AIVARI Agent: An Evidence-Grounded Agentic LLM for Variant Reportability and Interpretation

Clinical genomic variant interpretation is a patient-level multi-hypothesis reasoning task that integrates variant evidence, phenotype fit, inheritance, and database knowledge to determine reportability. We propose \textbf{AIVARI Agent} (AI VAriant Reportability and Interpretation Agent), an agentic LLM that performs one evidence-grounded rollout per retained candidate gene and jointly evaluates all associated gene-disease hypotheses. On a 300-case clinical cohort (6,460 hypotheses), AIVARI Agent achieves Group Sensitivity 0.905, Group NPV 0.933, and Row Precision 0.351. On a 235-case common subset, it improves over an operational hybrid pipeline by $+33$pp Group Sensitivity and $+40$pp Group NPV, with the largest gain on Inconclusive findings. These results support single-rollout agentic LLMs with on-demand evidence grounding.