Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers present a meta-learning method that models gradient descent training as a dynamical system using latent ordinary differential equations (ODEs) to predict optimal learning rate schedules. The method outperform…
Researchers introduced ArchEGraph, a large-scale benchmark dataset representing 5,481 buildings as heterogeneous graphs with aligned geometry, topology, weather, and zone-level thermal loads, including 49,326 validated b…
Researchers introduced GraphVerse, a benchmark for evaluating multimodal large language models (MLLMs) on visual graph reasoning, which tests perception, structural understanding, and multi-step reasoning over graph-base…
Researchers propose CoDAT, a Collaborative Dual-Attention Transformer for efficient edge action recognition, which replaces multi-head attention with a lightweight dual-branch module and a parameter-free TShift module fo…
Researchers introduced EntropyMoE, a Mixture-of-Experts architecture for byte-level large language models that routes tokens based on patch entropy, achieving the lowest held-out bits-per-byte among matched dense and spa…
Researchers introduced Graph Machine, an architecture with two explicit edge-based mechanisms—edge-augmented attention and edge-centric referral—that outperforms Transformer baselines on Sudoku reasoning tasks, with abla…
Researchers found that safety alignment in Diffusion Large Language Models (DLLMs) is sparse and transferable, enabling attacks that increase attack success rates from 2.6% to 73.8% on LLaDA and from 1.9% to 86.6% on Dre…
Global AI models trained on North American and European sounds fail to recognize Indian wildlife calls, prompting 59 volunteers to build the country's first open-access ecoacoustic dataset with 5,815 minutes of labeled r…
A team of 59 ecologists, researchers and wildlife enthusiasts has built India's first open, crowdsourced library of wildlife sounds, covering 518 species and nearly 5,815 minutes of recordings from 25 states and Union Te…
A researcher investigating whether fine-tuning breaks a transformer's induction head circuit found that a bug in the induction score measurement initially caused the known induction head L1H6 in the attn-only-2l model to…
Indiana University School of Medicine researchers developed a six-point explainable AI score that identifies patients at high risk of intramyocardial hemorrhage after a severe heart attack, with scores of 4 or higher ind…
The Federal Reserve Bank of New York and the European Central Bank use nowcasting models that estimate current GDP growth, inflation, and employment from high-frequency indicators, and AI is expanding what these models c…
Meta AI's DINOv2 (2023) scales the original DINO self-supervised vision model into a Vision Foundation Model that matches or beats text-guided models like CLIP on dense pixel-level tasks, introducing a patch-level object…
An independent research project demonstrated that three ESP32-S3 microcontrollers can jointly compute Graph Neural Network inference on traffic data using Replicated Secret Sharing, ensuring no single device sees the pla…
An independent researcher demonstrated that timing side channels in multi-agent reinforcement learning (MARL) policies deployed on microcontrollers can leak information about an agent's next action. By measuring inferenc…
NVIDIA's H100, H200, and B200 GPUs offer different trade-offs for AI workloads, and choosing among them depends on factors like model size, memory requirements, and whether the use case is training or inference. The H100…
Two July 2026 papers demonstrate that AI models improve by compressing information rather than accumulating it. The Chinese Academy of Sciences' Institute of Automation released PhiZero, a world model that compresses vid…
A developer's optimization campaign on an RTX 3080 Ti cut a GPT-2-small-class transformer's forward pass from 78.2ms to 1.60ms, a 49× speedup, beating torch.compile's 1.72ms and reaching 136,000 tokens/second. The single…
Sail Research has open-sourced HTDYM (How To Deploy Your Model), an internal performance modeling infrastructure that helps determine the optimal hardware and sharding configuration for serving large language models, aim…
A technical explainer from tensor.khalilli.ai demonstrates that the floating-point arithmetic error causing 0.1 + 0.2 to equal 0.30000000000000004 and training loss to become NaN at step 40,000 stems from the finite numb…