When should reasoning models translate? Learning to ask for help smartly

Reasoning language models handle English well but stumble on other languages—mostly because they don't understand non-English inputs reliably. Translating everything to English helps but adds unnecessary overhead. Luar uses reinforcement learning to train models to selectively invoke translation only when direct reasoning would fail, skipping translation when the original query is understandable. Across multilingual benchmarks, it outperforms standard approaches and generalizes to unseen low-resource languages.