From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

Large language models (LLMs) answer questions using pretrained knowledge, but the internal process connecting knowledge retrieval to answer generation remains only partly understood. We find that LLMs use routing information to retrieve relevant pretrained knowledge, then draw on knowledge represented in hidden states to produce an answer. Specifically, parameter routing guides access to pretrained knowledge, while answer-supporting hidden-state knowledge contributes to the answer. Deletion and replacement interventions provide causal evidence for parameter routing and the hidden-state knowledge used to answer. Tracing their effects across layers reveals an overlapping process of knowledge retrieval and use with model- and task-specific timing: parameter routing shapes the knowledge formed in subsequent layers, while hidden-state knowledge supports the answer. In later layers, hidden-state knowledge remains consequential even as parameter-routing deletion causes less answer damage.