Genomic language models fail the Mechanistic Invariance Test because they exploit AT content correlations rather than positional grammar, while a 100-parameter position-aware PWM achieves perfect scores.
15:−3.70) is∼8×largerthan the maximum positional effect (0.46)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
q-bio.GN 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
The Mechanistic Invariance Test: Genomic Language Models Fail to Learn Positional Regulatory Logic
Genomic language models fail the Mechanistic Invariance Test because they exploit AT content correlations rather than positional grammar, while a 100-parameter position-aware PWM achieves perfect scores.