Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
The FDA's public list of authorized AI-enabled medical devices reached roughly 1,451 entries by the end of 2025, with about 300 added that year, and radiology accounts for 76.5 percent of those authorizations. The strong…
Preference Model open-sourced Karotte, a framework it has used in production to build reinforcement learning environments for leading AI labs, hardened through more than a million evaluation runs and controlled red-teami…
Liquid AI released d1-3B, an open-weight text-and-image decision model scoring an average 74.1 across 11 public image benchmarks, and the experimental d1-omni-600M for text paired with images or audio, on October 7th via…
Liquid AI released two open decision models, d1-3B and d1-omni-600M, with d1-3B scoring 48.57 on the Decision Index 0.2.1 — the best result for any model under 10B, ahead of Decider 35B-A3B's 47.11. d1-3B answers a singl…
A paper submitted to arXiv on 5 Oct 2026 introduces "Stepped MoE," a unified framework that combines elastic nested sub-networks with sparsely gated mixture-of-experts routing so a single model can serve at 1, 2, 3, or 4…
A new method for neural network compression learns functional subspaces to select which low-rank components to remove from each transformer weight matrix, according to the paper "Learning Functional Subspaces for Neural …
A developer fine-tuned Qwen3-1.7B, an open model small enough to run on a consumer GPU, to control a home through Home Assistant, raising its score on the Home Assistant community's voice-command test from 75.0% to 89.8%…
Researchers studying one-step action generation in Robotic Foundation Models found that directly applying the MeanFlow framework causes performance collapse, according to a paper titled "Kinematic MeanFlow: One-Step Acti…
Developer Kanak Waradkar built an edge AI bus guardian that predicts bus arrival times to help a friend avoid being stranded, combining the TabPFN tabular foundation model with SerpApi for live transit data. The project …
A new arXiv paper submitted 6 Oct 2026 introduces BOTTLED, a benchmark testing whether LLM agents can autonomously convert general capabilities into cheaper, reusable task-specific solutions, an ability the authors call …
A new white paper introduces HiPHI, a large-scale motion capture benchmark for high-precision human motion and object interaction, aimed at closing the data gap limiting humanoid robot learning. The paper details how Fra…
A developer published a proof-of-concept showing how a machine-learning classifier can flag malicious Linux commands that rule-based detection would miss. The project uses scikit-learn's TfidfVectorizer to convert comman…
Abu Dhabi's Technology Innovation Institute reported that its 1.6-billion-parameter Falcon-ASR model reached a 22.73% word error rate on TII's internal Emirati Arabic evaluation, versus 26.80% for the 30-billion-paramete…
AdaptiveCpp has added a Vulkan backend and Android cross-compilation support to SYCL, bringing portable GPU compute to Android devices that ship with Vulkan implementations. The project's GitHub CI now tests the Vulkan b…
Instacart's machine learning engineering team published a case study on an agentic modeling loop in which autonomous agents formulate hypotheses, write training code, launch distributed training jobs, and evaluate result…
A study at 1.7B-class scale found that a shared Transformer can learn language modeling from fixed token identities rather than a trainable input embedding table, according to the paper's headline claim. The research com…
NVIDIA reported that fine-tuned versions of its Nemotron 3 model family reached gold-medal-level scores at both IOI 2026 and IMO 2026, with Nemotron-3-Ultra-CC scoring 535.4 out of 600 on IOI 2026 — above the 361.12 gold…
A developer's technical breakdown explains why full-parameter fine-tuning of an 8-billion-parameter model in 16-bit precision exceeds 80 GB of VRAM, since AdamW training consumes roughly 16 bytes per trainable parameter …
Apple Machine Learning Research reported that a single coding agent with shell access, called Malena, posted a 62.5% any-medal rate on MLE-bench, beating the best external multi-agent harness, AiScientist, by 15.4 percen…
Vision-language-action (VLA) models face a timing gap between low-rate inference and high-rate robot execution, according to a study that characterizes the gap through end-to-end latency measurements of model inference a…