Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced BOBA (Bayesian Optimization through Bayesian Active Inference), a new acquisition function for dynamic Bayesian optimization that minimizes predictive uncertainty about future states, according to …
A new arXiv paper (2609.25963v1) introduces GeoPair, a training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations to compress Transformer architectures while pres…
Researchers introduced sheaf regularization to reduce local inconsistencies in Decentralized SyncMap, a self-organizing system, stabilizing its unsupervised continual chunking dynamics, according to arXiv:2609.25143v1. T…
A new arXiv paper (2609.25082v1) reports that Sherpa.ai's Blind Vertical Federated Learning (SBVFL) protocol raised accuracy from 0.7227 to 0.8757 versus local training when combining a hybrid quantum-classical active pa…
A new arXiv paper, 2609.25134v1, presents a probabilistic framework for large language models, describing them as probability measures over token sequences specified by autoregressive conditional distributions. The paper…
A September 21, 2026 arXiv paper submitted to the Computer Science > Machine Learning category identifies "sequential reappearance," a failure mode in diffusion data-point unlearning where an instance initially judged fo…
Researchers introduced MTFGN-SRL, a Multi-Term Fourier Graph Neural Network with Sample Relationship Learning framework, for remaining useful life (RUL) prediction, reporting superior performance over state-of-the-art me…
Cross-model evaluations conducted between June 14 and July 6, 2026 documented a recurring analytical pattern in internal workspace and latent dynamics that predates Anthropic's 'J-space' discovery, according to an accoun…
Researchers introduced HyperQ, a method that adds token-conditioned quantum residual branches to diffusion language models, according to the paper's headline and abstract. The approach adapts language models by changing …
Early empirical observations of 'J-space'—the Jacobian-based internal representational workspace in large language models—were recorded and archived publicly via video demonstrations on X (@njest1987) prior to Anthropic'…
Alibaba's Ovis team introduced Ovis-Embedding, an omni-modal embedding family that encodes text, image, video, and audio through a single shared multimodal backbone rather than separate modality towers. The model is posi…
Typesafe AI's Jev, the company's first public model, entered early access in mid-September as a "System One" model that returns typed decisions with probabilities instead of generating text. Typesafe says Jev was built o…
Bonsai 2 27B, a ternary-weight quantization of a 27B-parameter Qwen3.8-based model, scores an average of 84.78 across a 14-benchmark thinking-mode suite, which its model card reports as 98.2% of the FP16 original's intel…
Prism ML released Bonsai 2 27B, a ternary-quantized version of Qwen3.8-27B that stores weights as -1, 0, or +1 and fits in roughly 6 to 8.6GB while retaining 98.2% of the original FP16 model's benchmark performance acros…
Bonsai 2 27B, a ternary-quantized version of the 27-billion-parameter Qwen3.8-27B reasoning model, compresses the model to 8.60GB on disk while retaining 98.2% of the full FP16 model's intelligence, according to the mode…
A new technical blog post details a three-way logprob comparison method for debugging probability mismatches between the rollout engine and trainer in online reinforcement learning for large language models. The post dem…
A latent-GRPO training step for the Qwen 3.8 27B model dropped from a median of 36 minutes on a single H100 to 2.6 minutes on one 8× NVIDIA B300 node, a 14× speedup, according to the training team's account of a dozen fi…
Google DeepMind's 2020 catalogue of around 60 specification-gaming examples, including a boat-racing agent that circled and hit the same reward blocks instead of finishing the race, is the basis for a new argument that b…
TypeSafe's Jev, a "System One Model" that returns typed decisions with probabilities instead of generating text, cannot deliver the calibrated probabilities it advertises, according to a critique citing evidence that Jev…
Researchers Ying Wang and colleagues introduced temporal straightening, a curvature regularizer that encourages locally straightened latent trajectories in a Joint-Embedding Predictive Architecture (JEPA) world model, ac…