AI Model Improves Dramatically by Arguing with Itself Repeatedly: Introducing CoRT (Chain of Recursive Thoughts)

GitHub user PhialsBasement has developed an innovative approach called Chain of Recursive Thoughts (CoRT), which significantly enhances the performance of AI models by making them engage in repeated self-debate. The method is surprisingly effective, especially for smaller models like Mistral 3.1 24B, where it has shown dramatic improvements in complex tasks such as programming. What is CoRT? CoRT is a technique that encourages AI models to think recursively about their responses. Instead of providing a single output, the AI generates multiple alternatives, evaluates them, and selects the most suitable one. This process is akin to the AI doubting its initial answers and iterating until it arrives at the best possible solution. Does It Actually Work? Yes, it does. When tested with the Mistral 3.1 24B model, CoRT transformed its performance from mediocre to exceptional. This improvement is particularly noteworthy given the relatively small size of the model, which typically limits its capabilities. Here’s how CoRT functions: Initial Response Generation: The AI produces its first response to a given prompt. Determining Thinking Rounds: The AI decides on the number of recursive thinking rounds it needs to refine its answer. Thinking Rounds: For each round, the AI generates new responses, evaluates them against the previous ones, and selects the best option. Final Selection: After all rounds are complete, the AI’s final response is the most refined and accurate one, much like the survivor of an AI battle royale. Examples Mistral 3.1 24B Without CoRT Without CoRT, the model's responses can be hit or miss, often lacking the depth and accuracy needed for complex tasks. Mistral 3.1 24B With CoRT With CoRT enabled, the same model exhibits a remarkable increase in precision and reliability. Its responses become more thoughtful and well-considered, making it a powerful tool for demanding applications. Try It Yourself PhialsBasement has made the CoRT implementation available on GitHub, allowing anyone to experiment with it. The repository includes detailed instructions and example code to help you get started. The Secret Sauce The effectiveness of CoRT lies in several key factors: Recursive Self-Doubt: By forcing the AI to reconsider its own outputs, it can identify and correct errors more efficiently. Iterative Refinement: Each thinking round allows the AI to build upon previous iterations, gradually improving the quality of its responses. Flexibility: The method can be applied to various tasks and models, making it a versatile tool in the AI developer's kit. Contributing If you find ways to improve CoRT, the developer is eager to accept pull requests. The project is open-source and licensed under the MIT License, so feel free to contribute and adapt the code to your own projects. In summary, CoRT represents a novel and highly effective approach to enhancing AI performance by leveraging recursive self-debate. Its ability to significantly boost the capabilities of smaller models opens up exciting possibilities for a wide range of applications, from programming to natural language processing. Whether you’re a seasoned AI researcher or just starting out, this technique is worth exploring.