NVIDIA and IIT Madras's AI4Bharat released the Nemotron-pretraining-Indic-v1 pretraining corpus, roughly 2.8 trillion tokens covering all 22 scheduled languages of India, with 2,601.0B natural-script tokens and 224.6B romanized tokens (Gemma-4 tokenizer statistics); it will soon ship alongside an open-source, GPU-accelerated data generation pipeline.
Microsoft Research released StudentSim, a research system for evaluating and improving AI tutoring tools before testing with real learners, training a personalized simulator for each student based on publicly available de-identified learner records. It also introduces the StudentSimEval benchmark, which evaluates along two dimensions, behavioral fidelity and instructional responsiveness, across three domains: chess, second-language English writing, and math, outperforming leading baselines. The system is for research-stage use only, is not a product, has not been deployed in classrooms, and is not intended to replace teachers.
The co-publishers with Hugging Face released Falcon-Emirati-7B, fine-tuned from the Falcon-H1-Arabic 7B variant to understand and generate Emirati Arabic dialect. The model uses a hybrid Mamba and Transformer attention architecture, trained on Emirati dialect web data, Emirati-culture MSA data, and synthetic data constrained by lexicon and style rules, with a context window supporting up to 128K/256K tokens.
Telecom operators are building their AI strategies on open models; in NVIDIA's latest State of AI in Telecommunications report, 89% of respondents said open-source models and software are important to their AI strategy.
Ai2 open-sourced AstaBrief 8B, a Qwen3-8B-based model, along with its training data. It generates cited reports in one pass from a research question and retrieved literature snippets, and is already live in Asta's Generate a report feature as the Fast mode alongside the Claude-powered Thinking mode.
Google has announced a next-generation federated learning system based on trusted execution environments (TEEs), providing remotely verifiable and auditable guarantees for data anonymization. Access policies are published to the public transparency log Rekor, the KMS and data processing binaries are reproducibly built from open source code, and encrypted training data is decrypted and processed only inside the TEE, with workload operators seeing only metrics and differentially private model weights.
ServiceNow CoreAI launched AutoSynthData, which uses a target model's failures in an environment and successful trajectories from stronger teachers to automatically generate executable, realistic, and hard training tasks for enterprise agent post-training. Each task consists of a system specification, user prompt, and verifier, and must meet feasibility, realism, and difficulty requirements. The team demonstrated the pipeline with a public dataset from EnterpriseOps Gym and dynamically adjusted the curriculum as models improved.
Returns on NVIDIA AI factories depend on three interrelated factors: revenue capability, lifespan, and demand. Per SemiAnalysis AgentX data, NVIDIA Vera Rubin NVL72 delivers over 30x higher throughput per megawatt than GB300 NVL72, with per-million-token costs as much as 45x lower on the DeepSeek V4 Pro model.
At the TSMC OIP Conference held in Santa Clara, California, SK hynix showcased its next-generation memory portfolio and won the Partner of the Year award (3DFabric-HBM Integration category) for the second consecutive year.
AWS proposed an architecture based on Amazon EKS, EFA and DeepEP to accelerate large-scale reinforcement learning post-training of MoE models, delivering 40% higher throughput. The solution targets workloads like RLHF, PPO and GRPO, addressing resource balancing between rollout generation and policy training, high-throughput communication across hundreds of accelerators, and the dynamic all-to-all token routing overhead introduced by Expert Parallelism.
The University of Manchester, working with the NVIDIA Earth-2 team, repurposed the Earth-2 CorrDiff generative downscaling model for air pollution forecasting and added Earth-2 StormCast, which can directly leverage air quality observations.
Google Research proposed ToolGrad at ACL 2026, an "answer-first" method for generating tool-use datasets: it first builds real tool-call chains, then iteratively proposes, executes, selects, and updates API workflows using textual "gradients."
OpenAI has launched the Data agent in ChatGPT Work, allowing users to connect company data using natural language, mine insights and build interactive dashboards. The feature helps users put their data to practical use in an AI-driven way.
A Hugging Face blog author used TRL and OpenEnv to openly reproduce Surya Narreddi's approach, training Qwen/Qwen3.5-35B-A3B to write p5.brush code that paints watercolours.
Why it matters: The author has open-sourced the training recipe, environments, reference pool, and model, so readers can reproduce the entire pipeline with a single command and learn from the troubleshooting notes.
Hugging Face published a tutorial on the blog showing how to use the TRL library to fine-tune LiquidAI/LFM2.5-350M with GRPO to improve structured output compliance, requiring only about 500 samples and 100 training steps, runnable on free-tier Colab or Kaggle GPUs.
Google Earth AI has introduced an experimental research capability called the Planetary Prediction Engine (PPE), which autonomously performs the full geospatial prediction workflow—data discovery, cleaning, feature engineering, model training, and evaluation—starting from natural language queries, compressing work that previously took weeks of manual data engineering down to a few minutes.
Google unveils GlucoFM, a self-supervised foundation model with a two-stream design that separates slow glucose baseline trends from short-term deviations. Across four cohorts, seven clinical prediction tasks and 14 evaluations, its mean PR-AUC improves from 54.7 to 58.8 over the strongest baseline CGM baseline, leading on all diabetes risk and beta cell dysfunction evaluations, and achieves the lowest MAE in postprandial glucose response prediction.
