Teaching language models new languages without forgetting old ones

Extending LLMs to low-resource languages usually causes catastrophic forgetting of general skills—the "alignment tax." This work replaces supervised fine-tuning with Group Relative Policy Optimization (GRPO), using embedding-level semantic rewards rather than likelihood maximization. The approach encourages meaning preservation through flexible surface realizations, reducing interference with pretrained knowledge. Evaluated on Tibetan-Chinese translation and generation, the method acquires new language capabilities while better preserving general competence than standard fine-tuning, with improved semantic quality and few-shot transfer performance.