Machine Learning News

Google released LiteRT.js on July 9, 2026, a JavaScript binding of LiteRT that runs .tflite models in browsers via WebAssembly, XNNPACK on CPU, ML Drift over WebGPU, and experimental WebNN for NPUs. Google reports up to …

A new approach using Quadratic Unconstrained Binary Optimization (QUBO) for evidence selection in retrieval-augmented generation (RAG) systems achieves competitive exact-match and token-F1 performance on the HotpotQA ben…

A new developer published an introductory article on neural networks aimed at beginners, explaining core concepts without requiring a mathematical background.

A developer built Vocalinux, an open-source system-wide speech-to-text tray app for Linux that works offline on both X11 and Wayland. The project uses whisper.cpp by default but supports multiple engines including VOSK a…

A compact Karpathy-style decoder experiment comparing real, complex, and quaternion Transformer projections found that quaternion projections achieve lower validation loss through 2 million training tokens, but real proj…

Researchers have introduced LOD-MSNO (LOD-Multiscale Neural Operator), a hybrid model that combines the Localized Orthogonal Decomposition method with neural operators to improve accuracy on multiscale problems while mai…

A new method called Class-Contrastive Influence (C2I) quantifies synthetic sample usefulness by gradient-based influence on classifiers, outperforming traditional realism-focused approaches in few-shot medical image clas…

A new evaluation technique called reversal-drop reveals that temporal video models often rely on positional encoding rather than understanding visual sequences, with Molmo2 failing to distinguish event order while Qwen3-…

A new study improving the Huth et al. fMRI language decoding pipeline achieved a mean METEOR score of 0.149 and BLEU-1 score of 0.200, an 11% relative METEOR gain over baseline, but the fMRIFlamingo model's 42.86% Top-1 …

A new investigation into transformer models for reinforcement learning reveals that larger embedding dimensions produce more sophisticated internal world models and improve interpretability, even though small embeddings …

Audio-language models like CLAP fail to handle negation, with accuracy on negation tasks dropping below chance levels, according to a new evaluation framework called NegEval-Audio. Researchers found that affirmation bias…

Evolution Strategy (ES) has achieved a 26% reduction in skin friction in turbulent channel flow, outperforming the gradient-based multi-agent controller's 17% reduction and the opposition control method's 22% reduction. …

Hi-LeWM, an extension of LeWorldModel (LeWM) that integrates high-level planning over latent subgoals while keeping the low-level LeWM frozen, outperformed its predecessor by 11.3 percentage points at medium-range horizo…

Researchers have successfully applied evolution strategies (ES) to fine-tune large language models at a billion-parameter scale without dimensionality reduction, marking a first that challenges reinforcement learning's d…

A study introducing PoPE (Popperian Placebo-controlled Evaluation) found that small frozen code models with 0.5 to 1.5 billion parameters do not meaningfully learn from their own errors when generating code. In prompt ch…

Researchers have released Deep4ge, a massive dataset containing 14,227 training runs from 59 deep neural network programs to help detect faults in deep learning systems. The dataset includes 9,845 faulty runs created usi…

A new Environment Parameter Gradient Theorem challenges traditional reinforcement learning by enabling simultaneous optimization of both policies and environments, according to researchers. The theorem introduces a gener…

A new framework, VMoGE (Variational Mixture of Graph Neural Experts), promises to enhance the accuracy of dementia diagnosis using multi-band EEG analysis, achieving an AUC of 0.89 in distinguishing healthy controls from…

A new lightweight model, the Lightweight MultiScale AutoEncoder (LMSAE), achieves competitive anomaly detection performance with a model size under 500 KB, slashing inference latency by 9x and cutting power consumption b…

SinAE, a new AI model from an unnamed research team, unifies molecules, crystals, and proteins into a single Transformer-based architecture, achieving near-lossless reconstructions via flow-matching decoding. By reducing…