Answering natural-language questions over knowledge graphs is hard because the underlying fact base is enormous and hard to navigate. A new preprint on arXiv proposes a way to ease that burden: instead of asking a large language model to reason over the whole graph, mine a focused subgraph related to the question and use it to rerank the model's candidate answers.
According to the abstract, the mined subgraphs are meant to serve as a potential solution to the scale problem. The paper does not provide full details in the abstract, but the direction is clear: use graph structure to guide or correct the output of an LLM rather than relying on the model alone.
Because this is a single preprint with a truncated abstract, there is no independent comparison to other approaches here. The claim is plausible but unverified until full results are available.