Large language models struggle to reliably identify and correct errors in their own reasoning. Unlike prior training-based approaches that fine-tune an LLM t…
机构:KAIST
来源:arXiv 2610.13672 | AI4Papers 论文推荐平台