# Arc Competition Predictions (Part II)

- Date

- Danny Geisz

In a previous post, I made a prediction about the form of the solution that will end up getting above 85% accuracy in the Arc Competition. However, that post was short and imprecise. I’m making this follow-up post to be more precise about my prediction.

### Prediction

Here’s what I think the winning system is going to look like. You have an LLM (or a multi-model model) that looks at representations of the input/output pairs, and this LLM generates hypotheses about the input/output pairs. These hypotheses could be about just about anything — a pattern in the input grids, a pattern in the output grids, a pattern in how the inputs/outputs relate, etc. The LLM additionally generates a test function that can validate the hypothesis. I presume most of the LLM’s hypotheses are going to be wrong, but if you crank the temperature and generate enough of them, I’d suspect the LLM would eventually find a couple that prove valid.

As the LLM produces more and more valid hypotheses, it gets closer and closer to being able to write a function that maps each input grid to the corresponding output. Eventually it generates this function, and then you use this function to make your final prediction.