 G7 instances, powered by NVIDIA’s RTX PRO 4500 Blackwell Server Edition GPUs, aren't just an incremental upgrade, they are a declared leap.
Stanford researchers present agentic AI 'scientists' at VB Transform 2026
At the top sits a chief scientist officer agent, an AI planner that doesn’t run experiments but orchestrates them. It delegates tasks to teams of specialized agents, each handling a discrete piece of the drug discovery puzzle.
Turning Logistic Regression Coefficients into Credit Score Grid
A client’s credit score isn’t born from thin air. It’s carved out of raw data using logistic regression, a model that spits out coefficients, not intuitive numbers. So how do you turn those messy decimal weights into a clean, actionable score grid?
Metric-Dependent Annotation Saturation for Learning from Label Distributions
Annotator disagreement is not noise, it is a signal. But how many annotators does it take to capture that signal? The answer depends on what you are trying to measure.