Hi-Q: Hierarchical Evidence-guided Query Refinement for Multi-Hop Question Answering
Hi-Q decides which parts of a multi-hop question need refinement by checking whether retrieved evidence already supports them.
The paper frames the problem as “retrievable granularity discovery,” where a query may be too broad or too fine for the evidence available in a corpus. Hi-Q builds a query tree dynamically: supported nodes stop, unsupported nodes are split and verified for semantic coverage. In full-corpus retrieval tests across three benchmarks, it reports 52.3 EM and 64.0 F1 on average, beating IRCoT and PropRAG on the cited comparisons. HF Daily Papers' note
The paper frames the problem as “retrievable granularity discovery,” where a query may be too broad or too fine for the evidence available in a corpus. Hi-Q builds a query tree dynamically: supported nodes stop, unsupported nodes are split and verified for semantic coverage. In full-corpus retrieval tests across three benchmarks, it reports 52.3 EM and 64.0 F1 on average, beating IRCoT and PropRAG on the cited comparisons. HF Daily Papers' note
score 4