We current the Multilingual Reasoning Gymnasium, an extension of Reasoning Gymnasium (Stojanovski et al., 2025), that procedurally generates verifiable reasoning issues throughout 14 languages. We translate templates for 94 duties with native-speaker validation in 10 languages and focused code or template diversifications to make sure linguistic naturalness. The Multilingual Reasoning Gymnasium preserves the core advantages of the procedural era method used within the authentic Reasoning Gymnasium, reminiscent of just about limitless drawback occasion era and adjustable problem, and stays straight usable for Reinforcement Studying from Verifiable Rewards and analysis settings. Issues within the Multilingual Reasoning Gymnasium are parallel throughout languages, enabling crosslingually parallel information era at large scale as a result of procedural nature of the environments. We launch our implementation to assist analysis into multilingual reasoning fashions.
- † Hasso Plattner Institute & ELLIS Unit Potsdam
- ** Work achieved whereas at Apple
- ‡ Equal contribution
