GRASP and GRISP for Question Answering over Wikidata

We describe our submissions to the WikiKGQA 2026 challenge on generating SPARQL queries for English and Spanish natural language questions over Wikidata. We use two methods: GRASP, an LLM agent that iteratively explores the knowledge graph and refines a query, and GRISP, a lightweight fine-tuned model that predicts a query skeleton and fills it with IRIs. We only entered the two without-mentions tracks, and we also report standalone Claude Code agent baselines to show what an off-the-shelf agentic LLM achieves without dedicated KGQA tooling. GRASP gets the best result of all submissions in both without-mentions tracks, matching the strongest Claude Code baseline at a fraction of its cost. GRISP is a much cheaper, fully local alternative, but its results are noticeably weaker.