Abstract Spatial transcriptomics measures gene activity while preserving the locations of cells within tissue. A key analysis step is to divide the tissue into spatial domains, but the performance of computational methods varies considerably across datasets, and no single method consistently performs best. Here we show that a large language model can compare spatial domain maps and identify the better-performing methods. We developed L-STAR, which visually compares and ranks candidate methods, then builds a consensus map from the highest-ranked results. Across datasets representing different tissues, species, and technologies, L-STAR matched expert annotations more closely than individual methods or consensus analysis using a fixed set of methods. Its rankings were reproducible and generally stable when image colors or presentation order changed. L-STAR also provided rationales for its comparisons and assigned biologically meaningful names to the detected domains. These findings show that large language models can make spatial transcriptomics analysis more adaptive, robust, and interpretable.