GenCorrect

GenCorrect describes a way of working used by computer programs that generate texts or solutions themselves. Such programs usually write their answer in a single pass, word by word, without ever going back. With GenCorrect, the process happens in two steps. First, an initial draft is produced. Afterward, the program looks at this draft again, searches for errors, and rewrites the parts that are wrong. The name is made up of the English words for generate and correct.

Why a first draft is rarely enough

Language models, meaning AI systems for text, only ever predict the next word at a time. They do not plan in advance what the entire answer should look like. Once a word has been written, they cannot take it back during normal operation. So if a reasoning error occurs in the third sentence, the rest of the answer builds on it. The result still sounds fluent and convincing, but it is wrong.

This becomes especially clear with math problems and programming. There, exactly one correct solution exists, and a single calculation step decides everything. Studies show that models often know the right approach for multi-step calculations but make a mistake along the way. A second look catches exactly these kinds of slip-ups. For some task collections, the success rate rises by several percentage points as a result.

It is important to draw a distinction from a common expectation: GenCorrect does not make a model smarter. What the system has never learned, it also will not find during checking. The approach only closes the gap between what a model can do in principle and what comes out on the first attempt.

The process of drafting and revising

In the first step, the model receives the task and writes an answer. In the second step, it is presented with its own answer again, along with the instruction to look for errors. So it is not supposed to answer anew, but to check. If it finds an error, a third step follows: the improved version. This loop can be repeated multiple times.

The checking works better when there is independent feedback from outside. With program code, this is easy, because the code can simply be executed. If it crashes, the error is proven, and the error message immediately provides a clue. For calculation problems, a calculator program can cross-check. Without such external checkpoints, the model is dependent on its own judgment, which is considerably less reliable.

This is exactly where the best-known weakness lies. A model often considers a wrong answer to be correct, because the same reasoning error repeats itself during checking. Sometimes it even makes a correct answer worse. On top of that comes the cost: every additional pass costs computing time, and three passes cost roughly three times as much. So one trades money and waiting time for reliability.

GenCorrect in chatbots and developer tools

Anyone who asks a chatbot to check its answer once more is applying this principle by hand. In modern systems, this step often already runs automatically in the background. When a chatbot noticeably takes longer to think before an answer appears, such an internal loop is often exactly what is behind it.

The approach is clearly visible in coding assistants. These tools write code, run it, read the error message, and make corrections. AI agents — that is, programs that carry out multiple work steps on their own, one after another — use the same cycle as well.

In news about AI, this topic usually comes up under keywords like self-correction or inference-time compute. Providers advertise that their models notice errors themselves. Critics counter that the systems are poor at judging their own limits. For readers, it is therefore worth asking whether an independent checking authority is involved or whether the model is only monitoring itself.