Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new paper, "StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training," addresses training stability in vector-quantized visual tokenizers, which underpin autoregressive and masked image generation …
Vincent and Jochen, founders of Sitefire (YC W26), launched a Show HN project training a model to identify AI-generated web content using page structure alone. The founders, who have backgrounds in RL/optimization at Sta…
Researchers introduced the Kronecker coVariance Neural Network (KVNN), a temporal graph neural network that represents the spatiotemporal covariance matrix as a sum of Kronecker products to decouple spatial and temporal …
A new arXiv paper (arXiv:2609.25397v1) presents an end-to-end, size-agnostic graph reinforcement learning framework for one-dimensional bin packing that lowers the mean optimality gap of a constructive heuristic from 2.6…
A study of 3,600 exact-rational word problems and 8,600 prompts across five open-weight language models found canonical accuracy of 0.969-0.996 but orbit correctness of only 0.848-0.981 and orbit invariance of 0.851-0.98…
An independent researcher pretrained a roughly 0.4B-parameter Bangla-first language model end-to-end in Rust for $164 in rented GPU time, documenting five defects in the Candle framework and three in Burn as training bac…
A new arXiv paper (arXiv:2609.25012v1) reports that diachronic word embeddings can track semantic change in Sanskrit, an ancient low-resource language, with 19 of 21 testable shifts moving in the philologically attested …
AIBuildAI-2.5, an agentic system that uses LLM-guided tree search to automate AI model development, ranked first on MLE-Bench with a medal rate of 73.3%, according to the arXiv paper 2609.25047v1. The system adds a judge…
A 3x3 mathematical-reasoning experiment on on-policy distillation (OPD) found that prompt breadth and rollout refresh interact, producing a 4.07 percentage-point effect with a 95% question-paired interval of [2.00, 6.28]…
A new benchmark called FrontierMath Erdős (FME), introduced in a paper submitted to arXiv on 6 September 2026, evaluates AI systems on 68 Erdős problems that remained open as of August 2026, requiring them to prove or di…
Researchers demonstrated the first cross-model handoff of persistent recurrent inference state between differently sized hybrid language models without target prefix replay, transferring live memory from a Qwen3.5 4B mod…
ChainDoRA, a weight-decomposed parameter-efficient fine-tuning framework that builds directional low-rank factors from a connected Tensor-Train chain, reached a seven-task average accuracy of 72.30% on commonsense reason…
A new arXiv paper (2609.25066v1) introduces ReliMap, a framework that decomposes LLM-based human behavior simulation into three structured layers and evaluates reliability at the individual level (R1) and population leve…
Token-matched experiments reported in arXiv paper 2609.22161v1 found that medical large language models trained on clinical data improve clinic-oriented tasks while staying competitive on knowledge-intensive ones, wherea…
A new arXiv paper (2609.22475v1) proposes a goal-driven approach to process variant categorization in which an organization's goal model is authored first to predefine the categorization axis, and a Large Language Model …
A population-supervised framework developed by researchers maps single 2D red-blood-cell images to latent biophysical quantities and aggregates them into mean corpuscular volume, red-cell distribution width and mean corp…
AutoGym, a framework detailed in arXiv paper 2609.22592v1, generates complete reinforcement-learning gyms — tasks, executable environments, and verifiers — from a minimal domain seed or prior model trajectories. The fram…
Researchers introduced GaitVista, a reliability-aware measurement layer that reduces average full-body and lower-body gait measurement error by 27.7% and 27.8% across seven clean and degraded sensing conditions on the To…
A study on arXiv (2609.22694v1) reframes dementia-related crash severity prediction as a decision-aware triage problem, using 4,781 Texas crash records with structured fields and police narratives under a stratified 70/1…
A new arXiv paper (2609.22691v1) proposes a human behavior control framework that jointly models goal-directed and habitual behaviors for embodied human-like agents, with a Habitual Controller retrieving cue-triggered be…