ReAct searched Wikipedia between thoughts and still trailed chain-of-thought on HotpotQA

A thought, an action, and a look at what came back

In 2022, Shunyu Yao and six coauthors at Princeton and Google published a simple recipe for getting more out of a large language model, a model trained to predict the next word across most of the web. Call the model in a loop. Let it write a thought, a sentence of plain reasoning about what to do next. Let it emit an action, a short command such as a search query. Feed it back the observation, the text the outside world returns. Then let it think again. The authors named the method ReAct, for reasoning and acting, and ran it on PaLM-540B, a 540-billion-parameter language model from Google.¹²

The model was not retrained for any of this. ReAct runs by few-shot prompting: you paste one or two worked examples into the prompt and the model imitates the pattern, its weights untouched. For the question-answering tests, the only tool it could reach was a Wikipedia interface with three commands, written as search[entity], lookup[string], and finish[answer].¹

Nothing in that run is special engine work. The model wrote the words Thought, Act, and Obs; a wrapper program read the Search[...] line, actually queried Wikipedia, and pasted the result back as the next Obs line. How a model's tool call runs is its own lesson. What ReAct added on top was the interleaving. A thought before each action, so the model plans, and an observation after it, so the next thought answers to something real rather than to the model's memory.¹

The search loop scored under chain-of-thought on HotpotQA

The obvious question is whether the loop helps. The paper's headline reasoning test was HotpotQA, a set of trivia questions that each need facts from more than one Wikipedia article. Its scoring is exact match: the model's answer string has to match the gold answer exactly, so the scores run low.¹

On HotpotQA, ReAct scored 27.4% exact match. Plain chain-of-thought, the same model prompted to reason step by step in words with no search and no observations, scored 29.4%. Standard prompting, with no reasoning shown at all, scored 28.7%. On the benchmark it is most associated with, the reasoning-and-search loop the paper is built around came in last of the three.¹³

FEVER, a fact-checking test where the model labels a claim supported or refuted, went the other way. ReAct scored 60.9% there against chain-of-thought's 56.3%. The paper states the split plainly, writing that ReAct "outperforms CoT on Fever (60.9 vs. 56.3) and slightly lags behind CoT on HotpotQA (27.4 vs. 29.4)." The paper's strongest HotpotQA mark, 35.1%, came from a hybrid that ran ReAct and let the model fall back to plain reasoning when search stalled. For scale, a model trained directly on HotpotQA scored 67.5%. Few-shot ReAct was nowhere near it.¹

Grounding every step drove invented facts to zero

So why did the paper matter, if its own loop lost at trivia? Because of what the loop did to hallucination, a model stating a fact that is simply false. The authors hand-labeled 50 correct and 50 incorrect runs each from ReAct and from chain-of-thought on HotpotQA, 200 in total, and sorted where the failures came from.¹

Chain-of-thought's single largest failure mode was hallucination: 56% of its wrong answers came from the model inventing a fact mid-reasoning. For ReAct that figure was 0%. Among the answers each method got right, chain-of-thought reached 14% of them through reasoning that held a made-up fact but landed on the right answer anyway. For ReAct that share was 6%. Reading a real Wikipedia line before the next thought kept the model honest.¹

Grounding cost something, and the paper says so. ReAct's own largest failure mode was reasoning error at 47%, well above chain-of-thought's 16%, often the model repeating the same thought and action without breaking out of the loop. Another 23% of its failures were search returning nothing useful, which left the model stuck. The loop trades some of the model's freedom to reason for a tether to the record.¹

The headline gains came on chores and a shopping site

The numbers the paper is quoted for are the two interactive tests, not the question-answering scores. On these the model does not answer a question. It takes a sequence of actions toward a goal.¹

ALFWorld is a text game of household tasks: the model reads a room description and issues commands to find and move objects, like putting a pepper shaker on a drawer. ReAct's best run there succeeded on 71% of tasks. The same loop with the thoughts stripped out, actions and observations only, managed 45%, and the previous published agent, BUTLER, managed 37%. Even ReAct's worst run beat both. WebShop is a simulated shopping site where the model has to buy an item matching a written request. ReAct succeeded 40% of the time against 30% for the best prior method, though expert humans hit about 60%. Those two gaps, 34 points on ALFWorld and 10 on WebShop, are the "34% and 10%" the abstract advertises, and they are what "ReAct started agents" points at.¹

Today's agents call tools in JSON, not hand-written prose

The blueprint stuck. The exact method did not. ReAct's loop lives on in every AI agent that thinks, calls a tool, reads the result, and thinks again. But the paper made the model produce that loop by hand, writing the literal words Thought, Act, and Obs in a text format copied from the one or two examples, and a wrapper parsed those strings.¹

Today's models are trained to emit tool calls as structured data. When a current model calls a function, it returns a typed tool_use block naming the tool and its arguments. The application runs the tool and sends the answer back in a tool_result block, and the model continues.⁴ OpenAI documents the same round trip as a five-step loop.⁵ The thought, action, and observation pattern now lives inside the model and the API instead of being prompted in as text.

The name drifted with the method. "ReAct" today is used for almost any agent that mixes reasoning with tool calls, a general pattern rather than the paper's specific prompted format.⁶ What carried forward is the plainest part of what the paper tested: reason before you act, and let what you observe shape the next thought.¹