Loop Correction

An AI language model normally produces an answer in a single pass: input goes in, output comes out. A loop correction changes this process. The model first generates an initial answer, then evaluates it itself for errors or gaps, and subsequently improves it. This cycle — generate, check, revise — can be repeated multiple times. The result is an answer that the model has worked through across several rounds, rather than spitting out in one go.

Why a single pass is often not enough

Language models are trained to predict token by token — that is, word by word. They don’t plan the way a human does when drafting, discarding, and revising an essay. They follow the most probable next word. This means that errors occurring early in the text get carried through to the end, because the model can no longer correct its own beginning.

For simple tasks, this is barely noticeable. For multi-step problems — long calculations, logical reasoning, or complex code — small errors quickly add up. Loop corrections are an attempt to work around this structural problem without retraining the model from scratch.

How a correction cycle works

In the simplest case, after its first answer, the model is given a second instruction: “Check your answer for errors and improve it.” The model reads its own output like a foreign text and comments on it. It then produces a corrected version. In more elaborate systems, a second, separate model takes on the reviewing role — similar to an editor who signs off on a journalist’s text.

It is important that the model finds different errors when reviewing than when generating. Anyone who writes something themselves and immediately proofreads it easily overlooks what they meant to say rather than what is actually there. That is why it helps to build in a clear separation of roles between the generation step and the review step — either through specific instructions or through a second model.

A well-known procedure from research is called Self-Refine. It systematically prompts the model to formulate feedback on its own output and then write a new version based on that feedback. In tests, Self-Refine performed significantly better on tasks such as code writing or text optimization than a single pass without a correction loop.

Loop corrections in current AI products

Anyone who has asked GPT-4o or Claude to find a bug and fix the code has experienced loop corrections in a basic form. In so-called AI agents — systems that independently work through tasks — correction loops are built in as a core feature. The agent performs an action, checks the result, and decides whether to continue or to try again.

OpenAI’s o1 and o3 models, as well as the Chinese model DeepSeek-R1, use a related idea: they think through a long internal chain before answering, one that includes self-corrections. What appears in the chat as a short answer is the result of many internal iteration steps. These models are therefore significantly ahead of models that operate without such loops in math and programming competitions.

The price is time and computational cost. Every additional round costs money and lengthens waiting time. Systems with loop corrections therefore need a smart stopping condition: when is good good enough? Otherwise the model spins in an endless circle — and eventually stops improving anything at all.