“Base LLMs such as GPT-4.5 and Llama 4 score 0% on the ArcGI 2 benchmark.”