To improve the faithfulness of model-generated reasoning, we study two other ways of eliciting reasoning from models in Radhakrishnan et al. These methods rely on question decomposition, or breaking down a question into smaller subquestions to help answer the original question.

→ View original post on X — @anthropicai