Synced
Edited by: Du Wei
Recursive Thinking + Self-Criticism, Can CoRT Bring a Leap in LLM Reasoning?
Everyone is familiar with CoT (Chain-of-Thought), which significantly enhances the reasoning ability of language models by imitating "human problem-solving ideas."
In the past few days, a concept called CoRT (Chain-of-Recursive-Thoughts) has become popular! From its name, it adds the step of "recursive thinking" to CoT.
Specifically, CoRT allows AI models to recursively think about their responses, generate alternative solutions, and choose the best one from them.
This is like giving AI the ability to question or refute itself and try again and again.
By combining "structured self-criticism" and a "recursive thinking mode," it enhances the reasoning ability of language models.
In just two weeks, CoRT's star count on GitHub is close to breaking 2k.
GitHub address: https://github.com/PhialsBasement/Chain-of-Recursive-Thoughts
From a technical principle perspective, compared to traditional CoT, CoRT allows language models not only to think step-by-step but also to repeatedly review and correct during the thinking process, forming a reasoning path similar to human "reflective thinking" or "introspection."
However, many netizens are not too excited about the appearance of CoRT. CoRT is a good technique to make LLMs think harder, but it's not a novel idea. Its working principle is like a meta-prompt with a recursive component added.
Some netizens also pointed out that this method appeared in the 2023 paper "Improving Factuality and Reasoning in Language Models through Multiagent Debate."
Some netizens asked: Isn't CoRT just the thinking mode of most LLMs now?
For example, Gemini 2.5 Pro configured in Cursor does this. The model thinks for a minute and refutes its own answer until it finds the least refutable one.
Another example is, isn't CoRT just the "but wait" mode in Qwen and R1? The model keeps thinking and refuting itself, and there seems to be no difference between the two.
Do you think CoRT is "old wine in a new bottle"? Please leave a comment in the comment section.
Project Introduction
According to the project introduction, the trick of CoRT lies in the following four aspects:
Self-evaluation;
Competitive alternative generation schemes;
Iterative optimization;
Dynamic depth of thought.
The workflow includes the following four steps:
First, the AI generates an initial response.
Second, the AI decides how many rounds of "thinking" it needs.
Then, for each round of thinking:
Generate 3 alternative responses;
Evaluate all responses;
Choose the best response.
Finally, the final response is the survivor of this AI melee.
Web Interface Usage (still in early development stage)
One, open start_recthink.bat
Two, wait a moment for it to install dependencies
Three, configuration successful
If you are running on a Linux system, follow these steps:
Open a new shell:
How are the results?
The author tested with Mistral 3.1 24B, and according to him, CoRT's performance on programming tasks went from "meh" to "holy crap."
Let's look at an example. The image below shows Mistral 3.1 24B + CoRT:
The image below shows Mistral 3.1 24B without CoRT:
From the results, using CoRT transformed the Tic-tac-toe game from a basic CLI (Command Line Interface) to full OOP (Object-Oriented Programming).
Reference link: https://x.com/omarsar0/status/1917401353061818478
© THE END
Please contact this official account for authorization to reprint
Submissions or requests for coverage: liyazhou@jiqizhixin.com