Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Cohere launched Embed 5 on September 30, 2026, with Pro priced at $0.12 per million text tokens and Fast at $0.08, while image tokens cost $0.40 per million for either variant, according to the company's launch announcem…
Unitary Foundation launched the QEC Challenge, a public GitHub leaderboard for quantum error correcting codes that now holds more than 1,500 verified codes from over 40 contributors. The board accepts parity-check matrix…
Liquid AI released d1, a decision model that returns calibrated probabilities across a fixed set of typed outcomes in a single API call with zero generated output tokens, available today as a hosted API under the model n…
Nandakishor M released Laya, an open-source library that replaces autoregressive LLM decoding with a 33-millisecond single-pass decision engine trained via reinforcement learning with strictly proper scoring rules (RLCD)…
A developer built a machine learning model that detects phishing websites from URL and page features with 96.7% accuracy, without opening the sites. Trained on the UCI Phishing Websites dataset of 11,055 labeled sites ac…
A developer released TurboGPT, an MIT-licensed byte-level GPT trainer written in CUDA C++ that trains a 22KiB transformer in 13 seconds. The tool builds on Windows with Visual Studio 2022 C++ tools and CUDA 13.4 via `bui…
Mark Vange of Autom8ly benchmarked seven open-source alternatives to TypeSafe's Jev decision model on a single 8 GB NVIDIA RTX 4060, adding 4-bit and 8-bit loading via bitsandbytes to three tools that only ship full-prec…
A hands-on test of Qdrant's scalar, binary, TurboQuant, and product quantization on 15,205 real Great Britain EV charging locations and a separate 100,000-point synthetic workload found that smaller vector copies cut mem…
Researchers introduced Allspark, a method that uses an alternating chain of thought to transfer reasoning improvements learned by a small, weak model to a larger, stronger model without generating large-model rollouts, a…
Google Research software engineers Chih-wei Hsu and Moonkyung Ryu introduced Diffusion Controller, a lightweight "steering damper" network that attaches to text-to-image models, including access-restricted closed-source …
OpenZL released v0.3.0, whose new LZ engine uses Marcin Żukowski's PivCo Huffman entropy backend to cut decompression time by 33% over v0.2.0 at compression level 1 with a 64KB window, reaching 3062 MB/s — 144% faster th…
Researchers submitted TokenCast, a method that learns a composable cost representation for each execution segment of an LLM agent run, forecasting token consumption before and during execution without additional LLM call…
Researchers Emiliano Penaloza and co-authors submitted KV-streams to arXiv on 28 Sep 2026, a plug-and-play context-compaction strategy for agentic reinforcement learning that streams the KV cache forward instead of flush…
OpenZL v0.3.0 introduces the Compression Transformer, a neural network that builds compression graphs automatically by selecting codecs one decision at a time, requiring no per-source training, no manual tuning, and no c…
A Peking University research team led by Haoran Ye trained the behavior of a specialized agent harness into Qwen3.5-9B via a method called harness distillation, raising the small model's average task success by 21 points…
A developer building a co-pilot for pediatric therapists found that a plain vector index ranked a four-month-old clinical note above one from 90 minutes earlier because embeddings encode semantic similarity, not recency …
NVIDIA released Kumo Tabular, an open foundation model for tabular classification and regression, on Hugging Face under the OpenMDW-1.1 commercial license, claiming first place on the TabArena, BeyondArena, TALENT and Sc…
Researchers introduced SCOPD (Sparse-Context On-Policy Self-Distillation), a method for efficient vision-language models that addresses the performance degradation caused by training-free token pruning of long visual-tok…
Qlty released SlopOne, a free research preview AI model that scores code maintainability by combining traditional static analysis statistics with classifications from TypeSafe AI's new Jev decision model, trained on huma…
Researchers present pruned CTC, a method that avoids materializing frame-by-vocabulary activations in memory to make large-vocabulary ASR training feasible. Conventional CTC implementations materialize these activations,…