Answers are grounded in this paper and cite their sources.
Graph-based Retrieval-Augmented Generation (GraphRAG) has emerged as a promising framework for multi-hop question answering by integrating graph-structured retrieval with large language model (LLM) generation. Nevertheless, its robustness to query perturbations remains underexplored, even though real-world information needs are frequently expressed through noisy, reformulated, or stylistically diverse queries. This paper is driven by two central objectives: (1) to quantify the impact of different types of query perturbations on GraphRAG performance, and (2) to mitigate perturbation-induced failures specific to the GraphRAG pipeline by addressing them at early stages, before they propagate to downstream components such as graph expansion. For the first objective, we perform a comprehensive robustness analysis on HotpotQA and 2WikiMultiHopQA using character -, word -, and sentence -level perturbations. The results demonstrate that GraphRAG is consistently vulnerable to such variations across a range of experimental settings. For the second objective, we propose R obust E ntity R eranking for G raph RAG ( RER-GRAG ), together with an outcome-supervised trained variant, Tr-RER-GRAG , as a targeted mechanism for improving the robustness of GraphRAG to query perturbations. Experimental results demonstrate that the proposed approach consistently outperforms the original GraphRAG pipeline as well as representative robustness baselines. In particular, relative to GraphRAG under perturbed queries, Tr-RER-GRAG consistently improves average accuracy across character-, word-, and sentence-level perturbations, with gains of 13.19%, 13.64%, and 17.12% on HotpotQA, and 21.78%, 18.95%, and 46.16% on 2WikiMultiHopQA, respectively (averaged over per-perturbation-type relative gains). Further analyses distinguish improvements in perturbed-query accuracy from reductions in perturbation sensitivity. On 2WikiMultiHopQA, Tr-RER-GRAG reduces both clean-normalized performance degradation and perturbation-induced failure rates. On HotpotQA, it reduces perturbation-induced failures and better preserves predictions that are correct for clean queries, although aggregate clean-normalized degradation does not decrease. Overall, these findings demonstrate that entity reranking improves GraphRAG accuracy under query perturbations across both benchmarks, while the extent to which these improvements translate into reduced sensitivity to query perturbations varies across datasets. The code is publicly available at https://github.com/halizadehn/RER-GRAG .
No takes yet. Share an insight, caveat, or question.
Noughabi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: