Large language models (LLMs) are usually evaluated by final-output correctness or safety, yet an endpoint label does not reveal how a failure develops throug…
机构:中科院
来源:arXiv 2610.37369 | AI4Papers 论文推荐平台