Hi! It's a custom algorithm based off of this paper: https://arxiv.org/abs/1606.02447
There are a limited number of "actions" the slime can do, such as selecting the left most and right most, selecting all of a certain color, and changing the colors. It generates a big list of guesses (a guess is a series of about 1-5 actions) based on your prompt that it cycles through.
Once it gets one correct, it uses the information from the correct and the incorrect answers to recognize patterns and try to generate a better list of guesses the next time!
This type of machine learning is optimized to learn very rapidly, but for a highly specific task, unlike Large Language Models that have to train on massive amounts of data for use at more generalized or generative tasks.
Thank you for asking!