Researchers have developed a method to improve the reliability of large language models (LLMs) in scientific decision support tasks. They focused on taxonomic reasoning, where LLMs are used to classify species based on characteristics such as morphology and genetics. The team used proxy features extracted by an open-source tool to train lightweight supervised estimators that can predict the correctness of LLM outputs at different levels of a hierarchical taxonomy. Their method outperformed a baseline approach in predicting correct classifications, with improved results achieved when considering the hierarchical structure of the output. The code for this research is publicly available. --- Why it matters: This work matters to engineers and researchers working on AI because it addresses a critical challenge in using large language models for scientific decision support: estimating their reliability. By improving uncertainty estimation, this method can help reduce errors and increase trust in LLM-based systems. Source: https://arxiv.org/abs/2608.22839