Read what ships this week.

Bayesian optimization via Gaussian Process surrogate — installs with pip install bayesian-optimization, runs in <5s with no GPU.

A from-scratch NumPy implementation of 20+ ML algorithms with runnable example scripts.

micrograd is a ~200-line pure-Python autograd engine + neural net library that backpropagates through scalar operations.

FedSIR is a federated learning framework that identifies noisy labels and clients by analyzing spectral properties of feature representations, then uses clean clients to help relabel corrupted samples. The method combines spectral analysis with noise-aware training strategies to achieve robust collaborative learning despite label noise.

This paper proposes CS-ARM-BN, a meta-learning method that uses negative control samples to adapt deep learning models to new experimental batches in biomedical imaging, closing the domain gap caused by batch effects. The approach achieves 0.935±0.018 accuracy on drug mechanism-of-action classification, recovering performance from 0.862±0.060 back to near training-domain levels of 0.939±0.005.

This paper presents a global dataset of Sentinel-1 SAR time series tracking offshore wind infrastructure deployment and operations from 2016-2025, with 15,606 time series and expert-annotated benchmarks for monitoring wind farm construction and operational dynamics. The dataset enables independent, high-resolution global monitoring of the rapidly expanding offshore wind sector.

Stream-CQSA enables exact self-attention computation on long sequences by decomposing attention into independent subsequence computations that fit within arbitrary memory budgets, allowing billion-token sequences to run on a single GPU without approximation. This removes the assumption that full query, key, and value tensors must fit in device memory, addressing the quadratic memory bottleneck of standard attention.

This paper reveals that different language model architectures (Transformers, RNNs, LSTMs) converge on learning similar periodic number representations with periods at 2, 5, and 10, despite being trained differently. The authors identify a two-tiered hierarchy of these features and explain when models learn geometrically separable representations useful for modular arithmetic.

This paper introduces gauge-equivariant graph neural networks that embed local gauge symmetries directly into message passing for learning on lattice gauge theories. The approach enables principled machine learning for systems with site-dependent symmetries and nonlocal observables.

This paper integrates working memory constraints into Transformer models through cognitively-inspired attention mechanisms and shows that these constraints improve grammatical accuracy and human alignment when training data is limited. The results suggest that mimicking human cognitive limitations can serve as a useful inductive bias for learning robust language representations.

This paper develops a physics-conditioned neural network to complete incomplete ice-layer thickness measurements from radar data by synthesizing missing layer traces. The approach combines geometric and temporal learning with physical climate model features to recover fragmented or absent layers while maintaining consistency with observed data.

This paper proposes a method to estimate long-term treatment effects and lifetime value changes in A/B tests for streaming platforms where user churn is costly. It uses inverse-variance weighted estimation across multiple cohorts and parametric decay modeling to capture both steady-state impact and cumulative user value within short experiments.

This paper presents a personalized BEV energy consumption framework that combines driver behavior prediction using Bidirectional LSTM with physics-based energy modeling and map contextual features. The system accurately estimates State-of-Charge depletion by capturing individual driving patterns across diverse road conditions.

This paper proposes semantic stratification, a structured evaluation framework for retrieval systems that organizes documents into entity-based clusters and systematically generates queries to ensure comprehensive coverage. It addresses hidden biases in current heuristic evaluation approaches and provides formal guarantees for more trustworthy retrieval assessment.

V-tableR1 is a reinforcement learning framework that trains multimodal language models to perform rigorous, step-by-step reasoning on tables rather than relying on pattern matching. It uses a critic model to provide feedback on visual reasoning chains and a novel optimization algorithm (PGPO) to improve performance.

This paper proposes a lifecycle-aware federated continual learning framework for distributed autonomous fleets that addresses both immediate forgetting during training and long-term cumulative drift through layer-selective rehearsal and rapid recovery strategies. The approach is validated theoretically and empirically, including on a real rover testbed.

AAC introduces a differentiable landmark-selection module for ALT shortest-path heuristics that maintains admissibility by construction through row-stochastic mixtures of triangle-inequality bounds. The method achieves near-optimal landmark coverage on road networks while being 1.2–1.5× faster than classical farthest-point sampling at matched memory.

F²LP-AP is a training-free label propagation method that adapts to both homophilous and heterophilous graphs using geometric median prototypes and local clustering coefficients. It achieves competitive accuracy with GNNs while being significantly more computationally efficient.

This paper proposes a Simulation-Based Inference (SBI) framework using neural networks for fast Bayesian condition monitoring of industrial equipment, specifically heat exchangers. The approach achieves 82× speedup over MCMC while maintaining comparable diagnostic accuracy and uncertainty quantification.

Near-Future Policy Optimization (NPO) improves reinforcement learning with verifiable rewards by learning from a policy's own future checkpoints, which are both higher quality and closer to the current policy than external sources. The method achieves significant performance gains on vision-language models, improving from 57.88 to 63.15 on Qwen3-VL-8B-Instruct.