“AI systems using static, single chain-of-thought reasoning achieve scores of approximately 1-2% on the ArcGI 2 benchmark.”