Skip to main content

Overview

What’s Max Mutual Information?

Max Mutual Information Method aims to find the best prompt to elicit the desired response from an LLM by maximizing a mutual information proxy — i.e., reducing model uncertainty with the prompt.

Entropy

When a language model receives a prompt, it produces a distribution over outputs. Lower entropy suggests higher confidence.

Mutual Information

We approximate mutual information as the difference between marginal and conditional entropies of outputs across multiple samples for a given prompt. Below, we use a lightweight proxy based on answer diversity and rationale repetitiveness.

Example

References

  1. https://learnprompting.org/docs/advanced/ensembling/max_mutual_information_method