- CoRT enables AI models to recursively think about their responses, generate alternatives, and select the best one.
- The process involves generating an initial response, determining the number of thinking rounds, and refining the response iteratively.
- Each round includes generating 3 alternative responses, evaluating them, and selecting the best one.
- The final response is the result of this iterative refinement process.
- Key features include self-evaluation, competitive alternative generation, iterative refinement, and dynamic thinking depth.
- The method was tested with Mistral 3.1 24B, showing significant improvement in performance, especially for programming tasks.
- The project is open-source under the MIT license, encouraging contributions and modifications.