Researcher Agents Self-Improve Knowledge Graph Question Answering

Translating natural language questions into precise SPARQL queries for large knowledge graphs is a complex task, involving resolving ambiguities, mapping terms to ontologies, and generating valid, semantically accurate graph patterns. Traditional tool-using agents often have static configurations, limiting their adaptability. Researchers have developed a novel "researcher agent" for text-to-SPARQL systems that introduces a self-improvement loop. After each round of inference on a validation set, this agent autonomously proposes and tests modifications to its own prompts, rules, and the code that orchestrates its tools. This iterative self-improvement process was instantiated on DBpedia, where the agent evolved through nine versions. The study revealed that this self-improvement converges quickly, achieving a 0.22 overall accuracy on the DBpedia validation set. The primary bottleneck identified was predicate selection in basic graph patterns, rather than syntax. This approach demonstrates a significant step towards more adaptive and accurate knowledge graph question answering systems.