Prompt-engineered LLMs rank first in Classical Latin Named Entity Recognition shared task

A University of Ottawa team reports placing first in both subtasks of the EvaLatin 2026 Named Entity Recognition (NER) shared task by using prompt engineering with commercial large language models (LLMs).

The shared task evaluated systems on two subtasks: coarse-grained NER with 11 classes and fine-grained NER with 28 classes, each under strict and fuzzy evaluation regimes. The team’s system achieved the best scores across all evaluation metrics and regimes among all submissions.

The authors attribute their performance to cross-lingual transfer learning, leveraging advancements from the broader LLM development community to address the underrepresented Classical Latin language.

The system description paper notes the use of prompt engineering with gemini-2.5-pro and claude-sonnet-4-5, commercial LLMs, to achieve these results without task-specific fine-tuning.

The work was submitted to arXiv on May 26, 2026, and describes participation in the EvaLatin 2026 shared task organized by LT4HALA@LREC, with results presented at the May 2026 conference in Palma de Mallorca, Spain.