Chain‑of‑Recursive‑Thoughts (CoRT): Boosting LLM Reasoning with Recursive Self‑Critique

Chain‑of‑Recursive‑Thoughts (CoRT) is a newly popular concept that extends the traditional Chain‑of‑Thought (CoT) approach by adding a recursive thinking step, allowing AI models to repeatedly evaluate and refine their responses.

CoRT enables models to generate alternative solutions, assess them, and select the best one, effectively giving the model a self‑questioning or self‑refutation capability that mimics human reflective reasoning.

The project’s GitHub repository ( https://github.com/PhialsBasement/Chain-of-Recursive-Thoughts ) has quickly attracted nearly 2 k stars, indicating strong community interest.

Technically, CoRT differs from classic CoT by allowing the model to loop back during the reasoning process, checking and correcting its own steps, which some observers describe as a “meta‑prompt with a recursive component.” Similar ideas appeared in the 2023 paper “Improving Factuality and Reasoning in Language Models through Multi‑agent Debate.”

Critics argue that CoRT may not be a fundamentally new idea, noting that many current LLMs already employ comparable recursive or “but‑wait” strategies, such as Gemini 2.5 Pro, Qwen, and R1.

Project Overview

The CoRT methodology focuses on four key aspects: self‑assessment, generation of competitive alternative solutions, iterative optimization, and dynamic depth of thought.

The workflow consists of four steps:

AI generates an initial response.

AI decides how many rounds of “thinking” are needed.

For each round, the model generates three alternative responses, evaluates all of them, and selects the best one.

The final chosen response is output as the result of the AI’s internal competition.

Web Interface (early development)

To run the demo, users open start_recthink.bat , wait for dependencies to install, and then follow the on‑screen instructions. Linux users can follow the provided shell commands (images omitted for brevity).

Performance tests using Mistral 3.1 24B showed a dramatic improvement in programming tasks, with CoRT turning a “meh” result into a “holy‑crap” one. Visual comparisons illustrate that a Tic‑Tac‑Toe CLI implementation evolved into a full object‑oriented program after applying CoRT.

Overall, the article invites readers to comment on whether CoRT is truly innovative or merely “new wine in old bottles.”