learning to prompt
x

u/adnanqidwai/thoughts/learning-to-prompt.md

learning to prompt

#agents#rl#prompting
  • as we become more and more agent driven, we need to make LLMs write better prompts for other LLMs (subagents, task handoff etc)
  • one way to know if your prompt is good (atleast in current day and age) is to see if it works across different model classes
  • the model which is executing on a task gets some rewards based on the environment
  • we can run n different models and get rewards for each, using the same fixed prompt
  • now what if i make a model write this prompt and reward it as a function of each executor model’s rewards (receiving the model generated prompt as input)
  • it gets better at writing robust prompts with training
  • something like:
Rtotal=α⋅mean⁡m∈Models(Rm)+(1−α)⋅min⁡m∈Models(Rm)R_{\text{total}} = \alpha \cdot \operatorname{mean}_{m \in Models}(R_m) + (1 - \alpha) \cdot \min_{m \in Models}(R_m)
  • feels like i am missing something? Apart from the cost ofcourse