This research introduces ReAct, a novel prompting method that enhances language models by synergizing reasoning and acting. ReAct prompts language models to generate interleaved reasoning traces and actions, allowing dynamic reasoning and interaction with external environments. Experiments across diverse tasks like question answering, fact verification, text-based games, and web navigation demonstrate ReAct's superiority over isolated reasoning or action approaches. The approach not only improves task performance but also enhances model interpretability and trustworthiness. Further analysis shows the importance of both reasoning to guide actions and acting to inform reasoning. Moreover, initial experiments involving the application of ReAct in closed loop systems for tasks like robotic action planning reveals that ReAct produces more robust results. The work shows the potential for human intervention and correction, making this method a promising step towards better human-machine collaborations.

Source: https://arxiv.org/pdf/2210.03629