Large language models (LLMs) can memorize and apply new information. However, it's unclear how they balance this new context with their pre-existing knowledge. This research analyzes how LLMs manage this conflict using a new counterfactual dataset.

The study investigates LLMs using Fakepedia, a dataset presenting contradictions between known facts and new information. Through Masked Grouped Causal Tracing (MGCT), the research deciphers LLMs' grounding mechanisms by contrasting neural activation patterns. Findings help understand the co-functioning of grounding with recall capabilities within LLMs.