PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs

发生了什么

Hallucinated information can propagate through multi-stage LLM systems and become part of the context for subsequent reasoning. Existing studies of post-hallucination reasoning (PHR) mainly characterize changes in final outcomes and aggregate reasoning dynamics, leaving how models resolve hallucinat

下一步怎么做

保存这条信号，打开原始来源核验关键细节；如果它与你当前项目相关，再和同主题的相关信号一起判断趋势是否持续。