Request Correction

SyntaxGym for French: Resource, Annotation, and Evaluation of French and Multilingual LLMs

Despite recent advances in large language models (LLMs), their syntactic competence remains insufficiently characterized, especially for languages other than English. While benchmarks such as BLiMP and SyntaxGym have enabled systematic syntactic evaluation in English and Spanish, no comparable resource exists for French. To address this gap, we present SyntaxGymFR, a manually curated evaluation suite for evaluating the syntactic abilities of French and multilingual LLMs. SyntaxGymFR consists of manually validated minimal sentence pairs targeting key syntactic phenomena in French. We describe the annotation methodology, the selection of linguistic constructions, and the validation procedures used to ensure the coverage of syntactic phenomena. Furthermore, we report experimental results obtained with several French and multilingual LLMs, analyzing their sensitivity to grammatical contrasts and cross-linguistic transfer effects. Our results provide new insights into the syntactic generalization capabilities of French LLMs and establish SyntaxGymFR as a benchmark for future research on language-specific evaluation of syntactic competence.