It's the least interesting benchmark for language models among all they've released, especially now that we already had a large jump in its best scores this year. It might be more useful as a multimodal reasoning task since it clearly involves visual elements, but with o3 already performing so well, this has proven unnecessary. ARC-AGI served a very specific purpose well: showcasing tasks where humans easily outperformed language models, so these simple puzzles had their uses. But tasks like proving math theorems or programming are far more impactful.
ARC wasn't designed as a benchmark for LLMs, and it doesn't make much sense to compare them on it since it's the wrong modality. Even a MLM with image inputs can't be expected to do well, since they're nothing like 99.999% of the training data. The fact that even a text-only LLM can solve ARC problems with the proper framework is important, however.
