CoRT (Chain of Recursive Thoughts) implements a multi-round self-reasoning loop where an AI model generates an initial response, creates multiple alternative responses, evaluates all options against each other, and iteratively refines its output through repeated self-debate. The system dynamically determines how many thinking rounds are needed and uses competitive evaluation to surface the best response from a pool of alternatives. It integrates with LLM APIs like OpenRouter to enhance model performance through forced self-reflection.