Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced BDH-CQ, a reasoning model combining in-context learning with recurrent latent reasoning, achieving 29.5% pass@2 on the ARC-AGI-1 evaluation set with a 150M-parameter configuration at a computed inf…
Developer Bhaumik Tandan built ThodaSa, a TikTok-style shopping feed for the Indian market with a recommender system that runs entirely in client-side JavaScript. The static site on GitHub Pages uses 13-dimensional produ…
Amazon Bedrock Model Evaluation enables teams to compare foundation models with data instead of guesswork, offering automatic metrics and LLM-as-a-Judge modes. A hands-on guide demonstrates building a dataset, running ev…
Hugging Face tests on 11 open-source automatic speech recognition models found several reproduced benchmark transcripts verbatim even when the audio contradicted them, exposing a 10–20% overstatement of real-world accura…
A single LDG.E instruction on an NVIDIA RTX 4090 takes 15 nanoseconds when served from L1 cache, 127 nanoseconds from L2, and 255 nanoseconds from DRAM, revealing a 17× latency gap for memory-bound workloads. The analysi…
Latent Space reports that the AI industry is increasingly replacing human-made components of the machine intelligence pipeline with model-made ones, a trend it calls 'human simulation' that is 10% worse but 100x cheaper …
Noctua, an open-source Kotlin SDK for the Oura Ring, provides on-device AI wellness insights and next-day readiness forecasts without uploading health data to the cloud. The SDK includes a typed Oura API v2 client, an on…
Heart Aerospace's X1 demonstrator, the largest battery-electric aircraft ever flown, completed a 27-minute flight on August 12 from Plattsburgh International Airport, powered entirely by batteries. The aircraft has a 106…
Researchers introduced EnvHarness, a programmable layer of plug-in components that reshapes static environments for LLM agent learning without modifying underlying logic, and EnvRigger, which synthesizes these components…
A study by researchers including Jon Saad-Falcon, submitted to arXiv, proposes intelligence per watt (IPW) as a unified metric for local AI efficiency, finding that local language models with up to 20B parameters answer …
Researchers introduced FlashMorph, a method that automates hybrid attention layer selection using only 20 million tokens and 2.1 GPU-hours, significantly reducing the cost of designing efficient models. The approach pres…
Two new research efforts challenge the trend of collapsing vision, language, and action into a single transformer backbone. VideoChat3 uses a dedicated video-centric backbone to achieve state-of-the-art video understandi…
A new study finds that lexical BM25 retrieval outperforms sophisticated agentic search methods on large corpora, achieving 50.5 accuracy versus 30.7 for File-System Agent at the largest scale. The gap widens with corpus …
Researchers have found that structural template tokens in diffusion transformers act as implicit semantic registers, enabling a training-free head-pruning rule that removes roughly 20% of attention FLOPs with only a ~1.4…
Researchers introduced Self-Geometry, a test-time adaptation method that enforces explicit epipolar consistency to improve vision foundation models' depth and pose predictions without full retraining. On the ETH3D benchm…
NVIDIA's open-source RAPIDS AI framework accelerates machine learning workflows by enabling GPU-based data processing, offering cuDF, a pandas-like library that performs operations on large datasets 100x-1000x faster tha…
Contrastive learning, a self-supervised machine learning technique, trains models to pull similar data points together and push dissimilar ones apart in feature space without human labels, using methods like instance dis…
Machine unlearning, the process of making AI models forget specific data without full retraining, is a difficult problem because neural networks store information in entangled weights, and exact methods like SISA are imp…
Researchers introduced a physical tokenizer that performs analogue light-to-token conversion directly at the sensor level for vision transformers, integrating a 32 × 32 photosensitive memory array of monolayer MoS2 float…
VLLM's disaggregated serving using AMD's MORI-IO connector achieved 2.5x higher goodput on the same 8-GPU node compared to standard collocated serving, according to an April 2026 blog post. The benchmark with Qwen3-235B-…