ReAct: Synergizing Reasoning and Acting in Language Models
Reasoning (e.g. chain-of-thought) and acting (action-plan generation) in LLMs had mostly been studied separately, limiting their synergy.
Interleaved generation of reasoning traces and task-specific actions so the model can update plans and interact with external sources.
ReAct outperformed prior methods on interactive decision-making tasks and improved human interpretability and trustworthiness.
Depends on prompt quality and in-context examples; scaling to very complex environments remains challenging.
The core pattern behind many modern tool-using agents. Directly relevant to the software-agent project that must reason and then call tools or applications.
📇 Summary flashcard — 13 analytical fields for this paper
خلاصه
Combines reasoning and acting in LLMs for better interactive decision making.
نمای سریع
ReAct – interleaved reasoning and acting.
یافتههای کلیدی
ReAct outperformed prior methods on interactive decision-making tasks and improved human interpretability and trustworthiness.
هدف
Synergise reasoning and acting in one loop.
روش
Interleaved generation of reasoning traces and task-specific actions so the model can update plans and interact with external sources.
نتایج
Higher success on ALFWorld and WebShop and better interpretability.
نتیجهگیری
Reasoning and acting should be generated together.
مفاهیم کلیدی
ReAct، function calling، agent، reasoning
مطالعهی بیشتر
https://arxiv.org/abs/2210.03629
تحلیل
Became the standard pattern for many subsequent tool-using agents.
محدودیتها
Depends on prompt quality and in-context examples; scaling to very complex environments remains challenging.
کارهای آینده
Scaling to more complex environments and reducing prompt dependence.
کاربرد عملی
The core pattern behind many modern tool-using agents. Directly relevant to the software-agent project that must reason and then call tools or applications.
◀ References (this paper draws on) (0)
No references recorded.
Cited by (0) ▶
No paper cites this one yet.
Prerequisite learning path for this paper