Segmenting X-ray sources in eROSITA sky tiles is a small-data, noisy-label problem: a detection threshold generates the training masks, and only eight sky tiles exist. What matters for a survey is generalization to unseen sky, so we judge every design choice on a held-out test region and by leave-one-region-out cross-validation, and use validation only to lock the decision threshold. A frozen evaluator scores foreground intersection-over-union (IoU) behind five guardrails, and we accept an effect as real only when the gap between two configurations exceeds their combined seed-spread. Testing one factor at a time (loss function, input channels, network architecture, augmentation, and hyperparameter search), we find that the loss is the one lever that beats the cross-entropy baseline on unseen sky: the Tversky loss raises foreground IoU from 0.33 to 0.37 past seed noise. Added model capacity, extra input channels, augmentation, and hyperparameter tuning show no measurable test gain. We trace this plateau to a label-noise ceiling set by the threshold-generated masks. The lever with real headroom is more data: pretraining on simulated eROSITA tiles and fine-tuning on the real ones gives the best score on the primary single-channel test region, whereas mixing the two sets directly collapses. Beyond one test region, leave-one-region-out cross-validation shows that controlling variance and overfitting generalizes: a smaller regularised model with seed-ensembling reaches 0.44 foreground IoU across held-out regions against 0.35 for the larger baseline and removes its regional collapse.

This paper studies a large language model (LLM), used through its chat interface with no API access, in three structurally different roles for budget-limited molecular optimization: a free-form inventor, a numeric regressor inside Bayesian optimization (BO), and a pairwise ranker inside preferential BO (PBO). The regressor and the ranker share the same SELFIES-VAE latent space and Gaussian-process acquisition machinery; only the LLM's role inside that loop changes. All three roles are evaluated under a fixed 15-evaluation budget on two chemically unrelated ground-truth oracles: aqueous solubility (ESOL) and JNK3 kinase inhibition probability. On ESOL, all three approaches comfortably beat both the seed bar and a random baseline, ranked by how much freedom each role gives the LLM (direct optimizer > regressor > ranker). On JNK3, that ranking collapses: only the free-form inventor still improves on the seed, while the regressor and ranker tie at exactly zero improvement, because the shared VAE essentially never decodes a real kinase-hinge binder near the seed region. The result is that the bottleneck in LLM-driven molecular optimization is not always the LLM's judgment quality; it is often the coverage of the generative model the LLM is paired with.

Speech recognition at the edge runs into three costs at once: converting sound from the analogue time domain into digital frequency features, extracting those features, and moving network weights between memory and processor. This paper reviews an architecture that attacks all three in the material itself. A reconfigurable nonlinear processing unit, a lightly doped silicon device working at room temperature, transforms raw audio in the time domain and reproduces three cochlear properties: frequency selectivity, tuneable nonlinearity and millisecond fading memory. A convolutional classifier trained on those features runs on an analogue in-memory computing chip whose phase-change memory crossbars perform matrix-vector multiplication where the weights are stored. The combined system reaches 96.2% on TI-46-Word spoken digits and 89.3% on Google Speech Commands keyword spotting, with feature extraction costing roughly 300 nJ per inference and the classifier adding 348 µs of delay. The two material systems remain physically separate chips.

The proliferation of large language models capable of generating syntactically correct code poses challenges for academic integrity and software provenance. We address SemEval-2026 Task 13, which formalizes machine-generated code detection as two sub-tasks: binary classification of human vs. AI-generated code (Task A) and multi-class attribution across 11 generator families (Task B). We propose the Fingerprint Effect hypothesis: stylistic artifacts such as whitespace, comments, and naming conventions are stronger discriminative signals than semantic code logic. We fine-tune CodeBERT on 20,000 stratified code samples, achieving an F1 of 0.9854 on Task A and a Macro F1 of 0.7214 on Task B, outperforming a TF-IDF baseline by 18.6% and 59.5% respectively. An ablation study validates our hypothesis, showing that even light preprocessing nearly doubles the binary error rate (1.46% to 2.80%) and causes a 3.59% Macro F1 drop in multi-class attribution.

Globally, chronic liver disease continues to be a major health concern that requires precise predictive models for prompt detection and treatment. Using the Indian Liver Patient Dataset (ILPD) from the University of California at Irvine's UCI Machine Learning Repository, a number of machine learning algorithms are investigated in this study. The main focus of our research is this dataset, which includes the medical records of 583 patients, 416 of whom have been diagnosed with liver disease and 167 of whom have not. There are several aspects to this work, including feature extraction and dimensionality reduction methods like Linear Discriminant Analysis (LDA), Factor Analysis (FA), t-distributed Stochastic Neighbour Embedding (t-SNE), and Uniform Manifold Approximation and Projection (UMAP). The purpose of the study is to investigate how well these approaches work for converting high-dimensional datasets and improving prediction accuracy. To assess the prediction ability of the improved models, a number of classification methods were used, such as Multi-layer Perceptron, Random Forest, K-nearest neighbours, and Logistic Regression. Remarkably, the improved models performed admirably, with Random Forest having the highest accuracy of 98.31% in 10-fold cross-validation and 95.79% in train-test split evaluation. Findings offer important new perspectives on the choice and use of customized feature extraction and dimensionality reduction methods, which improve predictive models for patients with chronic liver disease.

The purpose of this study is to determine the efficiency different categorization techniques work with sound samples in order to identify Parkinson's disease. The research explores classification strategies targeted at improving prediction performance in high-dimensional datasets by utilizing the UCI Parkinson dataset, which consists of 80 sound samples with 240 instances from 40 patients and 40 healthy individuals. This study focuses on the effectiveness of various classification strategies transform high-dimensional data and increase prediction accuracy. Key components of the methodology include Principal Component Analysis (PCA), Fourier and wavelet transforms, replication dependency accounting, and temporal feature extraction. KNN and logistic regression worked together to produce performance metrics and accuracy that were mediocre. Great performance was shown by Random Forest, especially when dealing with high-dimensional datasets. Remarkably, MLP achieved exceptional results of 95% in 10-fold cross validation, outperforming all other models in terms of accuracy, precision, recall, and F1-score.