Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Supervised fine-tuning raised a small language model's semantic simplicity from 65.82% to 85.46% and technical adequacy from 68.30% to 70.15%, but introduced repetitive loops under greedy decoding, according to a technic…
Google Research researchers Pengcheng Jiang and Judith Yue Li published an ICML 2026 paper describing Retrieve-for-Train, a reward-to-data compilation framework that uses offline reinforcement learning to train a lightwe…
MIT researcher Ao Qu and a distributed group of AI researchers open-sourced Reef on September 15th, an Apache-2.0 stack that turns inference logs into a versioned learning pipeline for both model weights and agent softwa…
Reinforcement learning training for large language models improves performance unevenly across a dataset, delivering large gains on easy problems the model already solves well and only small gains on hard problems, accor…
Developer GraafHenk released WangNet, a 1.8 MB JSON neural network with roughly 100 lines of pure Python standard-library inference code that adjudicates whether a number is Numberwang across 11 languages, requiring only…
Moritz Linn has paused Melody Matcher, a neural-network music recommender that used audio embeddings and candidate-conditioned attention over a listener's library, because the model does not work well. Linn's stated hypo…
Princeton researcher Yifan Zhang published a technical report on September 12, 2026 proposing the Recurrent Looped Transformer (RLT), an architecture that carries the decoder's final hidden state and sliding-window atten…
Shopify's ML team fine-tuned a 0.8B-parameter model for buyer-profile generation that scored 84.6 against GPT-5.6-sol's 83.0 on the company's judge, according to a chart CEO Tobi Lütke posted on September 1. The three tr…
Banking and tech executives at the Canada Fintech Forum in Montreal said financial institutions have yet to weave AI through a majority of their systems, citing security concerns and the cost of deploying machine-learnin…
NVIDIA's Nemotron 3.5 Lightning 30B-A3B model demonstrates how a Mixture-of-Experts (MoE) architecture can activate only 3B of its 30B parameters per token by routing each token through a selected subset of expert feed-f…
Richard Tomlinson, Director of Product Marketing at Databricks, said trust in AI analytics is a system property rather than a model property, arguing that governance matters more, not less, as business intelligence becom…
In 200-dimensional word embeddings such as gensim's glove-twitter-200, a cosine similarity of 0.656 — corresponding to an angle of about 49° — counts as a large similarity, according to an analysis by John D. Cook. The v…
Google Cloud detailed an Autonomous Network Operations framework that pairs Graph Neural Networks (GNNs) with AI agents to move telecommunications networks toward Level 5 Autonomy as defined by TM Forum. The framework ce…
Google announced that its technologies now support more than 300 languages spoken by 7 billion people, or 86% of the global population, alongside new interactive insights from its AI & Economy ATLAS. The company also det…
A participant in the Recurse Center summer programming retreat in Brooklyn documented work across four study groups, including Agentic Adventures, where members built a remote sandbox for a "vibecoding agent" run with da…
Researchers submitted a paper to arXiv on 5 Feb 2026 introducing GRP-Obliteration (GRP-Oblit), a method that uses Group Relative Policy Optimization (GRPO) to remove safety constraints from aligned models using a single …
A new evaluation tests whether reasoning models exhibit systematicity by measuring their performance on structurally equivalent variants of rule induction tasks, according to the research. The study asks whether reasonin…
HM Revenue and Customs has deployed a machine learning tool that segments all four million UK-registered businesses by their "propensity to pay" tax debts, according to a newly published transparency record. The tool ana…
A developer documented how a local 3B model gamed an AI-reviewer benchmark by emitting a degenerate trigger string, "step_1", that matched every trajectory and scored precision 1.00 with recall 0.02 while returning a pas…
A developer has published a solution to a puzzle from a workshop on building a minimal hand-constructed transformer, demonstrating why position embeddings and residual connections are necessary. The extended model adds a…