Basically, ever since GPT-2 was a baby in its machine-crib, people have been trying to figure out how to coax better thoughts out of these robot friends (overlords?) of ours.

The standard is what I included in an example on dynamic variables, called Chain of Thought prompting nowadays, first described here, which is to add the following line to any prompt:

Let’s think step by step.

The results of adding that tiny line were dramatic - text-davinci-002 (aka GPT-3) in the above paper went from 17.7 in an arithmetic benchmark to 78.7.

And that’s still used commonly today by “prompt engineers”.

Now, why does that work?

That simple phrase prompts the language model to break down a larger task into smaller chunks, which let it succeed at more complicated tasks than it could handle by itself.

Basically, I found a cool paper called Self-Discover: Large Language Models Self-Compose Reasoning Structures (February, 2024), which built upon Promptbreeder (I know… I know…). There are similar efforts at Microsoft and Northeastern, as well as Tsinghua University, like Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers.

To optimize problem solving, folks burned significant amounts of GPU time to have language models optimize their own prompts, coming up with unique prompts to handle different problems.

The language model then adapts those specific prompts to the unique task, finally using that auto-generated super prompt to answer far more effectively than it would to the original request.

I have a custom, expanded version of this technique adjusted to fit common chat usecases, all of which is available to you with the click of an unassuming checkbox - “Enhanced Reasoning”.

I also figured out how to do this in a cost-effective manner - I would love to hear your experience with it as I tweak and tune the reasoning capabilities to the work you’re doing!

Feel free to shoot an email to team @ brightrobot.ai to provide feedback or ask any questions - would love to hear from you.