Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A paper submitted to arXiv on 2 October 2026 reports that an observer at the split in split learning can rebuild most of a client's text from the traffic, with gradients adding measurable leakage: on GPT-2, an attacker h…
A developer launched acceptodds, a prediction market where users bet on whether each of the 42,000 ICLR 2027 submissions will be accepted or rejected, using Hanson's logarithmic market scoring rule (LMSR) to set prices. …
A new technique called In-Distribution Forcing targets drifting in autoregressive video diffusion models, where colors and textures shift and motion dynamics decay during long-horizon video generation. Existing approache…
Researchers introduced the Prior-Fitted Language Model (PFLM), a 300M-parameter byte-level transformer pretrained only on samples from a synthetic non-linguistic prior, which learns to predict real text in context with f…
A developer built trail-bird-id, a ~60-line Python wrapper around the open-source BirdNET V2.4 audio classifier that identifies bird calls fully offline from a laptop or Raspberry Pi, with no internet, account, or API ke…
Typesafe.ai released Jev, a new type of model it calls a "system one model" that accepts text input but outputs only floating-point numbers answering user-specified yes/no questions, choice lists and scoring requests, ac…
Four "System 1" decision models — Jev, Laya, Kev and Nimble — are now available as faster, cheaper alternatives to full frontier LLM calls for routing and typed decisions, according to a Towards AI article by Muhammad So…
DeepSeek-V3 uses sequential multi-token prediction (MTP) modules, in which predictions at one depth feed into the next depth, according to a technical explainer published via Towards AI. The article details the MTP compo…
Researchers at QLabs published Dust, a zeroth-order optimization method that matches backpropagation on transformer pretraining at scale, reaching a test loss of approximately 5.17 at 1 million tokens with 16,000 populat…
Developer krypticlogan released ZGC, a tensor graph compiler written in Zig, on GitHub after pivoting from a specialized neural network graph runtime. ZGC targets Zig versions 0.16.0 with 0.17.0 support planned, and the …
Researchers introduced TextReg, a method that regularizes text-space prompt optimization to mitigate prompt distributional overfitting in large language models. The work addresses a failure mode in LLM-feedback prompt re…
Researchers introduced a post-training method for diffusion language models (DLMs) that minimizes Maximum Mean Discrepancy (MMD) between generated and reference distributions in the feature space of a frozen pretrained D…
Karmendra Pandey, a Practice Architect in AI & ML at TEKsystems, has published a five-gate audit for reviewing machine learning designs before they ship, covering data provenance and leakage, decision-cost alignment, lat…
A developer built Touch Grass Birder, an open-source bird call identifier that runs entirely on-device, combining the BirdNET acoustic model with Gemma 3 1B via llama.cpp behind a small FastAPI server. The app records si…
A new paper finds that base models can reason competitively with their reinforcement-learning-tuned counterparts when particular starting token cues are fixed, showing that training data creates associations between a re…
Alibaba's DAMO Academy, with Renmin University of China and the University of Chinese Academy of Sciences, unveiled Elements Claw on July 3, 2026, an AI agent that screened 2.4 million candidate crystal structures in 28 …
Researchers introduced ALoDLM (Adaptively Looped Diffusion Language Models), a method that attributes the quality gap between diffusion language models and comparably sized autoregressive models to a computation-difficul…
Matas reported a 29 percentage point retention lift after using Subsets' explainable AI to target dormant Club Matas Plus subscribers, according to a customer story published by Subsets, the Copenhagen vendor that suppli…
IBM's Qiskit now exposes a C API to its Rust core, letting developers build and run quantum circuits natively from Fortran, C++, and Julia without a Python layer, the company said. The interface, available since Qiskit v…
Princeton researchers Adithya Bhaskar, Jeffrey Cheng and Danqi Chen built Queen, a 4-billion-parameter language model that climbed from 1782 Elo to 2697 Elo over seven rounds of training, a gain of more than 900 points, …