Brainstorm the role of different position embeddings and how they are affected by module coverage.

[purvapruthi]

• Why are RGs better than absolute embeddings?
• Why are both affected differently by module coverage?
• Is there a way to analyze the latent space of different compositions and measure the distance between them to characterize the CG behavior in diverse/uniform and direct/SBS models?