RELOOP: A Structure-Aware RAG Framework for Multi-Step Heterogeneous QA
RELOOP is a new retrieval-augmented generation framework designed to handle multi-step questions and heterogeneous evidence sources like text, tables, and knowledge graphs. It introduces Hierarchical Sequence (HSEQ) to linearize diverse documents into a reversible hierarchical structure with lightweight structural tags. The framework uses a Head Agent for guidance and an Iteration Agent for structure-aware retrieval, performing parent/child hops, table row/column neighbors, and KG relations. Evidence is canonicalized before answer synthesis, with an optional refinement loop to resolve contradictions. Experiments on HotpotQA, HybridQA, TAT-QA, and MetaQA show improved accuracy and efficiency over baseline RAG methods.
Key facts
- RELOOP uses Hierarchical Sequence (HSEQ) to linearize documents, tables, and knowledge graphs.
- The framework employs a Head Agent and an Iteration Agent for retrieval.
- It performs structure-aware actions like parent/child hops and table row/column neighbors.
- An optional refinement loop resolves detected contradictions.
- Tested on HotpotQA, HybridQA, TAT-QA, and MetaQA datasets.
- RELOOP aims to balance accuracy, latency, and token/tool budgets.
- The paper is available on arXiv with ID 2510.20505.
- The approach is designed for multi-step heterogeneous QA tasks.
Entities
Institutions
- arXiv