Researcher Agents for Knowledge-Graph Question Answering

A recent paper on arXiv (2608.07700) presents an innovative text-to-SPARQL system that enhances traditional static tool-using agents. This new researcher agent can iteratively refine its prompts, rules, and tool orchestration code following each inference round on a validation set. The implementation utilizes DBpedia, resulting in nine evolving versions of the agent, driven by a cost-effective reasoning model. The top-performing setup employs two more robust backbone models. Notable findings indicate rapid self-improvement, achieving an overall accuracy of 0.22 on the 2025 DBpedia validation set, with persistent challenges in basic-graph-pattern generation. The paper also tackles issues in converting natural-language queries into SPARQL, including lexical ambiguity and ensuring syntactic and semantic accuracy in graph patterns.