Machine Learning News

A study of PRM-Pruned Fragment Grafting (PPFG) on Qwen2.5-7B-Instruct with Math-Shepherd on full MATH500 (n=500, three seeds) found the inference-time intervention statistically indistinguishable from an independent para…

A study posted to arXiv (2610.00314v1) found that predictive credit for research-agent explanations remained unconfirmed across 336 prospective states in controlled learning, 12 Tox21 endpoints, and 24 OpenML tasks, with…

Fine-tuning large language models on math problems that share a reasoning approach with the target task outperforms training on same-topic problems, according to an arXiv paper (2610.00331v1) that tested two counterbalan…

A new arXiv paper (2609.38194v1) proposes a moving-horizon approximate branch-and-reduce method for training near-optimal deep classification trees on large-scale datasets with continuous features. The method combines a …

A data-free physics-informed neural operator for level-set interface advection, trained only on the transport residual and a geometric constraint with no reference solutions, reached 1.614 +/- 0.067% relative L2 error on…

A controlled experiment with 25 participants riding an instrumented e-scooter found that an entropy-based multi-class logistic regression classifier distinguished sober from alcohol-intoxicated riding with 85% overall ac…

A new arXiv paper (2609.38332v1) introduces Hermes, a family of configurable harnesses that shift decisions about context allocation and reuse from the harness to the model, plus Hermes-Learn, a two-stage framework for l…

Researchers introduced Continually-Recyclable Unit-Gating (CRUG), a continual learning method for dynamical systems reconstruction that achieves zero forgetting while conserving network capacity, according to an arXiv pa…

A new arXiv paper (2609.38342v1) introduces Activation-Conditioned Self-Distillation (ACSD), a method that extracts a steering vector by contrasting activations of self-generated trajectories reaching verified correct an…

Researchers introduced the Function-Space Transformer (FST), a framework that learns from functions via a spatially adaptive continuous latent representation, storing features at anchors whose locations are predicted fro…

A paper posted to arXiv (2609.38363v1) proposes Simulator-Refined Diffusion (SRD), a method that combines a low-fidelity differentiable surrogate with a high-fidelity non-differentiable full-wave electromagnetic simulato…

Researchers introduced GENSCRIPT, an inference-only synthetic data pipeline that eliminates model training by computing a deterministic statistical profile of source data and passing it to a language model to infer field…

Researchers introduced Spatial Lifting (SL), a method that lifts 2D image inputs into a higher-dimensional space and processes them with networks such as a 3D U-Net, according to the arXiv paper 2610.00017v1. The authors…

Researchers introduced MRFFU-Net, a deep learning architecture that adds a Multi-Resolution Feature Fusion (MRFF) module to U-Net-like encoder-decoder structures for MRI segmentation, according to an arXiv paper (arXiv:2…

Researchers introduced Gaussian Process Embedding Correction (GPEC), a pre-LLM error-correction method that improves cardiac ultrasound caption generation by VideoChat2, according to an arXiv paper (arXiv:2610.00196v1). …

A new arXiv paper (arXiv:2610.00294v1) introduces LENS-GRF, a permutation-invariant Lesion Evidence Network with Gated Residual Fusion for four-class acne severity grading. On the ACNE04 dataset, fully automated LENS-GRF…

A new arXiv paper (2610.00372v1) introduces the Causal Intervention Router (CIR), a lightweight policy that decides when an LLM agent's error-recovery harness should intervene, raising success on long-horizon ALFWorld ta…

A new arXiv paper (2610.01014v1) derives a sharp recoverability threshold for LLM voting under a fixed call budget, characterizing a "discovery-to-decision gap" in which a correct answer found early can still fail to acc…

A new arXiv paper (2610.00663v1) proposes Quarantined Expert Shutdown (QES), a backdoor-containment method that lets a backdoor form during training but routes trigger-conditioned behavior into a designated, quarantined …

Researchers introduced CompMat-Bench, a benchmark of 94 tasks derived from recently published computational materials studies, to evaluate AI agents on scientific research steps without rerunning expensive simulations. A…