IBM released the Granite 4.2 reasoning model family, comprising 3B, 8B and 30B dense decoder models built on the Granite 4.1 base (pretrained from scratch on about 15T tokens, with a five-stage strategy extending context to 512K), post-trained with SFT and multi-stage GRPO reinforcement learning, all open-sourced under the Apache 2.0 license.
Why it matters: The official disclosure covers the full training pipeline from pretraining and SFT to multi-stage GRPO and agentic RL, transferable to similar model-building workflows.
Hugging Face proposes Quantization-Aware Healing (QAH), compressing GPT-OSS 120B to a 60B-parameter structure and quantizing it to MXFP4, then distilling directly from the original pre-compression model rather than from a recovered bfloat16 checkpoint.
Hugging Face has published a paper that enables offline distillation by caching the teacher's top-100 logits, and introduces a fused chunked KL loss that avoids generating full vocabulary-by-sequence-length matrices.
Berkeley AI Research proposes the ABBEL framework, which replaces full interaction history with a natural-language belief state as the agent's working context and introduces belief grading to supervise summary content. On the CollabBench collaborative coding benchmark, ABBEL-rec-BG closes roughly 50% of the performance gap with full-context models, cuts training steps from 100 to 50, and uses significantly shorter peak context token lengths than the full-context setting.
Google Research released SensorFM, a Large Sensor Foundation Model pretrained directly from unlabeled wearable data. The training data comes from 5 million users who consented to participate in the research, covering over one trillion minutes of multimodal sensor signals across more than 100 countries and over 20 Fitbit and Pixel Watch devices.
Hugging Face details the data pipeline for its PRX image generation model: mixing public and internal datasets for pretraining, using a VLM to regenerate long captions, then converting them into streamable training data. The text encoder has been switched to Qwen3-VL, with text latents computed online during training at a throughput cost of only about 3–4%. Images are uniformly encoded as JPEG quality 92, and the data is built with Lance and streamed for training via MDS (Mosaic Streaming).
Dharma AI published an analysis of the 2026 paper by Goldfeder, Wyder, LeCun and Shwartz-Ziv, "AI Must Embrace Specialization via Superhuman Adaptable Intelligence".
Google released a vectorized dataset based on Farmscapes 2020, converting high-resolution raster maps of over 130,000 km² of England into actionable inventories of hedgerows, stone walls, and tree clusters, helping landowners expand forest habitats without compromising food security.
At AISTATS 2026, Google Research introduced Regularized f-Divergence Kernel Tests, a framework for auditing machine unlearning with greater sensitivity and flexibility.
Google Research has open sourced its hydrological modeling framework on GitHub under the Apache 2.0 license, allowing national weather and hydrology agencies to integrate AI flood forecasting into their own workflows.
Matt Onsum, AI/ML lead, and principal scientist Katherine Labbé at Calico Life Sciences use Google DeepMind's Co-Scientist to connect scattered research findings in aging biology and generate hypotheses worth testing.
Google DeepMind released the Decoupled DiLoCo distributed training architecture, which splits large model training into asynchronously communicating compute islands, isolating hardware failures and sharply reducing bandwidth requirements.
Google Research has published a paper in TMLR introducing Simula, a reasoning-first synthetic data generation framework. Drawing on mechanism design, it breaks data generation into four controllable axes — global diversity, local diversity, complication and dual-critic quality checks — without needing seed data.
Google Research released the ConvApparel dataset, containing over 4,000 human-AI multi-turn conversations spanning nearly 15,000 turns, to quantify the "realism gap" in LLM user simulators.
In the paper 'Forest vs Tree: The (N, K) Trade-off in Reproducible ML Evaluation,' Google Research studies the trade-off in AI evaluation between the number of annotations per item and the number of items annotated, and open-sources a simulator based on real datasets.
Mistral AI released Forge, letting enterprises train frontier-grade models on internal documents, codebases, and business data so models understand internal terminology, workflows, and compliance requirements. Forge supports pretraining, post-training, and reinforcement learning, works with dense and MoE architectures and multimodal inputs, and ships with data pipelines and Mistral AI's training methods, supporting continuous evaluation and iteration. ASML, Ericsson, and the European Space Agency, among others, have partnered with Mistral AI to train proprietary models.
Google Research launched Groundsource, using Gemini to extract structured disaster records from news reports in 80 languages, building an open dataset of 2.6 million urban flood events covering 150+ countries from 2000 to today.
Why it matters: The original gives the full method and validation numbers for extracting disaster data from news with Gemini, letting readers assess its research value.
Berkeley AI Research proposes a NeurIPS 2025 framework that evaluates and optimizes imaging systems directly on information content, using a mutual information metric to unify factors like resolution and noise. The method needs only noise measurements and a noise model to estimate information, and provides an upper bound on true information. Across color photography, radio astronomy, lensless imaging, and microscopy, the metric predicts decoder performance and its optimized results match SOTA end-to-end methods with lower memory and compute and no task-specific decoder.
Berkeley AI Research proposes an off-policy RL value-learning algorithm based on divide and conquer that no longer uses temporal difference (TD) learning, whose Bellman recursion error accumulation makes it hard to scale to long-horizon tasks.
Berkeley AI Research proposes a quantitative theory of the word2vec learning process: under several mild approximations, its learning problem reduces to unweighted least-squares matrix factorization, the gradient flow dynamics have a closed-form solution, and the final embeddings are equivalent to PCA on the target matrix M*.