Machine Learning News

A developer released a portable profiler that measures Apple Neural Engine KernelDMA prefetch behavior across M1, M3, M4, and M5 chips by timing the private _ANEClient evaluateWithModel call (eval_us) rather than host-do…

Cohere released North Small Translate, an open-weights mixture-of-experts machine translation model that scores 83.6 on the WMT26 All Languages benchmark, which Cohere says outperforms DeepL NextGen (81.37) and Google Tr…

IBM and NASA released the open-source NASA-IBM Lunar Foundation Model, a pretrained AI model and machine-learning-ready dataset for mapping lunar craters, volcanic formations, and potential ice deposits. The dataset cont…

A technical survey traces five rewrites of transformer attention since 2017, framing every variant as either storing less or reading less. It details how multi-query and grouped-query attention shrink the KV cache from 2…

Independent researcher Jinho Jang published a cross-scale mechanistic study of safety training in Mixture of Experts reasoning models, reporting 40+ novel findings from 200+ controlled experiments across nine models. The…

A University of California, Berkeley team led by computer science professor Trevor Darrell trained a robot model on 100 hours of high-quality tactile data covering more than 200 household objects, then fine-tuned it on a…

Researchers introduced FreeFlow, a bias-free hierarchical transformer for optical flow estimation that avoids task-specific inductive biases such as correlation volumes, feature warping, and iterative refinement. The wor…

Research published in PRX Life by a computational biophysicist's lab shows that a single E. coli bacterium can learn from past experience, store memories and use them to prepare for future conditions, with the team track…

Researchers propose DRG-MAPPO, a hierarchical dynamic role-graph multi-agent reinforcement learning method aimed at improving tactical coordination for cooperative air combat. The work addresses the difficulty of achievi…

Moonlake founders Chris Manning and Fan-yun Sun are proposing causal, action-conditioned world models that fuse game engine abstractions, symbolic reasoning, and learned neural priors, positioning the approach against Ya…

Cohere released North Small Translate, an open-weight sparse Mixture-of-Experts translation model with 218B total and 25B active parameters covering 50 languages, which scores 83.6 on Cohere's WMT26 all-languages evaluat…

Red Hat senior machine learning engineer Bartowski published a per-tensor layout map method for GGUF quantization on September 10th, replacing fixed name-, shape- and position-based precision rules with maps derived from…

Researchers Ayhan Suleymanzade and co-authors submitted a paper to arXiv on 10 September 2026 proposing "looped flows," a training method that uses local denoising objectives with progressively decreasing noise levels an…

A developer outlined disaster recovery strategies for web applications hosted on Azure Kubernetes Service (AKS) that depend on more than 50 on-premises machine learning inference services, adapting the classic four-tier …

Researchers at Rice University's PHI Lab released GRADE, a single-frame radar depth estimation system that grounds a pretrained generative prior in 4D mmWave radar geometry to produce metric depth under visual degradatio…

A new paper introduces MAAGL, a multi-agent agentic graph learning framework that partitions a graph into communities and assigns an independent agent to each community for region-specific specialization. MAAGL summarize…

A training-free framework combining reasoning and retrieval over Wikipedia improves multilingual multimodal entity linking by 6.9% overall and up to 23.3% on rare-entity slices on the MERLIN benchmark, according to an ar…

A new arXiv paper (2609.10789v1) proposes a two-parameter flow map learning framework for continuous-time diffeomorphic image registration that learns the solution of a non-autonomous ODE directly, without time discretiz…

A new arXiv paper (2609.10792v1) tests whether transformer-based models internalize the same linguistic features as traditional feature-based classifiers in Automatic Readability Assessment across Arabic, English, French…

A new arXiv paper (2609.10830v1) reports that language models' apparent memory of training data is largely confounded, using exact duplication counts from the published pretraining corpora of OLMo-2 and Pythia. Across fi…