For now, LLMs are a lot better when you express this kind of task in text, preferably as a 1D arrangement. GPT-4's performance on the Abstraction and Reasoning

For now, LLMs are a lot better when you express this kind of task in text, preferably as a 1D arrangement. GPT-4's performance on the Abstraction and Reasoning Corpus. https://arxiv.org/abs/2305.18354 https://arxiv.org/abs/2305.18354 Also, this model trained to predict image tokens seemed to be able to perform Raven's Matrices type questions when expressed as an inpainting task in context. https://yutongbai.com/lvm.html https://yutongbai.com/lvm.html