GRPO Boosts LLMs' Cross-Lingual Fact Recall Across 12 Languages

Large language models trained predominantly on English data often fail to recall facts consistently in other languages—a problem known as cross-lingual factual inconsistency. To systematically address this, researchers introduce PolyFact, a large-scale parallel multilingual QA dataset containing 100,000 Wikidata-grounded facts across 12 typologically diverse languages (including low-resource ones). They compare three approaches: light continual pretraining (CPT) on parallel data, supervised fine-tuning (SFT), and reinforcement learning via Group Relative Policy Optimization (GRPO), applied to Qwen-2.5-7B and OLMo-2-1124-7B.

Surprisingly, GRPO consistently outperforms both SFT and CPT, not only improving recall on seen languages but also generalizing to unseen languages. Mechanistic analyses show that GRPO reduces language-specific specialization in MLP layers and attention heads, effectively reorganizing multilingual routing toward shared cross-lingual representations. The findings suggest that reinforcement learning—particularly consistency-driven RL—can bridge gaps in multilingual factual recall without requiring massive parallel data. The dataset and code are open-sourced under arXiv:2606.06586, under review at EMNLP 2026.