在大型语言模型中诊断多重上下文知识更新下的检索偏差

Large Language Models (LLMs) are extensively utilized in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike previous studies that focus on one-shot updates or single conflicts, multi-update scenarios involve multiple historically valid versions that compete during retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, a Dynamic Knowledge Instance (DKI) evaluation framework is introduced, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assessing models through endpoint probing of the earliest versions.