Post-Training Degrades Grammatical Competence in Low-Resource Languages

Evaluation of six model families on the MultiBLiMP benchmark reveals that post-training reduces grammatical competence, specifically hitting low-resource languages hardest. While post-trained models retain latent knowledge, it is often unrecoverable through prompting unless using the model's native language.