“Base LLMs, without specialized reasoning layers, scored below 10% on the ARC-AGI V1 benchmark.”