A paper introduces ArcRAG, a retrieval method for graph-based retrieval-augmented generation that adapts to the knowledge graph at question time rather than…
The paper addresses a weakness the authors identify in GraphRAG systems: retrieval strategies are typically designed around the structure their own construction pipeline produces, and whether they survive a change of construction is rarely tested, according to the paper. The authors describe their study as one of cross-construction portability, asking whether one retrieval procedure, with no construction-specific schema, relation names, or retuning, stays effective when the graph is built differently.
The mechanism the paper describes works in three stages. ArcRAG induces a query-local graph interface at inference time, consisting of the node labels, relation descriptors, and property keys reachable from the entities a question names. It then plans multi-hop traversals against that interface and executes and repairs them as Cypher queries. The evaluation builds each of the MuSiQue, HotpotQA, and TriviaQA corpora twice: once as a schema-guided graph with typed nodes, a closed relation set, and source provenance, and once as an open-information-extraction graph with untyped nodes and free-text relations. The retriever's configuration is fixed across both versions. The paper reports no specific accuracy figures in the material available here, so the size of any performance gap between the two constructions cannot be quantified from this source.
The comparisons in the paper are of two kinds. The authors compare ArcRAG against what they call a construction-agnostic Direct-Cypher agent, and they ablate each stage across retry and pass budgets on both graph families. They also compare with results reported for prior GraphRAG systems, most of which, the paper says, design retrieval together with their own graph construction. That co-design is the practice the paper is arguing against: if retrieval must be retuned whenever the graph construction changes, then a GraphRAG system is locked to one builder. The headline finding, according to the paper, is that retrieval need not be co-designed with the graph, because query-time interface induction keeps a construction-agnostic retriever effective when the construction changes. The paper qualifies this immediately: construction still shapes absolute accuracy, meaning the choice of graph build affects how well the system performs even if the retriever itself does not need rewriting.
The evidence here comes from a single source, the paper's own description, and no independent second source was found for this event. The material available does not include the reported metrics, the ablation results, the retry and pass budgets used, or any author information, so none of those can be stated. The comparisons with prior GraphRAG systems are against results reported by those systems, not re-run experiments, which is a distinction the paper itself makes but which a reader should keep in mind. No peer-review status is stated in the material available.
For practitioners, the practical question the paper raises is whether GraphRAG deployments can swap graph construction pipelines without rewriting retrieval logic. The paper's answer, on its own evidence, is that query-time interface induction makes this possible in principle, while noting that construction still shapes absolute accuracy. Whether the effect holds at the scale and cost of production systems is not established by the material available.
Source: ArcRAG: Agentic Compositional Retrieval over Knowledge Graphs