Feature Request: Add support for Automatic Test-Time Scaling

[Macktek]

Recently it was shown in this paper: "Simple test-time scaling" by Niklas Muennighoff et al, that a model's performance can be vastly improved by injecting a token before it concludes to tell it to think more.
Could you consider adding support for this technique with some sort of "number" for how many times it should think before it concludes?

[LostRuins]

You might be able to achieve what you want with the Regex Replace features of Lite.

This allows you to match against any regex of the output and modify it to your preference.

[Macktek]

Yes, the regex would help. But, I need a way to tell it to stop after so many "iterations".