HyperReCo improves evidence retrieval for multi-hop question answering
HyperReCo: Retrieving and Connecting Evidence with Hypergraph Neural Networks for LLM Multi-hop Reasoning
Artificial Intelligence
Summary
Answering complicated questions often requires gathering clues from many documents. The authors introduced HyperReCo, a method that treats documents and their shared entities like a network of connections, helping find and link important facts more effectively. Their approach uses a special neural network to learn how clues relate depending on the question and then explains these relationships in a way large language models can use for better reasoning. Tests show HyperReCo finds helpful evidence more accurately and improves answering complex questions compared to other methods.
What this means in practice
- •For ai developers: Improve multi-hop question answering systems by retrieving and explicitly connecting complementary evidence through hypergraph-based methods.
- •For enterprise search teams: Enhance document retrieval tools to identify and combine related facts across documents for more accurate information answering.
Authors
Zicheng Zhao, Linhao Luo, Junnan Dong, Haoran Luo, Xiaoli Li, Shirui Pan, Chen Gong
Abstract
Large language models (LLMs) have shown strong capabilities, with retrieval-augmented generation (RAG) supporting complex multi-hop reasoning by retrieving evidence distributed across documents. Graph-based approaches exploit connections among evidence, and hypergraph-based retrieval further preserves higher-order entity associations within documents and connects documents through shared entities. However, existing hypergraph retrievers often rely on predefined structural expansion or diffusion, which may miss query-dependent interactions needed to identify relevant evidence. They also leave connections among retrieved evidence implicit, requiring LLMs to reconstruct these connections before reasoning. Therefore, we propose HyperReCo, a framework for retrieving and connecting evidence with a hypergraph neural network (HyperGNN). We represent each document as a hyperedge over its extracted entities, with shared entities connecting the hyperedges. Through hypergraph message passing with joint supervision over documents and entities, the HyperGNN learns query-dependent interactions to retrieve complementary evidence. We further introduce Gradient-Guided Hyper-Path Decoding (GGHD), which uses gradient attribution to interpret the learned interactions and translate them into explicit hyper-paths that help LLMs combine complementary facts for multi-hop reasoning. Experiments on six benchmarks show that HyperReCo achieves the best retrieval performance among the compared methods on all three multi-hop QA datasets, together with strong downstream QA performance. Case studies and further analyses demonstrate the utility of decoded hyper-paths for connecting retrieved evidence.