Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Japanese researchers at the University of Electro-Communications and confectionery maker Ginza Cozy Corner have developed AI-powered robots that replicate the hand movements of skilled confectioners, aiming to preserve t…
NVIDIA Research's Efficient AI Team and Singapore Lab introduced SANA-Video 2.0, a hybrid video diffusion transformer at 5B and 14B scales that generates 720p video on a single H100 GPU. The model achieves a VBench score…
Machine-learning practitioner Enze Xie unveiled SANA-Video 2.0 on July 24, 2026, a full-stack optimized video model designed for efficiency and high quality. The system uses a hybrid attention architecture mixing dense 3…
ArmanddeCacqueray released SquareNet, an open-source tool that transforms wild elephant images into flat grids for machine learning preparation, as shown in a demonstration GIF on GitHub.
A developer shares 10 Python libraries essential for building AI applications in 2026, including FastAPI, LangChain, Pydantic, OpenAI SDK, ChromaDB, Pandas, NumPy, SQLAlchemy, Requests, and Rich. The developer emphasizes…
PyTorch users must ensure exact CUDA version matching between driver, toolkit, and framework to enable GPU acceleration, according to a technical guide on local CNN training. The guide details a three-step deployment flo…
A new post-training method called Codec-Gauge, introduced in arXiv:2607.20538v1, learns small orthogonal channel transforms for Transformer KV caches to improve compression fidelity. Across six models at 3, 4, and 6 bits…
Researchers propose DecodeShare, a protocol that identifies a low-dimensional subspace shared across tasks in LLM decode-time hidden states, and find that disturbing this subspace degrades decision performance more than …
A new arXiv study (2607.20543v1) finds that reinforcement learning with verifiable rewards (RLVR) can improve one-sample accuracy while causing pass@k inversion, where the trained policy solves fewer distinct problems th…
Researchers introduced Lumera, a benchmark and pipeline for engine-native, light-aware 3D scene parsing from a single image, built from 2,513 UE5 projects with 3.73M components and 102.6K parametric lights. Lumera-Box ac…
Researchers have identified new polynomial-time tractable neural network architectures for training with ReLU and linear activation functions, pushing beyond previous state-of-the-art results. For ReLU networks, the team…
A study from arXiv (2607.20519v1) finds that adaptive depth in looped Transformers is a joint problem of trajectory formation and exit readout, not just gate learning. Researchers at Ouroboros (Ouro) evaluated 1.4B and 2…
Researchers propose Generative Bayesian Filtering (GBF), a filtering framework that replaces restrictive observation models with pretrained conditional generative models parametrized by conditional variational autoencode…
A new position paper on arXiv (2607.20532v1) argues that reactive AI maintenance flywheels, which patch models based on observed user errors, fail to preempt future edge cases and become statistically inefficient due to …
A new study from arXiv presents the first systematic comparison of methods to inject biokinetic ODE knowledge into neural networks for bioprocess modeling under data scarcity. The researchers found that a generic decoder…
A new study from arXiv (2607.20540v1) develops a statistical framework for optimal noise-level allocation in diffusion model training, proving that the optimized schedule concentrates on finitely many noise levels under …
A decision-aware machine learning framework developed by researchers and deployed with the Sierra Leone national government increased consumption of allocated essential medicines by 19% in treated districts, according to…
Researchers introduced explanation-based runtime verification for ML-driven optical networks, an approach that uses model explanations to assess decision soundness before execution in the network control loop. The method…
Researchers introduce StressGAT, a Graph Attention Network that achieves 88.62% accuracy in stress recognition by using Differential Action Units to normalize individual facial expressions relative to neutral baselines. …
Researchers propose Uncertainty-Guided Cascade Forward Refinement (U-CFR), a new inference-time framework for interactive image segmentation that enables models to autonomously self-correct after each user interaction. U…