Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning

In three linesTranslate-R1 learns via RL a single policy deciding when to translate inputs into the model's dominant language. Trained on Qwen3-4B across 22 languages and 5 domains, the system improves reward by +4.6 to +23.5 depending on language resources, while reducing translation costs by 37% without performance loss.

Summary generated by Claude — human-verified