Whether LLMs Genuinely Perform Latent Reasoning

Determine whether large language models internally perform latent multi-hop reasoning—retrieving intermediate bridge entities and propagating inference—rather than simply memorizing multi-hop question patterns as atomic facts.

Background

Understanding if LLMs execute latent multi-hop reasoning is crucial for knowledge updating: if models do not propagate edits through related facts, parameter-level updates may fail to generalize. The authors cite mixed and context-dependent evidence on latent reasoning, underscoring the need for a definitive answer.