Overview

ReAct can make failures easier to diagnose without making agents reliable by default. In a manually inspected sample, ReAct had 6% false positives in sampled successes, and its largest sampled failure categories were reasoning errors at 47% and search-result errors at 23%. The paper also identifies repeated loops, context pressure, decoding and stopping limits, portability gaps, and unsafe action permissions. Visible thoughts support inspection and intervention, but they are not guaranteed faithful explanations, and the experiments use constrained websites or simulations.

Learning outcomes

ReAct shifts some sampled failure risk from unsupported facts toward search quality, reasoning, and state updates.
Repeated actions, context pressure, decoding, and missing stopping rules can turn feedback into unproductive loops.
Visible trajectories create intervention points, but a useful thought is not proof of faithful hidden reasoning.

Keywords

ReActFailuresLoopsSafetyRetrievalOversight

Where ReAct Breaks: Search Errors, Loops, and Control Limits

ReAct's failures are easiest to misunderstand when every wrong answer is called a hallucination. The loop changes what the model can observe, but it does not guarantee that the observation is useful, that the next thought interprets it correctly, or that the action is safe to repeat. This chapter diagnoses the paper's failure surfaces—retrieval, reasoning, state updates, context, decoding, and action permissions—so the reader can judge what ReAct's visible trajectory helps with and what it leaves unresolved.

The distinction matters because grounding is a control input, not a magic correction. A search result can anchor an answer to evidence, but an empty result can send the model into a loop. A thought can expose an assumption, but it can also preserve a mistaken plan. The paper's strongest limitations are therefore part of its mechanism: once reasoning and acting share a loop, errors can be traced to the points where information crosses between them.

Grounding changes the shape of failure

The paper's manual analysis compares 200 sampled trajectories from ReAct and chain-of-thought prompting. In the sampled success cases, human judges label ReAct as a true positive 94% of the time and a false positive 6% of the time, compared with 86% and 14% for CoT. A false positive means that the final result is correct even though the trajectory contains a hallucinated fact or reasoning trace. The lower sampled false-positive rate is consistent with the paper's interpretation that external observations make ReAct more factual and grounded.

The failure cases show why that advantage should not be described as reliability in the abstract. ReAct's largest sampled category is reasoning error at 47%, including failure to recover from repetitive thoughts and actions. Search-result error accounts for 23%: the search returns nothing or fails to provide useful information. CoT's largest category is hallucination at 56%. In other words, ReAct and CoT do not simply differ in how often they fail; they fail for different reasons. ReAct moves some risk from unsupported internal facts to the quality of search, the flexibility of reasoning, and the ability to update the trajectory.

A search error is a failure at the evidence boundary. The model may ask for an entity under an imprecise name and receive similar results, or it may find a page that does not contain the requested detail. The observation is real, but it does not resolve the question. Recovery requires the model to recognize the gap, reformulate the query, and avoid treating the returned text as sufficient. If it cannot do that, more actions may make the trajectory longer without making it better.

A reasoning error is a failure at the interpretation boundary. The environment may have returned an informative result, yet the model can repeat the previous thought, fail to mark a subgoal complete, or choose an action unrelated to the remaining goal. Repetition is especially revealing because the action may be valid in isolation. The problem is that the model has not used the latest observation to change its plan. An action-only system can get stuck in the same way, but ReAct's explicit thoughts make the mistaken state update easier to see.

Published by DeepMox

FAQ

What are the main ReAct failure categories?

The paper’s manually inspected ReAct failures were led by reasoning errors at 47%, including repetitive steps, and search-result errors at 23%; these percentages come from a selected sample, not all trajectories.

Can a ReAct trace be treated as a faithful explanation?

No. A visible thought can be an editable control surface and help locate a mismatch, but it may be incomplete, strategic, or wrong.

What safety controls does ReAct still need?

It still needs bounded action permissions, repetition detection or step limits, stopping rules, and human review when consequences are high; the paper’s tests avoid dangerous operations.

Previous8 minReAct in ALFWorld and WebShop

Next9 minWhen ReAct Adds Decision Value

ReAct Failure Modes and Control Limits

Where ReAct Breaks: Search Errors, Loops, and Control Limits

Grounding changes the shape of failure

FAQ