Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced HydroGym, a solver-independent reinforcement learning platform providing more than 60 validated flow control environments for fluid dynamics, spanning Reynolds numbers up to Re = 4 × 10^5 and Mach …
Researchers trained supervised learning models to identify 20 iconic jazz musicians from 84 hours of recordings, achieving 94% accuracy with a multi-input architecture that separately represents melody, harmony, rhythm, …
A developer benchmarked nine object-detection models on a single GPU with a consistent measurement setup, finding that measured latencies were 1.6 to 11.6 times slower than published figures, significantly altering the s…
A critical review of on-policy self-distillation for language models finds that the method, which trains a model on its own generations with token-level teacher scores, requires a second, larger teacher model, limiting i…
Software developer softwarewrighter reported that SWTOS, a clean-room preemptive microkernel written in the PL/SW systems language for the 24-bit COR24 FPGA soft CPU, has booted on a physical COR24-TB board at 921600 bau…
A new Nix flake project, immich-nix by axonasif, enables running the Immich photo-management application natively on Apple Silicon macOS and aarch64/x86_64 Linux without Docker, using Nix to supply the build toolchain, P…
A developer explains the fundamentals of posterior inference in probabilistic models, detailing how joint distributions lead to conditional posteriors and how the EM algorithm leverages these posteriors for learning. The…
Falcons AI, a Ras Al Khaimah-based AI company, has seen its NSFW image classification model surpass 50.8 million downloads on Hugging Face in the past 28 days, ranking sixth among over 1,400 models and outpacing DeepSeek…
Researchers introduced FlowBalance, a method that improves reasoning models by combining sparse terminal verifier supervision with on-policy experience, addressing the fragility of self-improvement loops. The approach ai…
Researchers from Princeton University and the U.S. Department of Energy’s Princeton Plasma Physics Laboratory deployed a modular AI control framework called PACMAN on the DIII-D tokamak in San Diego, successfully control…
Google's LiteRT, the renamed TensorFlow Lite, has evolved with new package names and APIs, including the CompiledModel interface and litert-torch for PyTorch conversion. The TensorFlow Lite packages are now in maintenanc…
A developer claims to have found a method to migrate between embedding models with zero downtime, addressing the challenge of upgrading models when dealing with large document collections. The approach is presented as a …
A developer and teammate built DataLens, a data tool with a zero-dependency constraint, by implementing an autoencoder's neural network using only Python 3.14's standard library. They hand-rolled matrix multiplication, a…
Perplexity has open-sourced Lily, a Rust and Metal local inference engine that outperforms Apple's MLX framework by 1.35x on decode throughput for Qwen3.6-35B-A3B on Mac, with prefill speeds of 4,156 tokens/s versus MLX-…
Reducto released r-1, a single-pass document parsing model that cuts error rates by 20% compared to its legacy agentic pipelines and costs 1 cent per page, up to 6x cheaper. The model, available in preview via Reducto's …
A developer has released EmbedFlow, an open-source tool that enables zero-downtime upgrades of embedding models by retrieving top-K documents with the old model and reranking them with the new model, avoiding the cost of…
A developer seeking to convert LIBERO HDF5 datasets to LeRobot format asked how to map the 79-dimensional state vector to LeRobot's 8-dimensional observation.state. The response clarifies that the 79-D vector is a flatte…
Hugging Face released the Ultra-Scale Playbook, a resource on its platform detailing strategies for training large-scale AI models. The playbook provides guidance on ultra-scale training techniques, though no specific fi…
Researchers at an unspecified institution argue that Vision-Language Models (VLMs) lack spatial reasoning due to a dimensional mismatch from training on 2D projections, proposing a method to factorize 4D properties to re…
A new paper and model from researcher George Diamos show that a block-routed mixture-of-experts with 128 experts and 3.65 million active parameters achieves 93% on in-context induction and 90% on positional shift after t…