#spurious rewards

Qwen2.5-Math models improve math reasoning significantly even when trained with incorrect or random reward signals, highlighting unique reinforcement learning dynamics not seen in other models.