Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers identified a failure mode in reinforcement learning for large language models called Positive-Credit Contamination, where low-probability tail tokens receive undeserved positive credit. They propose TACO, a m…
Researchers propose Holographic Neural PCFG (Hol-PCFG), a new unsupervised constituency parsing model that uses holographic embeddings to score grammar rules with a closed-form mathematical structure. Hol-PCFG achieves s…
Researchers propose SHIFT, a missingness-aware survival model that predicts from incomplete genomic data without test-time imputation, using masked self-attention and variable-rate feature masking. Evaluated on glioblast…
A new review of remote sensing foundation models (RSFMs) finds that no single geospatial foundation model is universally best and that inconsistent evaluation hinders fair comparison and reliable deployment. The chapter …
Researchers introduced FedTR, a federated learning framework incorporating transfer learning for industrial visual inspection, achieving 95.5% word-level accuracy on homogeneous data and 94.2% on heterogeneous data for l…
Researchers propose a value-of-information router for block-sparse attention that dynamically doubles the kept key blocks when the top-k selection is uncertain, improving recall on LongBench-v2 from 0.47 to 0.75 and pres…
A new survey from arXiv explores post-training techniques for end-to-end autonomous driving models, addressing limitations of open-loop imitation learning in safety-critical environments. The paper categorizes existing m…
Researchers developed a modular speaker diarization front end combined with an adapted Qwen3-ASR-1.7B recognizer for multilingual two-speaker conversational speech, achieving a tcpMER of 17.97 on the final evaluation set…
Researchers introduced Lipschitz Scaling Training (LiST), a method that adjusts the Lipschitz constant of neural networks to achieve out-of-the-box calibration while balancing accuracy and robustness. LiST outperforms ba…
AudarAI released Audar-ASR-V1, an open-weight speech recognition model focused on Arabic and its dialects, with plans to expand to multilingual audio intelligence. The release includes Flash and Turbo models, a technical…
Hundreds of educators are gathering in Seoul to address surging demand for Korean-language instruction, with discussions focusing on digital acceleration and algorithmic alternatives to traditional classrooms. The confer…
A developer at GLM proposes a no-downgrade self-test for coding routes using lower-cost models, emphasizing independent verification over benchmark scores. The test includes small refactor tasks with clear pass/fail cond…
A developer analyzed how large language models develop recurring symbolic patterns like 'goblin' metaphors, identifying five mechanisms including RLHF reinforcement, cultural priors in training data, compression effects,…
An interactive guide demonstrates how Transformer models use attention, embeddings, and positional encoding to predict moves in a fading Tic-Tac-Toe game, making the architecture's inner workings accessible. The article …
A developer built the first open-source collection of 10 agent skills for AMD ROCm GPU workloads, filling a gap where NVIDIA had 428+ skills on skills.sh and AMD had zero. The skills cover GPU setup, Docker configuration…
A developer tested OpenCV Zoo's SFace face recognition model by generating 128-dimensional face embeddings and performing nearest neighbor search on two datasets: Olivetti Faces and Labeled Faces in the Wild. The test us…
Researchers warn that AI tools used to predict climate migration and allocate humanitarian aid rely on datasets biased toward wealthier nations, systematically underrepresenting the Global South communities most affected…
At the SciFM26 conference, researchers discussed the limitations of autonomous labs in capturing serendipitous discoveries like penicillin, highlighting that such systems optimize for predefined metrics and may miss unex…
Deploying AI models, even open-weight ones, introduces significant black box risk because current reverse engineering capabilities cannot predict model behavior or reliably detect backdoors. Research shows that limited d…
Hugging Face released the third part of its 'Profiling in PyTorch' series, focusing on profiling attention mechanisms in transformer models. The post demonstrates how to use PyTorch's profiler to analyze naive attention …