ReAct Under the Microscope: Grounding Answers in External Evidence
ReAct's first serious test is not whether a language model can issue a search command. It is whether the model can use what the search returns to change its reasoning and select a better next move. This chapter examines that question in HotpotQA and FEVER, where the model must answer or verify a claim without receiving the supporting passages in advance. The goal is to read the results as evidence about a particular loop—reason, search, observe, revise—not as a universal ranking of ways to make language models answer questions.
The setup is valuable because it separates two sources of knowledge that are easy to blur together. A model may already contain a plausible fact in its parameters, or it may obtain a fact through the environment during the trajectory. ReAct tries to make the second route useful without abandoning the first. The comparison with reasoning-only and action-only prompting shows what each route contributes and where each one fails.
A question becomes a sequence of evidence needs
HotpotQA is a multi-hop question-answering benchmark: a correct answer may require connecting information from two or more Wikipedia passages. FEVER asks the model to classify a claim as SUPPORTS, REFUTES, or NOT ENOUGH INFO. In both experiments, the model receives only the question or claim. It does not begin with the support passages already placed in its context. It must either rely on its internal knowledge or interact with the paper's simple Wikipedia environment.
That environment has three relevant actions. search[entity] returns the first few sentences of a matching entity page or suggests similar entities when the exact page is not found. lookup[string] seeks the next sentence containing a phrase in the current page, much like a text search. finish[answer] ends the task. The interface is intentionally limited: it can retrieve only a small part of a passage through explicit queries and is weaker than a modern lexical or neural retriever. That weakness is useful for this experiment because it exposes the model's information-seeking decisions instead of hiding them inside a separate retrieval system.
A ReAct trajectory can therefore be read as a chain of evidence needs. The model first states what intermediate fact it needs. It searches for the relevant entity. The returned sentences may answer the question, supply only the next entity, or fail to contain the needed detail. The model then interprets that observation, chooses a lookup phrase or a more specific search, and repeats the process until it has enough support to finish. The important transition is between the observation and the next action: a result becomes useful only when the model can say what it established and what remains unknown.
Suppose a question asks for a property of an entity that is connected to another entity through an event or relationship. A reasoning-only model may jump directly from a familiar-sounding name to an answer. An action-only model may search the first name, receive a page, and then finish without checking whether the page contains the required relationship. ReAct adds an intermediate judgment: the observation either closes one part of the question or points to the next search. The external source supplies the evidence, but the language trace organizes the route through that evidence.
The same mechanism applies to FEVER, although the final output is a label rather than an open-ended answer. If the retrieved passage supports the claim, the model can finish with SUPPORTS. If it provides a contradiction, it can finish with REFUTES. If it leaves a critical detail unresolved, the correct action is NOT ENOUGH INFO rather than a confident guess. That last category is especially important: interaction with a source does not guarantee that the source settles the claim. ReAct must preserve uncertainty when the observation is insufficient.
The state transitions clarify why ReAct is not just retrieval followed by generation. Retrieval changes the model's available evidence, but the model still has to decide whether the evidence is relevant, complete, and consistent with the question. A poor result should move the trajectory to reformulation, not silently become a fact. A strong result should move it toward synthesis, but only after the model has connected it to the requested answer or label.
What the benchmark comparisons show
The reported PaLM-540B prompting results make the trade-off visible. On HotpotQA, Standard prompting reaches 28.7 exact match, chain-of-thought reaches 29.4, action-only prompting reaches 25.7, and ReAct reaches 27.4. On FEVER, the corresponding values are 57.1, 56.3, 58.9, and 60.9 accuracy. ReAct therefore improves on action-only prompting in both tasks, but it does not beat plain chain-of-thought on HotpotQA. On FEVER, where a small factual distinction can decide whether a claim is supported or refuted, access to retrieved information is more valuable and ReAct leads the single-method comparison.
Those metrics should be kept in their own frames. Exact match asks whether the predicted answer matches the benchmark answer; FEVER accuracy asks whether the classification label is correct. A four-point difference across those columns does not have one common meaning. More importantly, the same method can be helped or hurt by the task's information demands. A question that can be answered from a model's internal knowledge may not benefit from the extra steps of searching. A claim whose truth depends on a precise passage may benefit from the ability to obtain and inspect that passage.
