It would be nice if we could conduct more extensive robust analyses in terms of the following perspectives:

• re-evaluate model on adversarial evaluation set
• evaluate the faithfulness of solution chains
• more others