Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose FL-MAESTRO, a multi-agent orchestrator that uses three specialist large language model (LLM) agents to jointly decide communication topology, per-client resource allocation, and aggregation rule in Fe…
Researchers introduced the Spatiotemporal Conditional Operator (STCO), a neural operator framework for time-dependent PDEs that handles prescribed conditions such as body motion, inflow, and forcing. Across twelve matche…
Researchers introduced Tydra, a hybrid Transformer-State Space Model (SSM) architecture for tabular in-context learning that interleaves attention and SSM layers. Across 30 OpenML datasets, Tydra reduces inference time b…
Researchers introduced a reinforcement learning framework for Continuous-Time Jump Markov Decision Processes (CTJMDPs) with general discrete state spaces, addressing a gap in existing continuous-time RL methods that rely…
Researchers introduced a Time-Aware transformer-based prediction model for Acute Exacerbation of Chronic Obstructive Pulmonary Disease (AECOPD) that uses only home ventilator respiratory data to enable timely detection. …
Researchers propose TH-GNN, a heterogeneous temporal graph neural network that detects LLM-agent shilling attacks on recommender systems by jointly modeling temporal, structural, and semantic signals. In evaluations acro…
Researchers introduced Winder, a joint-embedding architecture that uses a phase-equivariant self-supervised objective to organize cardiac representations into phase-invariant and phase-rotating subspaces, achieving diagn…
An engineer detailed the architecture of a production-grade HIPAA-compliant healthcare document processing pipeline built with .NET and Azure, emphasizing that orchestration, not AI models, is the primary cost driver. Th…
A Medium blog post by Rajendran22 presents Part II of a two-part series on linear algebra for machine learning, covering advanced concepts such as inner products, orthogonality, projections, and the Gram-Schmidt process,…
Scenematic's generative video pipeline uses cheap low-step sketches to select parameters, but out-of-distribution prompts can turn these surrogate scores into noise, leading to costly compounding failures. The team imple…
Microsoft's Vibevoice, a long-form multi-speaker text-to-speech model available on Replicate, synthesizes conversational audio up to 90 minutes in a single pass with support for up to four distinct speakers. The 1.5B par…
Xavriley and collaborators at CPJKU released Beat_this, a beat and downbeat tracking model from the ISMIR 2024 paper 'Beat This! Accurate Beat Tracking Without DBN Postprocessing'. The model detects precise beat position…
Elazarg's Nakdimon model, available on Replicate, is a character-level LSTM-based Hebrew diacritizer that restores diacritical marks to undiacritized Hebrew text without requiring dictionary resources. Built by elazarg, …
Cwalo has released the All-In-One-Music-Structure-Analysis model on Replicate, a unified deep learning system that predicts tempo, beats, downbeats, segment boundaries, and labels from music audio, and optionally demuxes…
Vufinder has released Vggt-1b-Depth, a feed-forward neural network based on research from Meta AI and the University of Oxford's Visual Geometry Group, which infers 3D scene attributes from single or multiple image views…
Aaronjmars has released Unirig-Ai, an AI model on Replicate that automatically generates skeletal rigs and skinning weights for 3D models across humans, animals, and objects. Based on SIGGRAPH'25 UniRig research from Tsi…
Adirik's Titanet-Large model on Replicate performs speaker identity verification by comparing two audio files and determining if they contain the same speaker. Built on NVIDIA's NeMo framework, the model extracts speaker…
Nateraw's Autotune model on Replicate performs pitch correction on voice audio, shifting vocal frequencies to target pitches using strategies like closest pitch or scale-based normalization. It is designed for music prod…
Meta's 7.8 billion-parameter automatic speech recognition model, Meta-Omnilingual-Asr-7b, supports 1,693 languages with character error rates below 10% for 78% of covered languages. Built by Subformer, the model combines…
A developer has completed a Google Summer of Code 2026 project for the sbi library, replacing its string-based network configuration with typed configuration objects across all estimator families. The new API uses frozen…