ReAct (Reason+Act)
A prompting framework that interleaves language model reasoning with grounded action execution.
ReAct is a prompting paradigm for large language models (LLMs) that interleaves chain-of-thought reasoning traces with discrete action steps, allowing a model to think through a problem and interact with external tools or environments in alternating fashion. Introduced in a 2022 paper by Yao et al. and gaining widespread adoption through 2023, the framework structures model outputs as sequences of Thought → Action → Observation tuples: the model reasons about what to do, issues an action (such as a search query or API call), receives an observation from the environment, and then reasons again in light of that new information. This loop continues until the model produces a final answer.
The core motivation behind ReAct is to address two failure modes common in LLM-based agents: hallucination and lack of grounding. Pure chain-of-thought prompting lets models reason but leaves them disconnected from real-world information, while pure action-based approaches (like tool-use pipelines) lack the interpretable reasoning that helps models recover from errors. By tightly coupling reasoning and acting, ReAct enables models to dynamically adjust their plans when an action returns unexpected results, producing more reliable and auditable behavior on tasks like multi-hop question answering, fact verification, and interactive decision-making benchmarks such as HotpotQA, FEVER, and ALFWorld.
ReAct has become a foundational building block in the LLM agent ecosystem. Frameworks like LangChain and LlamaIndex implement ReAct-style agents as a default pattern, and it has influenced subsequent agent architectures including Reflexion, Toolformer, and AutoGPT. Its significance lies not just in improved benchmark performance but in establishing a legible, modular structure for agentic behavior: because the reasoning traces are explicit, developers can inspect why an agent took a particular action and intervene when the logic goes wrong. This transparency is increasingly valued as LLM agents are deployed in higher-stakes settings requiring human oversight.
