LLMs4OL 2026 Overview: The 3rd Large Language Models for Ontology Learning Challenge
Abstract
We present the results of the 3rd Large Language Models for Ontology Learning Challenge, held at the International Semantic Web Conference 2026, a shared task designed to evaluate the effectiveness of large language models (LLMs) for ontology learning. The challenge comprised three complementary tasks: end-to-end ontology construction from text (Flagship), extension of an existing partial ontology (Reuse), and cross-domain taxonomy discovery (Taxonomy). Participants explored a broad range of approaches, including general-purpose and adapted LLMs, embedding-based retrieval, conventional NLP pipelines, and hybrid symbolic methods. Results show that constraining the label space through a supplied vocabulary matters more than model scale, and that retrieval grounding and task-specific adaptation of open-weight models transferred more reliably than post-hoc correction. Persistent challenges include candidate retrieval, entity normalization, and complete taxonomy recovery, while non-taxonomic relation extraction was rarely attempted under the current scoring. Building on this edition, we are integrating submitted systems into OntoLearner, our modular framework for ontology learning, so challenge contributions remain reusable and comparable.