Researchers studied the trade-off between intrinsic self-correction in language models and the potential for introducing errors. They found that refinement can improve accuracy but also change correct answers into incorrect ones, and propose selective invocation of revision as a better approach.