Reflection or Re-Generation? Why LLM Revision Fails Where Human Revision Succeeds

Researchers studying large language model reasoning found that current methods of prompted reflection often fail to mirror the iterative improvement seen in human revision. While these models are frequently instructed to revisit their prior steps, the study suggests they often struggle to achieve the same cognitive gains, highlighting a disparity between simulated self-correction and actual analytical refinement.