Hi-Q: Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering

Hi-Q. Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering. By Jueun Kim and two others, from the Pohang University of Science and Technology. Published on the thirty-first of August. No code repository is linked to it.

Multi-hop question answering has a mismatch at its centre. The granularity at which a question is expressed often differs from the granularity at which evidence in the corpus can actually be retrieved. Existing methods deal with this in three broad ways. Some impose a fixed graph structure over the corpus. Some reformulate the query, again and again. Some generate a program and execute it over the corpus. What none of them does explicitly is decide, for a given unit of the query, whether it is already supported by evidence, or whether it should be refined further. That decision is left implicit, or settled ahead of time by the structure each method commits to.

Hi-Q makes that decision the problem itself. The authors formulate it as retrievable granularity discovery: finding the level at which each part of a question can be answered from the corpus. Hi-Q is an evidence-conditioned framework for hierarchical query refinement, and the important word is conditioned. Whether to refine a piece of the question is decided by the evidence that retrieval returns, not by a plan fixed in advance.

The working structure is a tree of queries. At each query node, a resolution operator tests whether the retrieved evidence supports the current query unit. If it does, the node is resolved and terminates. If it does not, the node is expanded by a binary operator that preserves the dependencies between the parts it creates, and that expansion is then checked by a semantic coverage verifier.

The consequence is that nobody designs the shape of the tree. It grows, and its topology is determined by corpus support signals rather than by a fixed decomposition template or a graph built beforehand. A query unit the corpus already supports stops where it is. One the corpus cannot support at its stated granularity keeps being refined until it can be.

Hi-Q is evaluated on three multi-hop QA benchmarks, primarily under full-corpus retrieval. In that setting, dependent evidence has to be located among open-domain distractors, rather than within a small annotated pool. Averaged over the three benchmarks, Hi-Q reaches fifty-two point three EM and sixty-four point zero F1. That puts it ahead of the iterative retrieval baseline IRCoT by fifteen point one EM and eighteen point two F1 on the same average. On MuSiQue-full specifically, it leads the graph-based RAG baseline PropRAG by eleven point five EM and twelve point zero F1, and it does so without constructing a graph over the whole corpus.

The authors also report the restricted supporting and distractor setting that prior work used. There too, Hi-Q attains the best accuracy, averaging fifty-seven point nine EM and sixty-nine point three F1. Its margin over PropRAG in that setting is five point six EM and three point nine F1. Over IRCoT it is thirteen point seven EM and fifteen point eight F1.

The thread through all of it is a change in who decides the decomposition. Templates and pre-built graphs settle the structure before any evidence has been seen. Hi-Q hands that choice to the corpus, one node at a time, according to whether retrieval can already support the question as it is currently phrased. The gains hold in both settings, and the strongest margins come in the open, full-corpus one, which is where a fixed structure has the least to go on.