Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers present OriginBlame (ob), a record- and token-level data provenance system that enables precise identification of training records belonging to a specific author, addressing the gap between data contributor r…
A new study on arXiv (2607.13246v1) finds that Muon, an optimizer that reshapes gradient updates through approximate orthogonalization and has outperformed Adam and AdamW in large language model training, does not consis…
Researchers propose targeted parameter decomposition (tPD), a method that extracts only the neural network components responsible for specific inputs, reducing compute costs. In tests on transformer language models train…
Researchers propose C-Norm (Cell-Distribution Normalization), a method that decouples abnormal and normal cells from ThinPrep Cytologic Test (TCT) images and re-synthesizes them to ensure uniform cell distribution, impro…
Researchers propose Differentiable Geometry Image (DiffGI), an end-to-end 3D-to-2D mapping framework that replaces binary occupancy maps with a continuous 2D Truncated Signed Distance Function (TSDF) to eliminate stairca…
Researchers present AffectFlow-DINO, a multi-task learning system for the 11th ABAW challenge that uses a conditional rectified-flow head to model ambiguity in facial behavior, achieving a multi-task learning performance…
A Transformer-based Masked Autoencoder achieves 91.3% Hungarian matched accuracy in unsupervised grouping of steel surface defects, according to a new arXiv preprint (2607.13178v1). The method, which masks 75% of input i…
Researchers from an undisclosed institution introduced CoDiffGRN, the first co-evolutionary discrete diffusion framework for gene regulatory network inference, which outperforms existing methods in discovering novel regu…
Researchers propose MGFace, a mask-gated face identification pipeline that predicts mask status and conditionally routes similarity computation, achieving over 80% accuracy with FaceNet and over 90% with ArcFace on the L…
Researchers propose UrbanAgent, a multi-agent framework that reframes urban region profiling as a reasoning-driven inference problem, achieving an average improvement of 8.1% in R2 over existing baselines on global datas…
Cogni-ML, a Crystal machine learning library, now supports native Apple Silicon GPU acceleration with Metal inference for GGUF models, including nomic-embed-text-v2-moe embeddings and Qwen 3.5 9B text generation. The lib…
A 200-step, ~1.7B-parameter ablation run in OLMo-core comparing Engram on a standard attention Transformer versus a 3 GDN-layer + 1 attention-layer hybrid found that Transformer + Engram reached slightly lower cross-entr…
Researchers present Fleet, a hierarchical task-based abstraction for multi-die GPUs that introduces Chiplet-tasks to exploit chiplet-level locality and synchronization. On AMD Instinct MI350 with Qwen3-8B, Fleet achieves…
Researchers present a method to accelerate block low-rank inference for foundation models on memory-constrained GPUs, improving efficiency without sacrificing accuracy. The technique reduces memory bandwidth and computat…
A developer successfully ran Google's Gemma 4 E2B and E4B models on Apple's Neural Engine (ANE) using CoreML, achieving a memory footprint under 1 GB on an iPhone 17 Pro. The project overcame ANE's lack of native matrix …
Researchers have released SynapticOS, an open-source inference-first runtime for neural processing units on resource-constrained microcontrollers, built on Zephyr. The runtime achieves constant-time memory allocation (~1…
USC researchers will present three papers at the 2026 Robotics: Science and Systems (RSS) conference in Sydney, Australia, covering robotic foundation models, humanoid loco-manipulation, and robot manipulation learning. …
A developer argues that AI database agents need human review queues, not just approval buttons, to handle uncertainty in real workflows. The review queue should capture ambiguous metric definitions, missing filters, part…
LFortran and Enzyme enable exact gradients from legacy Fortran code without rewriting, turning validated physics solvers into differentiable layers for JAX. The experimental stack compiles Fortran to LLVM IR, differentia…
AMD has tagged ROCm 7.14 "TheRock" tech preview, bringing AI training enhancements, performance improvements of 10-16% for select workloads like Comfy UI, and continued Microsoft Windows support for the open-source GPU c…