The action-only comparison isolates the value of reasoning to guide acting. ReAct is better than Act on both HotpotQA and FEVER in the paper's results. An action-only model can have access to the same Wikipedia interface but fail to choose the next query, fail to synthesize several observations, or stop after a result that is merely related to the question. ReAct supplies language for decomposing the task and checking progress. The improvement is therefore not evidence that every tool call helps; it is evidence that tool access becomes more useful when the model can state what it is trying to find and why.
The comparison with chain-of-thought isolates a different trade-off. CoT is not restricted to the narrow retrieval interface, so it can formulate a clean reasoning structure from the full context. ReAct is more closely tied to observations, which can reduce unsupported factual steps but also constrain the form of reasoning and expose it to search errors. The paper's conclusion is not that one method should replace the other. It is that internal and external knowledge can be combined.
The combined methods use simple fallback rules. In one direction, ReAct backs off to chain-of-thought self-consistency—CoT-SC—when it does not return an answer within seven steps on HotpotQA or five on FEVER. Self-consistency samples multiple reasoning traces and uses the majority answer, so a strong majority is treated as a sign that internal reasoning may be sufficient. In the other direction, the system backs off from CoT-SC to ReAct when the majority answer appears fewer than half the time, treating weak agreement as a reason to seek external evidence.
The order matters because the two methods compensate for different weaknesses. ReAct can ground a shaky internal answer, while CoT-SC can provide a more flexible reasoning fallback when the retrieval route stalls. In the reported experiments, the two combined approaches each have an advantage on one task, and both outperform CoT-SC across the tested numbers of samples. They reach the performance of 21 CoT-SC samples with only 3–5 samples in the reported comparisons. These thresholds and sample counts are experimental settings, not a general controller that can be copied unchanged into another environment.
The table also contains an important negative result for inflated claims. Supervised task-specific systems score much higher than the prompting methods—67.5 exact match and 89.5 accuracy in the table's reference row. ReAct is a useful general prompting paradigm under the paper's conditions, but it is not presented as a substitute for every specialized retriever, verifier, or trained task system.
Grounding changes the error profile
Aggregate scores do not show how a method arrives at an answer. To examine that question, the authors manually inspected 200 sampled trajectories: 50 correct and incorrect examples from ReAct and the corresponding set from CoT. In the sampled success cases, they labeled ReAct as a true positive 94% of the time and a false positive 6% of the time; the corresponding CoT figures were 86% and 14%. A false positive here means that the result was correct even though the reasoning trace or facts contained a hallucination. The difference supports the paper's claim that retrieved observations make the ReAct trajectories more fact-driven in this sample.
The sampled failures reveal the cost of that grounding. ReAct's largest failure category is reasoning error at 47%, including failure to recover from repetitive steps. Search result error accounts for 23%: a search returns nothing or does not contain useful information. CoT's largest category is hallucination at 56%. ReAct's sampled failures are instead concentrated around the structure of the retrieval process. External evidence reduces one kind of error while making the system dependent on finding and interpreting useful evidence.
This distinction changes how a failure should be diagnosed. If a search returns an empty or unrelated result, the model may need a better query or a related entity. If the result is informative but the model repeats an earlier thought and action, the problem is not retrieval alone; the model has failed to update its working trajectory. If the retrieved passage supplies only part of the requested fact, the model should preserve the unresolved part instead of turning a partial observation into a complete answer. The error profile therefore separates the quality of the evidence from the quality of the model's update.
The manual categories are informative, but their scope matters. They come from a selected human-inspected sample, not from every trajectory in the benchmark, and they do not establish a causal estimate of how much retrieval changes hallucination in general. “Grounded” means that the answer is tied to observations from the environment in these experiments. It does not guarantee that the source is complete, current, or correct, and it does not prove that the visible thought is a faithful account of hidden computation.
The result is a conditional advantage
The knowledge-task evidence supports a precise conclusion. ReAct is more capable than action-only prompting when the model must use an external source and synthesize what it finds. It is not uniformly better than internal reasoning, because retrieval introduces cost, limited action space, and new failure paths. The strongest reported results come from allowing the model to move between modes: use internal reasoning when it is confident, seek external evidence when confidence is weak or facts need checking, and stop when the evidence is still insufficient.
That conclusion also explains why the next application matters. In a retrieval task, the loop may contain only a few searches and observations. In a long-horizon environment, the model must preserve subgoals across many actions and recognize state changes that are not reducible to a single fact. The same reasoning–observation relationship will be tested under a larger action space, where the question is no longer only “Which evidence supports this answer?” but also “What must happen next, and how will the agent know that it happened?”