Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A developer built an open-source early-warning system that uses public ADS-B data from the OpenSky Network to flag unstable approaches by general aviation training aircraft near Daytona Beach International Airport (KDAB)…
A team of AI agents using Opus 5.5 designed one room-temperature magnetic semiconductor candidate and identified a second, a material first made in 1999, whose calculations predict it has the targeted Luttinger compensat…
A pair of arXiv pre-prints, Spatial Memory Intelligence (SMI) and Latent Spatial Memory (LSM), introduce an understanding-driven long-term memory that couples semantic meaning with 3-D location, replacing the replay buff…
Reflection introduced Beam, its first open-weight model, a sparse Mixture-of-Experts system with 501 billion total parameters and 23 billion active parameters built for coding, reasoning, and agentic workloads. Reflectio…
LeRobot's dataset visualization tool lerobot-dataset-viz inspects recorded episodes — camera frames, robot state and action streams, and episode timing — but does not track training loss, according to a community breakdo…
A Tamil-speaking developer built a 60-question benchmark comparing 13 AI models on English, Tamil script, and romanized Tanglish to measure what he calls the "script penalty." Google's Gemini 3.5 and 3.7 Flash scored 100…
Google DeepMind released AlphaProtein Novo (AP Novo), a generative diffusion pipeline for de novo enzyme design via structural motif scaffolding that co-generates protein structures and sequences conditioned on a catalyt…
Liquid AI released d1, its first decision model, adding image support to a system it says became the first to rival Jev on text decisions with last week's experimental release. Liquid AI reported that d1 matches or beats…
A developer released two Qwen3.5 fine-tunes, Jiwo, that top the decision index leaderboard in their size categories, with the 0.8B model beating the best other Qwen3.5-0.8B fine-tune by 45.7% in the sub-1B class and the …
Yandex Music replaced its production recommender's 15+ candidate generators, pre-ranker and ranker with a single transformer model called Sona, according to an A/B test reported by the team. The company said the end-to-e…
The Coalition for Secure AI introduced the Artifact Integrity Maturity Model (AIMM), a framework for assessing trust and provenance in AI/ML supply chains, according to a post on the organization's site. The post opens w…
MintFlow's training-free minimal-intervention strategy enforces constraints on flow-based generative models with a 2x reduction in KL drift versus projection after the final step, according to the MintFlow research, whil…
Orange tested an AI system that targeted a €10 monthly Livebox Fibre discount only at hesitant visitors, and Kameleoon's case study reports the targeted version lifted conversion by 11.6% versus a no-code control and 7.6…
A study measuring order dependence in LLM scorers finds that models with equal ranking quality can produce different decisions depending on the order candidates appear in a single prompt, affecting passage reranking, res…
A machine-learning "speech clock" trained on nearly 3,000 people in Latin America estimated age from acoustic and linguistic speech features, and a larger gap between estimated speech age and chronological age was linked…
A Rust port of Super Smash Bros. Melee (NTSC-U 1.02) reproduces the GameCube game's simulation bit-for-bit, matching the original on 25 of 26 characters and all six tournament stages across every recorded test scenario, …
Developer RobViren released Parakeetogo, a standalone C++ CPU inference engine for moondream's parakeet-redux ternary (1.58-bit) speech recognition model, which runs about 2x faster than moondream's own Photon runtime on…
Developer lcmialichi released PHP-GPU-tensors 0.1.0-beta.4, a native PHP extension that runs GPU tensor operations and runtime-compiled CUDA kernels without a Python runtime. In a PatchCamelyon MLP training workload on a…
A Stochastic Blog post builds a multilayer perceptron on raw MNIST pixels, verifying a hand-derived gradient against autograd before running a controlled sweep over seven activations, four initializations, three normaliz…
Efficient reasoning training does not always harm chain-of-thought faithfulness and monitorability, according to research addressing the concern that training large language models to solve tasks with fewer tokens degrad…