Yuetian Lu, Yihong Liu, Hinrich Schütze
Relational linearity in language models is strongly correlated with hallucination rates, suggesting that how models store relational data affects their ability to self-assess knowledge accuracy.
This study explores why large language models, like those used in AI, sometimes give incorrect answers, a problem known as hallucination. The researchers focused on how these models respond to questions about made-up entities, finding that certain types of information are stored in a way that makes it harder for models to recognize when they're wrong. Specifically, information stored in a more abstract, linear manner tends to cause more hallucinations. This insight could help improve AI models by changing how they store and process information.
