Political polarization reflects systematic differences in how individuals process identical political information. Although motivated reasoning has been linked to biased evaluation, it remains unclear whether partisan bias emerges during early semantic encoding or later stages of rational evaluation. The present study addresses this question using an existing fMRI dataset in which strongly left- and right-leaning participants evaluated structured political arguments. We apply representational similarity analysis (RSA) to examine how arguments are encoded within language-related and reasoning-related neural networks. Model representational dissimilarity matrices based on semantic embeddings and bias-corrected argument strength are compared to neural representational geometries separately for attitude-congruent and attitude-incongruent arguments.