LaboroAI at LLMs4OL 2026 Tasks Flagship and Reuse: Less is More for LLM-based Ontology Learning
Abstract
This paper presents team LaboroAI's participation in the LLMs4OL 2026 Challenge, covering the Flagship task (end-to-end ontology learning from text) and the Reuse task (extending a given seed ontology). Both systems share one deliberately simple design: a single LLM call per document, grounded by few-shot exemplars retrieved from the training pool, followed by deterministic symbolic post-processing. We evaluated a range of more complex alternatives, including decomposing the task into specialist passes, LLM self-critique, and ensembling over model and prompt variants, and none outperformed the simple system on held-out data. Our best submissions achieved test exact graph similarity of 0.6440 on the Flagship task and 0.6989 on the Reuse task. We additionally perform a systematic ablation study across both tasks and analyze the remaining error modes.