CONSTRAINED LANGUAGE-GUIDED REFINEMENT FOR ZERO-SHOT SPATIAL ANNOTATION

Spatial transcriptomics enables the analysis of cellular organization by measuring gene expression in situ, but assigning coherent spatial region labels remains challenging across platforms due to heterogeneous resolution, incomplete marker panels, and ambiguous boundaries. Existing approaches typically rely on super-- vised training, dataset-specific tuning, or deep graph models, which can oversmooth structure, generalize poorly across technologies, and offer limited interpretability. We introduce NicheAgent, a training-free structured prediction framework that casts spatial annotation as a constrained decision problem with selective language-based verification. NicheAgent first performs deterministic prototype-based assignment using curated region prototypes (“nichecards”) encoding canonical marker genes and expression centroids. Only for low-confidence cases, a lightweight large language model (LLM) is invoked as a closed-world verifier, arbitrating among a fixed set of candidate labels using marker semantics and local neighborhood context under a strict ontology. A single round of spatial smoothing enforces local coherence without blurring anatomical boundaries. Across Visium, MERFISH, and STARmap datasets, NicheAgent consistently out- performs supervised, graph-based, and prior LLM-driven methods on standard spatial annotation metrics, while remaining transparent and interpretable. More broadly, our results highlight a general design pattern in which LLMs act as con- strained adjudicators over symbolic hypotheses, improving structured prediction in high-ambiguity settings without end-to-end learning or loss of interpretability.