mirror of
https://github.com/PhialsBasement/Chain-of-Recursive-Thoughts.git
synced 2025-04-29 22:56:02 +03:00
49 lines
1.4 KiB
Markdown
49 lines
1.4 KiB
Markdown
# CoRT (Chain of Recursive Thoughts) 🧠🔄
|
|
|
|
## TL;DR: I made my AI think harder by making it argue with itself repeatedly. It works stupidly well.
|
|
|
|
### What is this?
|
|
CoRT makes AI models recursively think about their responses, generate alternatives, and pick the best one. It's like giving the AI the ability to doubt itself and try again... and again... and again.
|
|
|
|
### Does it actually work?
|
|
YES. I tested it with Mistral 3.1 24B and it went from "meh" to "holy crap", especially for such a small model, at programming tasks.
|
|
|
|
|
|
## How it works
|
|
1. AI generates initial response
|
|
2. AI decides how many "thinking rounds" it needs
|
|
3. For each round:
|
|
- Generates 3 alternative responses
|
|
- Evaluates all responses
|
|
- Picks the best one
|
|
4. Final response is the survivor of this AI battle royale
|
|
|
|
## Examples
|
|
Mistral 3.1 24B non CoRT
|
|
|
|
|
|
Mistral 3.1 24B + CoRT
|
|
|
|
|
|
|
|
## Try it yourself
|
|
```python
|
|
pip install -r requirements.txt
|
|
export OPENROUTER_API_KEY="your-key-here"
|
|
python recursive-thinking-ai.py
|
|
```
|
|
|
|
### The Secret Sauce
|
|
The magic is in:
|
|
|
|
- Self-evaluation
|
|
- Competitive alternative generation
|
|
- Iterative refinement
|
|
- Dynamic thinking depth
|
|
|
|
### Contributing
|
|
Found a way to make it even better? PR's welcome!
|
|
|
|
### License
|
|
MIT - Go wild with it
|