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.
References
The inner workings of LLMs are themselves a controversial topic, and some unconfirmed issues about these mechanisms are fatal to the task of knowledge updating: one such question is whether LLMs genuinely perform latent reasoning.
The results indicate the two-layer setting improves success and that the full system can complete repository-level repair without pre-built entity relation edges. This is compatible with the hypothesis, but does not directly verify the latent graph realized inside the model.
In a ``closed-book'' setting, we do not know if an LLM knows a fact because it simply memorized the fact from the pre-training corpus, or perhaps the model made a multi-hop inference based on other memorized facts.