Google DeepMind published a paper proposing Decoupled DiLoCo, splitting large-scale training into decoupled compute islands that communicate through asynchronous data streams, isolating local hardware failures.
Google proposed Simula, a reasoning-first synthetic data framework reframing dataset construction as mechanism design. It builds datasets from scratch without seed data through global diversification, local diversification, complexification and dual-critic quality checks.
Google Research proposed MoGen (Neuronal Morphology Generation), using PointInfinity point-cloud flow matching to generate synthetic neuronal shapes and supplement PATHFINDER reconstruction training. It reduced reconstruction error by 4.4% on held-out mouse axons, mainly by reducing merge errors.
Google Research studied trade-offs between evaluation item counts and annotators per item in Forest vs Tree: The (N, K) Trade-off in Reproducible ML Evaluation. Common counts of one, three or five annotators are often insufficient; reflecting human disagreement frequently requires more than ten.
S2Vec uses S2 Geometry partitions and rasterised features to turn buildings and roads into multilayer images, learning general embeddings with masked autoencoding (MAE). It beats SATCLIP and GEOCLIP on zero-shot geographic socioeconomic extrapolation such as US population density and median income, but needs improvement on tree cover and elevation.
Mistral AI launched Forge, a system for enterprises to build frontier-class AI models on proprietary knowledge, supporting pre-training, post-training and reinforcement learning. It handles dense and MoE architectures and, when needed, multimodal inputs, with training and governance on companies' own infrastructure.
Mistral AI joined NVIDIA Nemotron Coalition as a founding member. They plan to jointly develop frontier open AI models, with Mistral contributing architecture, multimodal capabilities and fine-tuning tools, and NVIDIA providing compute, model development tools and synthetic data pipelines.
Berkeley AI Research proposed SPEX and ProxySPEX to identify key interactions driving LLM outputs at scale in feature, data and model-component attribution. SPEX turns interaction search into sparse recovery using sparsity and low order; ProxySPEX exploits hierarchy to match SPEX with roughly ten times fewer ablations.
Google introduced urban flash-flood forecasts on Flood Hub, using a new AI approach to predict risks in urban areas up to 24 hours ahead.
Google Research launched Groundsource, using Gemini to extract structured historical disaster data from global news. Its first open dataset contains 2.6 million urban flash-flood records across over 150 countries, spanning 2000 to the present.
Google Research released WAXAL, a large open speech dataset initially covering 27 sub-Saharan African languages spoken by more than 100 million people, under CC-BY-4.0.
Berkeley AI Research proposed a mutual-information framework to evaluate and optimise imaging from noisy measurements and noise models. A NeurIPS 2025 paper validates decoder performance predictions for colour photography, radio astronomy, lensless imaging and microscopy. Designs match end-to-end state-of-the-art approaches with less memory and compute, without task-specific decoder design.
Berkeley AI researchers proposed the first theory quantitatively predicting word2vec learning. Under realistic training conditions, it reduces to unweighted least-squares matrix factorisation with a closed-form gradient-flow solution, and the resulting representation is equivalent to PCA of target matrix M*. From small initialisation, word2vec learns orthogonal linear subspaces in discrete steps, increasing embedding rank and reducing loss. These features are M*'s leading eigenvectors and can be calculated in advance from corpus statistics and hyperparameters.
Mistral fine-tuned Pixtral-12B with LoRA, substantially outperforming the untuned baseline on Aerial Image Dataset (AID) satellite classification and reducing hallucinated invalid class names. Fine-tuning uses Mistral's API or LaPlateforme UI without extensive hyperparameter tuning, with 8,000 training and 2,000 test samples.
Mistral AI, Carbone 4 and France's ADEME completed the first full-lifecycle AI model analysis. By January 2025, Large 2 training and 18 months of use produced 20.4 ktCO₂e, consumed 281,000 cubic metres of water and 660 kg Sb eq of resources.