LLMs Struggle with Fact Recall Due to Entity Order

Google researchers have published a study that reveals why Large Language Models (LLMs) struggle to recall specific facts. The team found that while LLMs encode around 95-98% of tested facts, they often fail to directly recall 26-34% of them.

The issue is not that LLMs lack sufficient information; rather, the bottleneck lies in accessing that information. One major reason for this difficulty is the order in which subject and object entities are presented in a query.

According to the study, when questions reverse the subject/object entity order from what was encountered during training, LLMs have more trouble recalling the relevant fact. This phenomenon is observed even among strong models like GPT-5.2, where recall failures account for over 70% of its errors.