Expanding the Capabilities of Reinforcement Learning via Text Feedback https://arxiv.org/abs/2602.02482