Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A paper submitted to arXiv on 28 Sep 2026 introduces "harness learning," a method that trains a proposer model via reinforcement learning to revise a solver language-model agent's harness — the executable program organiz…
TensorViz launched as a VS Code tool that lets developers explore 29 interactive PyTorch architectures, including ResidualClassifier, without installing Python or any additional software. The tool offers guided tours of …
A developer implemented a FlashAttention kernel in CUDA on an NVIDIA RTX 4050 GPU, achieving a 41.6x speedup over standard attention by keeping computation on-chip in SRAM instead of materializing the N×N attention matri…
A paper posted to arXiv (2610.12845v1) proposes formal postulates and a conceptual mathematical definition of creativity grounded in Boden's three criteria — newness, value, and surprise — and validates the formula in tw…
ByteDance deployed ByteExpress, a software communication layer over hardware RDMA, in its parameter-server-based inference and training systems, cutting mean and P99 model parameter fetch latency by 35.7% and 43.7% respe…
A new arXiv paper (2610.12776v1) presents a coordinate-free automatic differentiation operator that expresses geometric differential operations entirely in terms of existing AD primitives, addressing the awkwardness of s…
CacheScope, a training-free method for accelerating diffusion transformer inference, achieves up to 3.57x acceleration and reduces median LPIPS by up to 70.1% over the strongest probe-based baseline, according to an arXi…
A new arXiv paper (2610.12538v1) proves that Cox regression can be efficiently learned under outcome-dependent truncation and right-censoring using first-order optimization when the baseline hazard is known, despite the …
Researchers introduced algebraic integer quantization (AIQ), a lossy compression format for large language model inference that achieves competitive accuracy against the industry-standard MX format at equivalent bit-per-…
Researchers developed CircuitTree, a data structure built on a Θ(N log N) tree-compose schedule, to compose arbitrarily many neural circuits into a single model, according to an arXiv paper (2610.12520v1). Composing circ…
Researchers posted arXiv:2610.12679v1, introducing a one-step implicit sampler (OSIS) that uses short-time MCMC dynamics to define a local target for an implicitly modelled distribution, which in the asymptotic limit bec…
Researchers combined holomorphic neural networks with geometry-derived singular functions to solve 2D and 3D Laplace boundary-value problems, reporting that the enriched model was the most accurate method tested on every…
Researchers proposed Sample-Adaptive Zeroth-Order Optimization (SAZO), a zeroth-order fine-tuning method that learns sample importance online and allocates queries adaptively, in a paper published as arXiv:2610.12566v1. …
A benchmark of statistical, machine-learning, recurrent neural-network, and pretrained TimesFM models on a university building's day-ahead electricity demand found that TimesFM 3.0 with covariates achieved the lowest ove…
A new arXiv paper (2610.12544v1) establishes end-to-end statistical guarantees showing that transformer-based masked diffusion models (MDMs) learn structured distributions with sample complexity governed by latent struct…
A team of researchers has built an all-transistor physical-learning transformer, detailed in arXiv paper 2610.12571v1, that restores the bidirectional information flow broken by nonlinear components in equilibrium propag…
DiffusionART, a new reinforcement-learning fine-tuning method for diffusion models from the arXiv paper 2610.12572v1, reaches the late-stage mean reward levels of DiffusionNFT and Flow-GRPO-Fast on HunyuanVideo-13B using…
A new arXiv paper (2610.12586v1) presents a benchmark for multi-horizon event sequence forecasting that measures both prediction accuracy and distributional realism, finding no existing method excels at both and revealin…
STAPLE, a recipe built on STEM for integrating static sparse memory into pretrained LLMs, improves over continued dense training on OLMo-2 1B by 4.27 points on HumanEval, 2.80 on MBPP, 2.27 on GSM8K, and 1.65 on DROP, ac…
A new arXiv paper (2610.12651v1) introduces SynAct, an active learning framework that allocates scarce expert feedback to individual agent decisions rather than whole tasks, ranking them by the mutual information between…