Establishing correlations between laser-scanned pavement texture features and skid resistance using machine learning models is crucial for rapid assessment of road safety performance. However, due to measurement limitations, datasets in related studies are often constrained in size, adversely affecting the predictive accuracy of the models. To address this issue, Tabular Prior-data Fitted Networks (TabPFN), a model integrating pre-training and in-context learning, is applied in this study due to its optimization for small-sample tabular datasets. TabPFN exhibits robust generalization capabilities for unseen data by leveraging extensive pre-training on millions of datasets. In this study, 500 paired samples of 3D laser-scanned pavement texture data with corresponding friction coefficients were collected through field measurements. From the 3D data, 19 geometric features, 18 multiscale features based on 2D wavelet decomposition, and a reflectivity feature were extracted. Comparative evaluations between TabPFN variants and state-of-the-art machine learning baselines revealed that all TabPFN variants consistently outperformed existing methods. Notably, the hyperparameter-optimized TabPFN achieved the highest predictive accuracy (R² = 0.902), surpassing the best-performing baseline model by 3.7%. Feature importance analysis identified Mean Intensity, Energy of 2D-wavelet decomposition at level 6 (3.2–6.4   mm), and S ku (kurtosis of height distribution) as the three most influential predictors, representing key indicators across their respective categories. This study successfully improves pavement friction coefficient estimation by applying a large-model framework to tabular data, validating its engineering applicability and offering a viable solution for data-driven challenges with limited sample sizes.