The reviewers are overwhelmly positive on the paper. The paper provides a theoretical analysis on why DARTS tends to bias towards architectures with skip connections and then leverage the theoretical insight to introduce a regularization term to reduce the bias. The paper is well written and has valuable contribution on both theoretical analysis and practical algorithm.