Multi-hop question answering (QA) often faces a granularity mismatch: questions are expressed at a certain level of detail, while corpus evidence is retrievable at a different level. Traditional methods impose fixed graph structures, iteratively reformulate queries, or execute generated programs, but they rarely decide when a query unit is already supported by evidence and when it should be refined.
To address this, researchers at Hugging Face propose Hi-Q, an evidence-conditioned framework for hierarchical query refinement. At each query node, a resolution operator tests whether retrieved evidence supports the current query unit. Resolved nodes terminate, while unresolved nodes are expanded by a dependency-preserving binary operator and checked by a semantic coverage verifier. This process grows a query tree whose topology is determined by corpus support signals rather than a fixed decomposition template or a pre-built graph.
Evaluated on three multi-hop QA benchmarks under full-corpus retrieval—where dependent evidence must be located among open-domain distractors—Hi-Q achieves 52.3 EM and 64.0 F1 on average. This surpasses the iterative retrieval baseline IRCoT by 15.1 EM / 18.2 F1 and the graph-based RAG baseline PropRAG by 11.5 EM / 12.0 F1 on MuSiQue-full, without requiring corpus-wide graph construction.
In the restricted supporting/distractor setting used by prior work, Hi-Q also attains the best accuracy, with 57.9 EM and 69.3 F1 on average, ahead of PropRAG by 5.6 EM / 3.9 F1 and IRCoT by 13.7 EM / 15.8 F1. The project page is available at hi-q-project.github.io.