LLM "Reflection" Often Fails to Improve, Unlike Human Revision

This research introduces the Human-LLM Reflection Framework (HRF) to systematically compare how humans and large language models (LLMs) revise their answers. The study reveals that while humans consistently improve their responses through reflection, LLMs often fail to gain information. On objective tasks, LLM reflection is akin to re-sampling, yielding minimal improvement. For subjective tasks, it can even lead to worse outcomes. The findings suggest that LLM "reflection" is more accurately described as conditioned re-generation rather than genuine error-driven revision. This limitation stems from the models' inability to reduce uncertainty about the target without external information, even when provided with high-quality initial responses. The failure is localized to the revision step itself, highlighting a fundamental difference in how humans and current LLMs approach iterative improvement.