Knowledge Leakage

A phenomenon in which a large language model simulating a low-knowledge user cannot fully isolate its pretrained knowledge, causing response bias. It manifests as systematic score differences between real entries and fictional twin entries.