MARS: Multi-hop Adaptive Retrieval and SPARQL Generation for KGQA

As large language models (LLMs) are prone to hallucinations and struggle with complex multi-hop reasoning over structured data, the authors sought to develop a more reliable and scalable method for Knowledge Graph Question Answering (KGQA). The goal was to create a system that leverages the reasoning capabilities of LLMs while grounding them in explicit, up-to-date symbolic knowledge from knowledge graphs (KGs) without the need for costly fine-tuning or unpredictable agentic exploration.

The authors propose MARS, a pipeline that replaces open-ended agentic exploration with a structured, iterative retrieval process. The system begins by identifying entities in a natural language question and then enters a reasoning loop. In each iteration, MARS retrieves and ranks relevant triple patterns—rather than raw subgraphs—and enriches them with typological information and instance counts. An LLM then evaluates this context to either generate a final, executable SPARQL query or decide to perform further graph traversal. This approach allows the model to adapt its retrieval depth to the complexity of the specific question while maintaining a predictable execution flow.

MARS achieves state-of-the-art performance on established KGQA benchmarks, including QALD-10, where it outperforms existing agentic alternatives. The authors demonstrate that by focusing on pattern-based retrieval, the system effectively manages context size and reduces the likelihood of error propagation common in agent-based frameworks. Furthermore, the study highlights the importance of reproducibility in KGQA, providing a transparent, open-weight model setup and updated datasets to address inconsistencies in previous research.

This work provides a robust, efficient framework for bridging the gap between natural language queries and structured knowledge bases. By avoiding fine-tuning and relying on a deterministic, pattern-guided retrieval loop, MARS offers a more reliable and interpretable solution for knowledge-intensive tasks, making it highly applicable for real-world scenarios where data is frequently updated and accuracy is paramount.