LLM-Driven Inference of Demographic History: The GHIST 2026 Tournament Results

ABSTRACT

Inferring complex evolutionary and demographic histories from genomic data is plagued by model misspecification and unconscious bias. We report the results of the 2026 Genomic History Inference Strategies Tournament (GHIST), which introduced a new challenge category: LLM-Driven Inference Agents. Competitors deployed autonomous agents to analyze site frequency spectra and linkage disequilibrium patterns, iterative building and testing demographic models. The results demonstrate that Agentic LLMs, capable of dynamically adjusting parameters based on statistical feedback, significantly outperformed traditional static machine learning approaches in resolving complex secondary contact and archaic admixture scenarios.