Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A schema-adaptive action-conditioned Joint-Embedding Predictive Architecture (SAAC-JEPA) for CNC dynamics achieved zero-shot RMSE of 0.546, R² of 0.012, and NLL of 0.52 on a sealed target machine sharing only 10 of the s…
A new parameter-pruning scheme derived from the Fisher information metric's geodesic distance in model space outperforms magnitude pruning and local Fisher information pruning across every architecture and dataset combin…
A corpus of 2,507 head-to-head comparisons across 27 scientific disciplines from papers published between 2000 and early 2025 shows AI is not a universal replacement for existing scientific methods, according to a new ar…
Researchers introduced RAIL, an interface exposing polyhedral enclosures of nonlinear dynamics to LiRPA-style bound propagation, and CLIPPER, a branch-and-bound algorithm that jointly refines enclosures and splits contro…
A new arXiv paper (2609.16338v1) introduces BITCOS, a distribution-adaptive storage layout for ternary LLMs that costs 2 − z bits per weight given a zero density z, after measuring 29 ternary models and finding zeros acc…
A new arXiv paper (2609.16454v1) finds that supervised fine-tuning (SFT) is not inherently biased toward mode collapse or its opposite, over-dispersion, in large language models. The authors derive a bias-variance decomp…
A study posted to arXiv (2609.16493v1) proposes the TAISE framework, which repurposes AI weather forecasting models to generate coherent extreme weather sequences for catastrophe risk modeling at an order-of-magnitude lo…
A study comparing three self-supervised pretraining objectives for lung ultrasound found that the best-performing objective on one dataset does not transfer to another, according to research posted to arXiv (2609.16551v1…
A new arXiv paper (2609.16306v1) proposes a method to recognize the motion sequences of Bharatanatyam Adavus, achieving 99% accuracy with a Convolutional Neural Network for Key Posture recognition and 84% with a Support …
A new training framework for Visual Question Answering (VQA) achieves 61.64% accuracy on the bias-sensitive VQA-CP v2 benchmark while maintaining 62.80% on the standard VQA v2 dataset, a generalization gap of 1.16%, acco…
Researchers introduced FLAT (Flexible-Length Aligned Transmodal representations), a pre-training framework that jointly optimizes a shared multimodal encoder with text-to-image and image-to-text decoders, mapping visual …
Researchers proposed FRPSS (Feature Rearrangement in Pre-Shape Space), a single-image generation method whose core MSR-FAGS module replaces randomly initialized low-scale generator features with rearranged Pre-Shape feat…
Researchers developed BrainVLM, a vision-language foundation model that classifies all 12 World Health Organization 2021 brain tumor types from preoperative multimodal MRI data, training on 40,043 individuals and validat…
A new arXiv paper (arXiv:2609.16768v1) proposes a coverage-aware virtual IMU augmentation framework to improve human activity recognition (HAR) when labeled wearable-sensor data is scarce. The method selects diversity an…
A new arXiv paper (2609.16752v1) introduces the Cognitive Field Network (CFN), a recurrent Transformer in which an organized hidden field re-enters subsequent inference via the update rule Phi_{n+1}=F_theta(X_{n+1},Phi_n…
Researchers proposed an Affect-Prototype-Conditioned Fusion (APCF) framework for open-vocabulary multimodal emotion recognition that handles incomplete and unsynchronized modal data, according to an arXiv paper (arXiv:26…
Researchers introduced AntennaFlow, a three-stage generative flow model that performs phaseless, offset-vector-free near-field to far-field reconstruction for large-aperture antenna testing, according to a paper publishe…
Researchers investigated how specialist models trained solely on question-answer pairs, without explicit reasoning supervision, generate reasoning trajectories for domain expert distillation. The work examines what gover…
A study of Transformer representation evolution decomposes learned additive updates into parallel and perpendicular components, finding that pretrained models preserve their current direction more than they redirect it. …
An international research team in the I3LUNG project trained AI models on data from 2,396 patients across six medical centers in Italy, Germany, Spain, Greece, the United States and Israel, raising doctors' accuracy in p…