Tag
The study finds that rank truncation in matrix-CODI models does not impact accuracy on reasoning tasks, indicating that rank may not correlate with parallel reasoning paths in latent representations.