Machine Learning News

VIDRAFT released Darwin-180B-RSI, a 180-billion-parameter reasoning model trained with a weight-level Recursive Self-Improvement framework that updates model weights directly without human-annotated chain-of-thought. The…

Sktime released TServe, an open-source inference server that loads time-series foundation models including Chronos, TimesFM, Moirai, TTM, and TiRex once into memory and serves forecast requests over HTTP via a single POS…

Laya released a multilingual, non-autoregressive "System 1" decision engine that returns typed decisions across 100+ languages in a single forward pass in 33 ms, installable via `python -m pip install laya` on Python 3.1…

A hands-on one-semester course walks students through building a machine-learning compiler from scratch, progressing from elementwise operations to training a real LLM on GPUs. Over roughly ten weeks, students implement …

A September 30, 2026 arXiv paper (2609.39838) finds that large language models learn steganographic reasoning — concealing reasoning inside innocuous-looking text — far less readily than the neighboring capabilities of s…

Allegro's AlleCompanion machine-learning retrieval system lifted attributed gross merchandise value by as much as 21.25% in the in-cart placement but left GMV flat in the pre-cart step immediately before it, according to…

The Chinese Academy of Sciences' Institute of Artificial Intelligence for Industries released Maxwell, an embodied AI model that scored 91.9 on the Meta-World benchmark for robot physical tasks, the highest score ever re…

NVIDIA released Kumo Tabular, a family of tabular foundation models in Small, Medium, and Large sizes spanning roughly 28M to 215M parameters that predict new rows in a single forward pass with no training, hyperparamete…

TypeSafe AI, a San Francisco lab founded by former OpenAI researcher Diogo Almeida, released Jev, a decision-only model that returns typed probabilistic answers instead of generated text, priced at $0.042 per million inp…

David AI's DAI-ASR-I18N benchmark, built by CEO Tomer Cohen and CTO Ben Wiley, compared 14 speech-recognition systems on unscripted conversation across 21 languages and found error rates varied by an average of 3.7 perce…

Researchers introduced Ataraxos, an AI for the hidden-information board game Stratego, which defeated the most decorated human Stratego player of all time by a large margin, achieving what the Nature paper calls the firs…

NVIDIA published a fine-tuning workflow for its Nemotron 3.5 ASR streaming 0.6B model that adapts the 40-language-locale speech recognizer to Saudi Najdi and Hijazi dialects using the NeMo framework, the SADA 2022 and FL…

EXPLORENET, a learned adaptive exploration policy for group-relative diffusion RL, improves held-out GenEval2 by 14% over Flow-GRPO on Stable Diffusion 3.5 Medium and reaches a 67.2% human preference win-rate, according …

TrackFish3D, a geometry-driven self-supervised framework for dense multi-camera 3D tracking of schooling fish, improves 3D Multi-Object Tracking Accuracy from 87.7% for the strongest baseline to 95.8% on its benchmark, a…

An AI agent rediscovered a degree-four Blaschke-curve invariant from numerical data, fitting a homogeneous cubic whose frozen coefficients predicted 480 lines from 80 unseen parameter values with a recorded RMS scale-fre…

Researchers proposed MoFlow, a method that generates agentic workflows optimized across multiple objectives at once by formulating workflow generation as a multi-objective Markov decision process and solving it with Conv…

AREX-2, an LLM agent built on Qwen3.8-27B and trained on synthesized long-horizon improvement trajectories from machine learning and algorithmic programming tasks, scored 81.8 on MLE-bench Lite and 70.7 on Frontier-CS, a…

Researchers introduced EVOKE, a post-training method that elicits pretrained world knowledge in LLM agents by holding environment state and interaction history fixed while ranking the same candidate actions under alterna…

DEdit, a diffusion-based speculative decoding drafter from an arXiv paper (arXiv:2609.38510v1), achieves macro-average speedups of 5.72x on Qwen3-4B and 5.97x on Qwen3-8B over autoregressive generation under greedy decod…

LoopVL, a recurrent vision-language model that combines Module-Loop and Model-Loop computation to iteratively update a unified vision-language state through shared modules, outperforms similarly sized and larger non-recu…