The History of AI

ReAct

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Agents

Lineage
Acting in the world
Kind
Publication

Who

Princeton University, Google Brain

What happened

Shunyu Yao and colleagues at Princeton and Google propose ReAct: a language model that alternates reasoning steps with actions, such as looking something up on Wikipedia or moving through an environment, and uses what it observes to decide the next step.

Why it mattered

Reasoning and actions interleaved in a single loop, a pattern that many agents built on language models will adopt.

The full story

In 2022 there were two separate lines of work. Chain of thought got models to reason in writing, but only with what they already knew, which is why they sometimes made up facts and carried the error forward. Other work used models to choose actions in an environment, but without reasoning about them.

ReAct brought the two together. The model writes a thought, such as what it needs to find out, carries out an action, such as searching for a term on Wikipedia, reads the result, and thinks again with that information. The reasoning serves to plan, track progress, and revise the plan, and the actions bring in new information from outside.

On questions that require combining facts, such as those in HotpotQA, and on fact verification, querying a simple Wikipedia interface reduced chain of thought's fabrications and error propagation, and the trajectories were easier to interpret. In two interactive environments, ALFWorld and WebShop, it beat imitation and reinforcement learning methods by 34 and 10 points of success rate, with only one or two examples in the prompt.

The paper was presented at ICLR 2023. Its loop of thinking, acting, and observing is the one that many later agents follow, with variations.

Three diagrams. A line of four nodes, the chain of thought. A branching tree, the tree of thoughts. A cycle of three boxes, think, act, and observe, that loops back to think: ReAct.

ReAct combines chain of thought's written reasoning with actions that bring in information from outside.

Videos

Builds on

  • Chain of thought (January 28, 2022)

    Documented direct influence

    ReAct starts from chain-of-thought reasoning, which it cites as an example of reasoning cut off from the outside world, and interleaves it with actions (abstract)

Leads to

  • Tree of thoughts (May 17, 2023)

    Inherited, with no proven direct link

    Both papers take the model off a single line of reasoning, one with actions and the other with search

  • Toolformer (February 9, 2023)

    Independent convergence

    Both papers give a language model external tools, from different directions: one with examples in the prompt, the other with self-supervised data

  • Computer use in Claude (October 22, 2024)

    Inherited, with no proven direct link

    Computer use follows the ReAct loop: the model looks at a screenshot, decides on an action, and looks again at the result

Threads that pass through here

Sources

Original source

How to cite

“ReAct.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/react

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline