Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers introduced PhyMo, a physics-grounded multimodal framework that organizes heterogeneous measurements through PDE-associated operators, according to an arXiv paper (arXiv:2609.27554v1). PhyMo uses a three-stage…
Researchers released TNLearn, an open-source Python package that automates the construction of task-based neurons and networks and supports training of task-based networks, according to the arXiv paper 2609.27564v1. TNLe…
Researchers posted arXiv:2609.27244v1, a paper introducing a full-covariance smoothing method for Bayesian neural networks that propagates Gaussian moments through nonlinear activations via a cross-covariance identity, r…
A study of 2.6 million real option contracts found tree-based ensemble models detected deviations from Black-Scholes pricing with 93.8% accuracy, beating Kernel PCA dimensionality reduction by 21.5 percentage points (72.…
Researchers proposed LLMAE, a method that repurposes a pretrained decoder-only language model as a continuous text autoencoder by exposing a fixed-length latent bottleneck within its internal activations. Instantiated wi…
FWBench, a new benchmark detailed in arXiv paper 2609.27385v1, evaluates how language-model agents select and use time-series foundation models (TSFMs) for operational decisions across 1,251 electricity and cycle-hire ca…
SR-Fraud, an outcome-supervised reflective LLM agent framework described in arXiv paper 2609.27287v1, improves all detection metrics over its frozen decision agent and posts higher point estimates than static and periodi…
Researchers introduced the Neurogenesis Network (NGN), a differentiable parameterization that learns how many ordered structural components a model should use by training a single learnable boundary per component group t…
Researchers introduced DualRes, a compact oscillatory state-space model with 39,528 parameters for machine fault diagnosis from vibration data, reporting state-of-the-art performance among nine evaluated methods at six o…
Researchers introduced KV-Invariant Transformer Expansion (KITE), a scaling paradigm that trains a model from a smaller to a larger size while placing newly added parameters in regions that do not affect attention KV, ac…
Researchers introduced Pheno-GS (Phenoscape-scale Geodesic Sinkhorn), a method that computes geodesic optimal transport distances between all pairs of patient single-cell datasets in a single heat diffusion, running over…
Researchers introduced MENO (Memory-Efficient Neural Operator), a PDE neural solver built on the Manifold Function Encoder, in arXiv paper 2609.27739v1. MENO's memory footprint is independent of data resolution, it accep…
A new arXiv paper (2609.27355v1) proposes a quantization-robust unlearning framework that keeps forgetting intact after LLM compression by targeting sensitive weights identified through a curvature-based criterion. The m…
A new installment in the "How To Scale Your Model" series lays out a theory of sharded matrix multiplication for training large ML models across thousands of accelerators, using named-axis notation to describe how arrays…
Repository-level coding benchmarks such as SWE-bench suffer from data leakage because they are built on popular open-source repositories repeatedly used for LLM training, meaning strong performance may reflect memorizati…
Researchers introduced WhatWorkedBench, a benchmark that measures "experimental understanding" in AI agents by testing the accuracy of their predictions about component changes after budgeted experimentation. Agents insp…
A September 22, 2026 arXiv paper reports that greedy decoding in large language models is not precision-invariant: the same model, prompt and decoding algorithm produced different outputs in BF16 versus FP16 on identical…
An independent review of TypeSafe AI's Jev, a hosted "System One" model launched September 15, 2026 that returns typed answers (choice, score, noul) rather than generated text, finds it level with mid-price LLMs but 6.5 …
A developer built a next-character language model from a neural cellular automaton — a 16×16 grid of cells holding 16-number vectors updated by one shared 3×3 convolution rule, with no attention or transformer blocks — a…
Apple researchers compressed an on-device streaming neural audio tokenizer by 2.8× using latent-space distillation, keeping the distilled student within 1.9% relative word error rate of its teacher on five of six teacher…