Search papers, labs, and topics across Lattice.
The paper introduces C2RAG, a framework to improve multi-hop reasoning in Graph Retrieval-Augmented Generation (GraphRAG) by addressing knowledge graph (KG) quality issues like noise and incompleteness. C2RAG uses constraint-based retrieval with fine-grained constraint anchoring to filter candidates and suppress retrieval drift, along with a sufficiency check mechanism to prevent retrieval hallucinations by activating textual recovery when needed. Experiments on multi-hop benchmarks show C2RAG outperforms existing baselines by 3.4% EM and 3.9% F1, demonstrating improved robustness under KG imperfections.
GraphRAG gets a robustness boost: C2RAG tackles noisy and incomplete knowledge graphs, outperforming baselines by up to 3.9% F1 on multi-hop reasoning tasks.
Graph Retrieval-Augmented Generation enhances multi-hop reasoning but relies on imperfect knowledge graphs that frequently suffer from inherent quality issues. Current approaches often overlook these issues, consequently struggling with retrieval drift driven by spurious noise and retrieval hallucinations stemming from incomplete information. To address these challenges, we propose C2RAG (Constraint-Checked Retrieval-Augmented Generation), a framework aimed at robust multi-hop retrieval over the imperfect KG. First, C2RAG performs constraint-based retrieval by decomposing each query into atomic constraint triples, with using fine-grained constraint anchoring to filter candidates for suppressing retrieval drift. Second, C2RAG introduces a sufficiency check to explicitly prevent retrieval hallucinations by deciding whether the current evidence is sufficient to justify structural propagation, and activating textual recovery otherwise. Extensive experiments on multi-hop benchmarks demonstrate that C2RAG consistently outperforms the latest baselines by 3.4\% EM and 3.9\% F1 on average, while exhibiting improved robustness under KG issues.
