data-augmentation

This paper investigates scenarios where adding correctly labeled data can harm model performance, introducing insertion-stability and examining the limits of dimension-based theory in machine learning generalization.

This paper proposes a geometric filtering framework that selects high-quality LLM-generated samples by evaluating their Euclidean distance to real class examples in an embedding space, improving few-shot text classification performance by +2.61 percentage points over SMOTE.

This paper proposes a simulation-based methodology to generate augmented traffic datasets by replacing physical sensors with virtual ones, extending sensor coverage in urban networks while preserving traffic patterns.

This paper introduces techniques for the second MLC-SLM Challenge, including random leading-silence cropping and synthetic data generation to enhance multilingual conversational speech tasks, achieving improved accuracy and reduced error rates.

The paper introduces S^2VOPD, a self-supervised method that improves vision-language models by distilling from original images into strongly augmented student views, achieving performance surpassing GPT-5.4 on benchmarks.

TailBooster is a dual-layer generative framework that synthesizes operationally valid extreme air-transport events using statistical tail extraction and autoencoder-based cleaning, significantly improving extreme-event prediction accuracy.

This paper systematically evaluates 15 machine learning models, including the TabPFN foundation model, for post-wildfire debris-flow prediction using USGS basin-scale data, finding TabPFN achieves the best performance (threat score 0.637) and that synthetic data augmentation improves most models.

This paper presents a novel unsupervised data augmentation method combining Gaussian Mixture Models and Large Language Models to improve clustering on imbalanced text datasets by generating synthetic documents for underrepresented clusters.

Proposes a GAN-based framework (GT-GAN) for synthesizing high-fidelity data from incomplete LEO satellite Internet observations, showing robustness even with 40% missing data.

This paper presents a three-stage neural-symbolic pipeline for gaming toxicity detection, combining transformer ensembles with rule-based mediation, achieving top accuracy in the EEUCA 2026 shared task.

This paper proposes a cross-lingual data augmentation strategy that uses machine translation to transfer expert-annotated difficulty labels from high-resource languages to low-resource languages. Experiments with BERT-based regression models show that augmenting scarce native data with translated corpora significantly improves the accuracy of text difficulty assessment.

Introduces K-IPO, a generate-then-select oversampling framework that preserves the original data's feature importance ranking (measured by Kendall's tau) during augmentation for imbalanced tabular data, showing improved preservation, explanation consistency, and predictive performance across 20 datasets.

Researchers from MIT and others developed GIFT, a system that teaches vision-language models to automatically convert 2D designs into accurate CAD programs for rapid prototyping, using model-generated data to correct mistakes and improve performance with less computation.

This paper evaluates the effect of frame sampling rate on sequence-based classification of autism-related self-stimulatory hand behaviors using LSTM and GRU models, achieving up to 98.75% accuracy at a 15-frame interval, and analyzes data augmentation strategies for small behavioral video datasets.

This paper isolates Fixed-Source Synthesis (FSS) from Source Expansion in synthetic data scaling, proposing a rectified scaling law that predicts performance at high budgets from low-budget fits. Empirical results show FSS is bounded and that adding seed questions outperforms increasing response budgets at large scales.

The article compares the text-albumentations library to the new Autodata paper, which adds review mechanics with a weak/strong resolver and an external judge to maintain synthetic dataset quality.

AGVBench is a reliability-oriented benchmark for data augmentation in vein recognition, evaluating 30 augmentation strategies across multiple datasets and backbones, revealing decoupling between accuracy and security.

This paper investigates the mechanisms behind self-alignment methods in diffusion transformers, revealing that performance improvements from methods like Self-Flow primarily come from data augmentation along the noise dimension rather than token interactions between noise levels. The authors introduce Attention Separation to demonstrate this and propose an effective design combining self-representation alignment with dual-timestep augmentation.

Introduces the capability slice, a unit for linking evaluation failures to data interventions in LLMs, enabling a closed-loop process that diagnoses and fixes model weaknesses. Demonstrated on two case studies, showing recovery from training regression and significant math reasoning improvements.

This paper investigates the distributional gap between synthetic and real speech in LLM-based ASR systems, identifies where the LLM separates them, and proposes using layer-selection and RIR augmentation to match real-data baselines with less real data.