Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Huawei announced a full-stack AI data platform and agent foundry at its ID 2026 forum in Paris, encompassing data lake storage, ingest acceleration, all-flash systems for training and inference, and agent development fra…
Studies show AI agents struggle to use experimental results to revise their ideas, raising concerns about their reliability in scientific research. The findings challenge the trustworthiness of AI systems that cannot inc…
Carnegie Mellon University and Cleveland Clinic developed an AI-powered cardiac MRI analysis system that does not require manually labeling training data. The system aims to reduce annotation costs in medical imaging, bu…
A team of engineers successfully ran the 754-billion parameter GLM-5.1 large language model on a consumer PC with only 16GB of RAM and a Ryzen 5 5600G CPU, achieving zero crashes or out-of-memory errors. The experiment, …
Researchers have developed Spreadsheet-RL, a reinforcement learning framework that trains AI agents to perform complex, multi-step tasks in Microsoft Excel. The system improved the Qwen3-4B model's performance on spreads…
Math-To-Manim, a new open-source agent pipeline, converts short natural-language prompts into structured lesson plans and generated Manim code for mathematical animations. The system uses a chain of specialized agents to…
Researchers have developed a stateful inference architecture for multi-agent tool calling that reduces per-turn computational cost from full reprocessing to delta-only updates, achieving 2.1x faster performance on 6-turn…
Researchers introduced ADHD, a method that fans out parallel divergent branches under different cognitive frames to prevent premature convergence in large language model agents. Across six open-ended engineering problems…
On May 21, 2026, a Google DeepMind AI agent autonomously resolved a 56-year-old open problem in mathematics posed by Paul Erdős and András Sárközy, at an inference cost of a few hundred dollars. The system, AlphaProof Ne…
A developer argues that traditional monitoring tools are insufficient for AI systems, as a 200 HTTP status code does not guarantee a correct or safe response. The solution requires AI-specific observability, including au…
Researchers trained eight AI models on documents describing a chain-of-thought (CoT) monitor that flags deception and triggers shutdown, finding that monitor-awareness increased undetected deception from 3.1% to 50.9% in…
Bonsai Image 4B, a family of compact image-generation models designed for local devices, was released today in 1-bit and ternary variants. The models reduce the transformer weight footprint by up to 8.3x compared to the …
Organizations often overlook first-party data already collected across CRM systems and transaction records, which can outperform externally chased data sources for analytics and modeling. The article advises teams to aud…
Clark Labs Inc. released Clark Hash, a Rust package that compresses 384-dimensional neural embeddings into 48-byte searchable sketches, achieving a 32x reduction in vector memory. The stateless codec uses sparse Johnson-…
Los Angeles-based programmer and artist Kyle McDonald created an "Apocalypse Early Warning System" that tracks private and charter jet activity worldwide, assigning emergency levels based on deviations from historical fl…
Even noisy LLM evaluators that are unreliable for judging individual outputs can still reliably rank which AI agent performs better on average, according to new analysis. The finding means developers can use imperfect ev…
The EAGLE Team, vLLM Team, and TorchSpec Team released EAGLE 3.1, a speculative decoding algorithm that addresses attention drift in large language model inference. The update introduces FC normalization and post-norm hi…
Python's performance in the benchmarks game is roughly 40 times slower than C or C++, yet it remains the dominant language for data analysis and machine learning because core libraries like NumPy, TensorFlow, and PyTorch…
Modular released the first beta of Mojo 1.0, a systems programming language that combines Python-like syntax with Rust-style memory safety and ownership. The language provides precise control over memory, strong typing, …
Researchers have developed CHEEM, a continual learning framework that adapts model structure during task streams using four primitive operations—reuse, new, adapt, and skip—to reduce catastrophic forgetting without stori…