SpecLLM feature-based VNIR representations improve selected LLM classification tasks

A prototype evaluation reported in a Scientific Reports abstract found that compact, feature-based representations of visible and near-infrared (VNIR) spectra substantially improved blood-substance classification for selected large language models compared with raw numerical encodings. The tested approach, named SpecLLM, remained less effective than a Random Forest operating directly on raw spectra.

SpecLLM was designed to give language models a concise representation of spectral evidence rather than complete spectra serialized as comma-separated values, JSON or XML. The method extracts summary statistics and discriminative extrema, then represents these features either in natural language or structured JSON. The natural-language format is human-readable and inspectable, and the abstract reports that it reduces prompt length relative to full-spectrum encodings. Natural-language and JSON feature representations were broadly comparable in the evaluated comparison.

The evaluation covered prototype-based substance classification in two complementary scenes drawn from an open-access VNIR hyperspectral blood dataset. This was a computational prototype study, not a review, clinical study, animal experiment, cell experiment or human-outcome trial. The results therefore describe classification performance for the selected models, scenes and dataset, rather than establishing that large language models outperform conventional machine-learning methods in other settings.

The available evidence includes the article abstract and publisher information. The abstract identifies the dataset type, evaluation scope and relative performance pattern, but it does not provide numerical performance measures, model configurations or the full paper’s stated limitations. Accordingly, the reported findings cannot be used to infer performance across additional datasets, substances or applications. The publisher describes the 17 October 2026 version as an early-access article that is citable, peer reviewed and subject to further edits before replacement by the final Version of Record.

Subject of Research: Prototype evaluation of feature-based VNIR representations for LLM blood-substance classification

Article Title: SpecLLM: LLM-based VNIR spectral interpretation

Article References: Romaszewski, M., Strzoda, A., Halama, M., Głomb, P., & Masłowski, P. (2026). SpecLLM: LLM-based VNIR spectral interpretation. Scientific Reports. https://doi.org/10.1038/s41598-026-75562-5

Image Credits: AI Generated

DOI: 10.1038/s41598-026-75562-5

Keywords: SpecLLM, visible and near-infrared spectroscopy, large language models, spectral interpretation, blood classification, feature representations, machine learning, Random Forest, hyperspectral data, LLM-based, VNIR, spectral