Meta Prompt Design

You're an artificial intelligence research specialist, familiar with the details of the neural network, good at writing.Prompt。

I need you to write one.Meta-prompt，A simple demand description for subsequent receivers can generate a high-quality version of a large-language modelPrompt。

Suppose we have a language model.MAnd a mission.T。Our goal is to find one.promptp，MakeM(p)Yes.TIt's the best. Traditional methods may require extensive searches or complex optimization algorithms. And the fraction reasoning allows us to define a parameterized one.promptDistribution\(q(p|\theta)\)，And then optimize it.\(\theta\)Jean.\(q(p|\theta)\)As close as you can to reality.promptDistribution.

Sample from Basic Language ModelNindividual output, and select the one with the highest reward value as the final output. This approach does not require fine-tuning of the model to significantly improve the quality of the output while maintaining similarities with the underlying model. The current research on this methodology is also focused mainly on model alignment.

Suppose we have a basic language model.\(p(y|x)\)，of whichxIt's input tips.yis the output generated by the model. We have a reward function.\(r(x,y)\)，To evaluate outputyOrganisationxQuality below.Best-of-NThe workflow of the strategy is as follows:

1.For given inputx，From\(p(y|x)\)Central and independent samplingNOnce, got\(y_1,y_2,\ldots,y_N\)。

2.Calculate the reward value for each sample\(r(x,y_i)\)。

3.Select the highest reward value sample as the final output:\(y^*=\operatorname{argmax}_ir(x,y_i)\)。

Please design one for me on the basis of your ideas.Meta-prompt。