GENEB: Why Genomic Models Are Hard to Compare

Frozen representations from 40 genomic foundation models evaluated across 100 tasks in 13 functional categories reveal that aggregate leaderboard rankings are unstable — model rankings shift sharply across task categories, and scale yields only modest, inconsistent gains. GENEB uses a unified probing protocol including few-shot regimes to enable controlled comparisons across architecture, tokenization, and pretraining data choices.