No evidence of Semitic-specific cross-lingual transfer in large language models

Researchers fine-tuned seven large language models ranging from 4B to 671B parameters on Arabic and evaluated zero-shot reading comprehension across Semitic and non-Semitic languages. The study included both dense and Mixture-of-Experts architectures.

Across architectures, the authors report no evidence of Semitic-specific transfer. Models with weak baselines improved dramatically across all languages after fine-tuning, while models with strong baselines showed only marginal gains regardless of language family.

A chain-of-thought ablation reinforced these findings: the same models that benefited most from fine-tuning also benefited equally from inference-time reasoning. The authors interpret this as evidence that both mechanisms address task-format alignment rather than cross-lingual knowledge transfer.