LLMs Use Lifted Representations for Memory Updates

Understanding how Large Language Models (LLMs) store, select, and revise information is crucial for improving their reliability. When LLMs answer queries, they often map specific observations to more general, rule-like structures. However, the underlying mechanism for this generalization remains unclear. Researchers introduce the "lifted representation hypothesis," which posits that LLMs update their memory by leveraging shared latent structures rather as than individual, isolated facts. This perspective frames "lifting" as an efficient way to exploit symmetries across different instances, while "shattering" refers to the process of refining these broad, lifted structures into more specific subtypes when exceptions arise. Through controlled experiments on exception-learning across various training methods (in-context learning, LoRA, fine-tuning), the study found that LLMs are prone to "shattering failures" when dealing with data governed by nested rules and exceptions. Furthermore, lifting often occurs prematurely, indicating challenges in how LLMs balance generalization and specificity.