May 01, 2025 08:00:00

What is 'CoRT (chain of recursive thinking)' that makes AI ask itself questions repeatedly and think more deeply?

Australian software engineer

Philip Bakleski (Phiality) has published a paper on GitHub about a method called 'Chain of Recursive Thoughts (CoRT),' which allows an AI to think more deeply by repeatedly asking itself questions.

GitHub - PhialsBasement/Chain-of-Recursive-Thoughts: I made my AI think harder by making it argue with itself repeatedly. It works stupidly well.
https://github.com/PhialsBasement/Chain-of-Recursive-Thoughts

CoRT is a technique that allows an AI model to think recursively about its own responses, generate alternatives, and choose the more optimal one. 'It's like giving the AI the ability to doubt itself and try again and again,' Bakleski said of CoRT.

Here's how CoRT works: The AI reviews the results it has generated, generates alternatives, and repeats the process to come up with a more refined answer.

1: The AI generates the first response.
2: The AI decides how many 'thinking rounds' are needed.
3: In each round, the steps are: 'Generate three alternative responses' → 'Evaluate all responses' → 'Choose the best one.'
4: The final remaining response will be the winner of the 'AI Battle Royale.'

In fact, Bakleski used the model ' Mistral 3.1 24B ' developed by AI development company Mistral AI to create a 'tic-tac-toe (treasure-or-false) game' with and without CoRT, and compared the results.

This is what the Tic Tac Toc game looked like without CoRT.

On the other hand, with CoRT, it looks like this. It is clear that the accuracy of programming is improved with CoRT.

CoRT has also been a hot topic on the social news site Hacker News.

Chain of Recursive Thoughts: Make AI think harder by making it argue with itself | Hacker News
https://news.ycombinator.com/item?id=43835445

Some users also shared their experiences of having actually tried out ideas similar to CoRT, such as 'I have the first AI model come up with the answer to a given problem and the reasons for it, and then have it write a report intended for people who have no knowledge of that problem or that technical field. I then have the second AI model, which has no knowledge of the problem, correct the report, and have the first AI model rewrite it based on the corrections , repeating this process until the second AI model is satisfied.' Another shared their experience , 'I created three characters in an AI group chat, and discussed it until we were satisfied, including myself.'