Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting thei…
机构:Microsoft
来源:arXiv 2609.35706 | AI4Papers 论文推荐平台