EPSpec: An Evidence‐Guided, Prior‐Retrieval Agent for Near‐Infrared Spectral Band Selection

ABSTRACT To address band selection in high‐dimensional continuous near‐infrared spectra for quantitative analysis, the evidence‐guided, prior‐retrieval spectral interval selection (EPSpec) framework is proposed. The method partitions the full spectrum along the wavelength into a set of contiguous candidate intervals and computes multidimensional statistical indicators for each interval to construct a structured evidence table. At the same time, it introduces mechanistic prior knowledge based on the vibrational band evidence index (VIBEX) functional group–band prior library and task‐relevant mapping. Subsequently, under the joint guidance of interval evidence and mechanistic prior knowledge, a large language model ranks the candidate intervals by importance and outputs interpretable rationales. Finally, band selection and modeling are completed through Top‐K subspectrum search and nested cross‐validation without information leakage. Experimental results on three near‐infrared regression datasets show that EPSpec combined with partial least squares regression (EPSpec+PLSR) consistently improves upon full‐spectrum PLSR while substantially reducing the number of input wavelength variables and maintaining relatively low model complexity. Specifically, on the shootout, soil, and corn datasets, EPSpec+PLSR increases the coefficient of determination ( R 2 ) from 0.9298, 0.9428, and 0.8092 to 0.9372, 0.9657, and 0.9273, respectively, and reduces the root mean square error of prediction (RMSEP) from 4.4856, 2.5369, and 0.2833 to 4.2272, 1.9538, and 0.1671, respectively. Compared with traditional variable selection methods, EPSpec+PLSR achieves overall comparable or locally slightly better results. In particular, on the soil dataset, it achieves the best result while retaining only about 85/1050 wavelengths on average. Ablation experiments further show that interval‐level statistical evidence and mechanistic priors are key to stable performance gains, and that sliding‐window partitioning can further improve predictive performance.